Nothing Special   »   [go: up one dir, main page]
Next Article in Journal
Spatiotemporal Variation in Aerosol Optical Depth and Its Potential Effects on Snowmelt in High Mountain Asia from 2004 to 2023
Previous Article in Journal
Potential Impact of Sea Surface Temperature Variability on the 2007 Sudden Bloom of Ulva prolifera in the Southern Yellow Sea
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 16
Issue 23
10.3390/rs16234408
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

A Framework for Separating Climate and Anthropogenic Contributions to Evapotranspiration Changes in Natural to Agricultural Regions of Watersheds Based on Machine Learning

by
Zixin Liang
1,2,3,
Fengping Li
1,2,3,*,
Hongyan Li
1,2,3,
Guangxin Zhang
4 and
Peng Qi
4
1
Jilin Provincial Key Laboratory of Water Resources and Environment, Jilin University, Changchun 130021, China
2
Key Laboratory of Groundwater Resources and Environment, Jilin University, Ministry of Education, Changchun 130021, China
3
College of New Energy and Environment, Jilin University, Changchun 130021, China
4
Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences, Changchun 130102, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2024, 16(23), 4408; https://doi.org/10.3390/rs16234408
Submission received: 30 September 2024 / Revised: 14 November 2024 / Accepted: 18 November 2024 / Published: 25 November 2024
Browse Figures
Figure 1
<p>The geographical location of the Songhua River Basin.</p> "> Figure 2
<p>The flowchart of Random Forest Regression.</p> "> Figure 3
<p>Conceptual framework of the method for quantifying evapotranspiration influenced by climate change and that influenced by human activities of human-managed land cover types (taking the rainfed agricultural transition areas in saline-alkali land as an example).</p> "> Figure 4
<p>Land cover change in the study area: (<b>a</b>) Songhua River Basin in the 1980s, (<b>b</b>) Songhua River Basin in the 2015s and (<b>c</b>) Land use transfer contribution. F, forest area; G, grassland area; M, marshland area; SA, saline-alkali land; R, rainfed agriculture; I, irrigated agriculture; S, settlement; W, water.</p> "> Figure 5
<p>Annual anomaly and cumulative anomaly of evapotranspiration (ET) for forest (<b>a</b>), grassland (<b>b</b>), marshland (<b>c</b>), and saline-alkali land (<b>d</b>) in the Songhua River Basin from 1980 to 2015, along with the spatial distribution of both the average annual ET (<b>e</b>–<b>h</b>) and its changing trends (<b>i</b>–<b>l</b>) across the basin.</p> "> Figure 6
<p>Cross-validation of ET<sub>n</sub> prediction for the four types of natural areas.</p> "> Figure 7
<p>The importance of the variables for four regional ET<sub>n</sub> prediction models.</p> "> Figure 8
<p>Spatial distribution of ET<sub>m</sub> and ET<sub>h</sub> in the natural (forest, grassland, marshland, and SA) to rainfed agriculture areas from 1980 to 2015. SA is short for saline-alkali land.</p> "> Figure 9
<p>Spatial distribution of ET<sub>m</sub> and ET<sub>h</sub> in the natural (forest, grassland, marshland, and SA) to irrigated agriculture areas from 1980 to 2015. SA is short for saline-alkali land.</p> "> Figure 10
<p>Climate and anthropogenic contributions to evapotranspiration changes from 1980 to 2015 in the natural to agricultural areas.</p> "> Figure 11
<p>Component planes of the seven training parameters in the SOM of the ecological–agricultural transformation region and average <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>ET</mi> </mrow> <mrow> <mi mathvariant="normal">h</mi> </mrow> </msub> </mrow> </semantics></math> in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).</p> "> Figure 12
<p>Component planes of the seven training parameters in the SOM of the natural–agricultural transformation region and average <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>ET</mi> </mrow> <mrow> <mi mathvariant="normal">m</mi> </mrow> </msub> </mrow> </semantics></math> in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).</p> "> Figure 13
<p>Cross-validation of ET<sub>n</sub> prediction for the four types of natural areas based on XGBoost algorithm.</p> ">
Versions Notes

Abstract

:
Evapotranspiration is a crucial component of the water cycle and is significantly influenced by climate change and human activities. Agricultural expansion, as a major aspect of human activity, together with climate change, profoundly affects regional ET variations. This study proposes a quantification framework to assess the impacts of climate change (ETm) and agricultural development (ETh) on regional ET variations based on the Random Forest algorithm. The framework was applied in a large-scale agricultural expansion area in China, specifically, the Songhua River Basin. Meteorological, topographic, and ET remote sensing data for the years of 1980 and 2015 were selected. The Random Forest model effectively simulates ET in the natural areas (i.e., forest, grassland, marshland, and saline-alkali land) in the Songhua River Basin, with R2 values of around 0.99. The quantification results showed that climate change has altered ET by −8.9 to 24.9 mm and −3.4 to 29.7 mm, respectively, in the natural areas converted to irrigated and rainfed agricultural areas. Deducting the impact of climate change on the ET variation, the development of irrigated and rainfed agriculture resulted in increases of 2.9 mm to 55.9 mm and 0.9 mm to 53.4 mm in ET, respectively, compared to natural vegetation types. Finally, the Self-Organizing Map method was employed to explore the spatial heterogeneity of ETh and ETm. In the natural–agriculture areas, ETm is primarily influenced by moisture conditions. When moisture levels are adequate, energy conditions become the predominant factor. ETh is intricately linked not only to meteorological conditions but also to the types of original vegetation. This study provides theoretical support for quantifying the effects of climate change and farmland development on ET, and the findings have important implications for water resource management, productivity enhancement, and environmental protection as climate change and agricultural expansion persist.

1. Introduction

Actual evapotranspiration (ET) is defined as the total amount of water that evaporates and transpires from plants under specific climatic, soil, and crop conditions. It plays a crucial role in the terrestrial hydrological cycle, influencing the allocation of water resources and the demand for irrigated agriculture [1,2]. In recent years, an intensification of the global hydrological cycle has been observed [3,4]. Previous studies indicate that changes in ET are primarily driven by meteorological and land use changes [5,6]. In the context of rapid global warming, there has been an increase in research focused on the impacts of climate change on ET [7]. For example, variations in meteorological factors, such as precipitation, temperature, incident solar radiation, wind and atmospheric teleconnections regulate evapotranspiration and its variability across different terrestrial ecosystems [8,9,10]. Understanding the mechanisms through which climatic factors influence the ET process is essential for accurately predicting ET trends and changes [11]. Anthropogenic land use/land cover changes, along with water regulation and management have altered the original hydrological environment, directly impacting the amount and timing of ET [12,13,14]. Additionally, anthropogenic greenhouse gas emissions and aerosol production can alter the energy balance both at the surface and in the atmosphere, affecting climate patterns on global and regional scales, thereby indirectly influencing ET [15,16]. As research progresses, increasing attention has been directed toward distinguishing the effects of anthropogenic perturbations from those of climate change on evapotranspiration [17,18]. However, the diversity and complexity of human activities complicate direct quantitative assessments of their impacts on ET [19]. Therefore, indirect methods are commonly used to quantitatively assess the contribution of human activities to ET. These methods typically begin by calculating natural ET (ETn) and subsequently derive anthropogenic ET (ETh) by subtracting ET n from the total regional ET [17,20].
The advent of remote sensing technology has made accurate estimation of ETn more feasible. Currently, the dominant approach involves utilizing Gravity Recovery and Climate Experiment (GRACE) data in conjunction with hydrological modeling. Pan et al. [21] and Zheng et al. [22] identified that human activities have led to a 12% increase in ET in the Haihe River Basin, China , whileLiu et al. [23] reported a 13.5% increase in total ET due to human activities in the Ziya–Daqing Basins. However, this method presents several significant drawbacks: (1) The GRACE product has a coarse spatial resolution of 100,000 square kilometers, leading to substantial errors in ET estimation; for instance, in the Haihe River Basin, the estimated error reached 32.6 mm [21]. (2) The regional gravity change signal monitored by GRACE reflects not only changes in water storage but also gravity changes resulting from the extraction of energy minerals and the movement of goods in and out of the region [24]. (3) This approach can estimate the overall contribution of human activities to ET within a basin but cannot quantify the ET variation attributable to specific human activities. Scenario assumptions represent another method for estimating ET. This approach calculates the total ET of a watershed by subtracting ET contributions from various natural environments (e.g., forest, grassland, marshland) separately [25]. However, this method does not resolve the challenge of distinguishing between ET generated by different species.
In recent years, machine learning and intelligent algorithms have gained prominence and are widely used across various disciplines including geography, ecology and hydrology [26]. For instance, these techniques are commonly employed in the studies about hydrological variables forecasting [27] and flow coefficient estimation [28]. The Random Forest algorithm, proposed by Breiman (2001), is a popular machine learning method which has been widely adopted for both regression and classification tasks due to its low computational cost and availability of large amounts of data. As an ensemble learning method, Random Forests effectively address the limitations of single-model approaches. For instance, Support Vector Machines (SVMs) perform well when dealing with small datasets. However, their performance deteriorates and the training times increase when dealing with large or noisy datasets. Similarly, linear regression algorithms, although intuitive and straightforward, are primarily effective only when there is a linear relationship between variables and perform poorly with nonlinear data. RF alleviates these issues by aggregating the predictions of multiple decision trees, thereby enhancing the model’s generalization ability and stability [29]. Additionally, the Random Forest algorithm uses a specialized bagging technique (bootstrap aggregating), which greatly reduces the model variance and increases the accuracy, particularly when the decision trees exhibit low correlation [30].
In this study, we aim to harness the computational power of machine learning to differentiate between natural and anthropogenic ET, using land cover as an intermediary. The objectives of this study are as follows: (1) to investigate the intrinsic relationship between ET and environmental variables in the Songhua River Basin through machine learning, and to establish a framework for distinguishing between evapotranspiration influenced by climate change and that influenced by human activities; (2) to quantify the contributions of climate change and agricultural activities to the ET variation in natural areas that experienced agricultural expansion from 1980 to 2015 in the SRB; (3) to interpret the degree of dependence and driving mechanisms of ET in relation to environmental factors. The findings provide essential insights into assessing the hydrologic effects of farmland development, supporting sustainable strategies for regional land and water management, food productivity, and environmental protection under the changing climate.

2. Data and Methods

2.1. Study Area

In this paper, the SRB was selected as the study area (Figure 1). The Songhua River Basin is one of the seven major river basins in China (41°42′–51°38′ N, 119°52′–132°31′ E). It geographically covers 24 prefectural-level cities (leagues) in four provinces (autonomous regions), namely Heilongjiang, Jilin, Liaoning, and the Inner Mongolia Autonomous Region, with a total watershed area of about 56.12 thousand square kilometers. The Songhua River Basin is rich in water resources, with an annual runoff of about 733 billion cubic meters. The main water systems in the basin include tributaries such as the Songhua River, the Second Songhua River, and the Nenjiang River. These water resources not only provide important support for agricultural irrigation, but are also a significant source of hydropower generation in the region. Large and medium-sized water conservancy projects, such as the Fengman Power Station and Baishan Reservoir, have significantly improved the deployment of regional water resources and flood control capacity. Additionally, the basin has a typical seasonal rainfall pattern, characterized by concentrated precipitation in summer and periods of freezing in winter. The average annual precipitation in the basin ranges from 341.3 to 1010.9 mm, and the ratio of the maximum annual precipitation to the minimum annual precipitation is about 3:1. This variability makes the region prone to periods of continuous multi-year high rainfall or low rainfall and frequent occurrences of extreme weather events. Statistical data up to 2020 show that the primary land use types in the basin are arable land and forest land, followed by grassland. The basin is dominated by deep black soils and receives ample sunlight, conditions that provide a favorable environment for the growth of major food crops such as rice, maize, and soybeans on both sides of the river.

2.2. Data

2.2.1. Evapotranspiration

Actual evapotranspiration (ET) data are difficult to obtain, so there has been a gradual shift towards focusing on satellite products to fill the observational gaps for more efficient and wider access to relevant data [31]. In this study, integrated ET products from 1980 and 2015 with a spatial resolution of 0.25 degrees (quoted) (https://data.tpdc.ac.cn (accessed on 20 September 2024)) were used. The dataset was generated by reanalyzing and fusing ERA5, MERRA2 and GLDAS-Noah data [32]. The ET products were interpolated to a 1 km grid by three convolutions in this study.

2.2.2. Environment Factors

Meteorological, topographic, and geographic location data were used as input variables to build the ET separation model in the Random Forest Regressor (RFR). All these data were interpolated into grid data with an accuracy of 1 km × 1 km using the cubic convolution method.
(1)
Meteorological data
In this study, the meteorological factors including precipitation (Pre), surface downward longwave radiation (Lrad), surface downward shortwave radiation (Srad), near-surface air temperature (T), near-surface wind speed (Wind), near-surface air pressure (Pres), and near-surface specific humidity (Shum) were selected as the influencing factors of ET. Monthly precipitation data were generated from A Big Earth Data Platform for Three Poles data (https://poles.tpdc.ac.cn (accessed on 20 September 2024)). This dataset spans from January 1901 to December 2021 and features a spatial resolution of approximately 1 km (0.0083333°). It was validated using data from 496 independent meteorological observations, confirming its reliability [33]. Additionally, six near-surface meteorological elements, including 2 m air temperature (Temp), surface pressure (Pres), specific humidity (Shum), 10 m wind speed (Wind), downward shortwave radiation (Srad) and downward longwave radiation (Lrad) were generated from the China Meteorological Forcing Dataset (CMFD) (https://data.tpdc.ac.cn (accessed on 20 September 2024)). CMFD is a gridded near-surface meteorological dataset that was developed specifically for studies of land surface processes in China [34]. The dataset was made through fusion of remote sensing products, a reanalysis dataset and in situ observation data at weather stations. The time span of the dataset is from January 1979 to December 2018, with a temporal resolution of 3 h and a spatial resolution of 0.1°.
(2)
Topographic data
The topographic dataset used in this study includes elevation, slope, and aspect. Digital elevation (DEM) data were obtained from the Shuttle Radar Topography Mission (SRTM) data (version 4.1), which have a spatial resolution of 90 m (https://www.gscloud.cn (accessed on 20 September 2024)). The slope (Slo) and aspect (Asp) data were derived from these DEM data through the spatial calculations.

2.2.3. Land Cover

The land cover data were sourced from the China Multi-Phase Land Use Remote Sensing Monitoring Dataset (CNLUCC), which offers a spatial resolution of 30 m. These data are available on the Resource and Environmental Science and Data Platform (https://www.resdc.cn (accessed on 20 September 2024)). The dataset integrates land use and land cover information across China, categorizing land use into three levels and 66 types. This classification aids in enhancing the prediction and forecasting of land use change trends. In this study, the original land cover classes from the LUCC were subdivided into seven categories, including irrigated agriculture, rainfed agriculture, settlement, forest, grassland, marshland, water, and saline-alkali land (this land use type primarily comprises saline-alkali land and wasteland, with saline-alkali land constituting over 95% of the area; therefore, this paper specifically refers to saline-alkali land), among which forest, grassland, marshland, and saline-alkali land were regarded as the natural land use types and irrigated agriculture, rainfed agriculture, and settlement were regarded as the human-managed land use types.

2.3. Methods

2.3.1. Trend Analysis

In this paper, the Theil–Sen estimator test was used to analyze the trends of meteorological elements and multi-year evapotranspiration from 1980 to 2015. The Theil–Sen estimator, also known as Sen’s slope estimator or the Theil–Sen median slope estimator, is a method used in statistics to find a robust line of best fit for a set of two-dimensional points [35]. This method is particularly useful because it is resistant to outliers in the data. It is often preferred in situations where the data might have a significant amount of noise or when outliers are likely to be present.
β = Median x j − x i j − i ∀ j > i
where Median() represents the median value; if β is greater than 0, it indicates an increasing trend in the meteorological element, and vice versa for a decreasing trend.
To further determine the significance of the trends of evapotranspiration and other meteorological elements, the Mann–Kendall (MK) test was also employed in this study. MK test is a non-parametric statistical test used to analyze trends in a dataset [36]. This test does not assume that the data follow a particular distribution, so it is resistant to outliers and is applicable to data that are not normally distributed. The MK test is commonly used in environmental science and meteorology to determine if there is a significant upward or downward trend in a time series.
Calculation of statistic S is as follows:
S = ∑ i = 1 n - 1 ∑ i + 1 n sgn x j − x i
where sgn () is the sign function (the sign used to calculate the difference between x j and x i ), calculated as follows:
sgn x j − x i = + 1 x j − x i > 0 0 x j − x i = 0 − 1 x j − x i < 0
A trend test is performed using the test statistic Z. The Z value is calculated as follows:
Z = S Vars S S > 0 0 S = 0 S + 1 Vars S S < 0
where Vars (S) is calculated as follows:
Vars S = n n   − 1 2 n + 5 18
In this paper, the bilateral trend test is used, and the critical value Z1−α/2 is found in the normal distribution table at a given level of significance. When |Z| ≤ Z1−α/2, the original hypothesis is accepted, i.e., the trend is not significant; if |Z| > Z1−α/2, the original hypothesis is rejected, i.e., the trend is considered to be significant. Given the significance level α = 0.05 and α = 0.01, the critical value Z1−α/2 is equal to ±1.96 and ±2.58, respectively. Therefore, when the absolute value of Z is greater than 1.96 and 2.58, it means that the trend passes the significance test with a confidence level of 95% and 99%, respectively. The criterion of the significance of the trend is shown in Table 1.

2.3.2. Land Use Change Analyses

Transfer matrix analysis was used in this study to explore the overall land use changes in the Songhua River Basin and identify the regions where natural land use types have changed to human-managed land from 1980 to 2015. The land use transfer matrix is often represented by a matrix with the following mathematical expression:
S ij = S 11 ⋯ S 1 n ⋮ ⋱ ⋮ S n 1 ⋯ S nn
where Sij is the area that was transferred from land i to land j during the study period, km2; n is the land use type.

2.3.3. Introduction to Models

(1)
Random Forest Regression (RFR)
In this paper, RFR is applied to estimate the evapotranspiration for the natural land covers (i.e., forest, grassland, marshland, and saline-alkali land) in the Songhua River Basin. The RFR workflow involves a bootstrap sampling method to repeatedly sample the training data, generating multiple independent data subsets, each of which is utilized to train a distinct decision tree. Following the identification of optimal hyperparameters and the application of out-of-bag (OOB) data to evaluate the model’s generalization capacity, predictions can then be made on new data. In regression tasks, the predicted value of the dependent variable is derived from the average value of that variable across the results of these trees. Evapotranspiration is significantly influenced by meteorological parameters, which are closely related to atmospheric conditions and are affected by topography and geographical location. Therefore, Met (Pre, Lrad, Srad, Temp, Wind, Pres, Shum), Top (DEM, Slo, and Asp), and Geo (Longitude (Long) and Latitude (Lat)) were employed to correspondingly develop the ET n models for the four natural land covers (Figure 2).
These models were split into a 70% training group and a 30% validation group, with the validation dataset comprising 122,480 data points for forest, 41,690 for grassland, 14,715 for marshland, and 5902 for saline-alkali land to ensure the models’ reliability. A grid search with cross-validation optimized the RFR’s hyperparameters.
(2)
Evaluation of the Random Forest Models
The performance of the Random Forest models for the natural land use types in estimating evapotranspiration was assessed using four metrics including the correlation coefficient (R2), mean absolute error (MAE), and root mean square error (RMSE).
(1)
Correlation coefficient (R2)
R2 indicates the degree of linear correlation between the predicted values and the actual values and reflects the model’s ability to explain the variation in the data series. It is expressed as follows:
R 2 = 1   − ∑ i = 1 n ( y i − y ^ i ) 2 ∑ i = 1 n ( y i − y - ) 2
where y i represents actual values of ET n obtained from remote sensing products, y ^ i is the predicted values of ET n by RFR models, y - is the average of actual values, and n is the number of data points.
When the model perfectly fits all the data points, the R2 value reaches 1, indicating complete explanatory power over the observed variability.
(2)
Mean absolute error (MAE)
MAE measures the average of the absolute differences between predicted values and actual values and is calculated as follows:
MAE = 1 n ∑ i = 1 n y i − y ^ i
A lower MAE indicates that the model’s predictions are closer to the actual data.
(3)
Root mean square error (RMSE)
RMSE, also known as the standard error, more intuitively captures the average error between the predicted values and the actual values.
RMSE = 1 n ∑ i = 1 n y i − y ^ i 2
When the predicted value is completely consistent with the actual value, RMSE equals zero.

2.3.4. Quantification Framework for ETm and ETh

Typically, in the absence of human intervention, evapotranspiration in the natural areas (i.e., forest, grassland, marshland, and saline-alkali land) was predominantly affected by natural factors, including meteorology, topography, and geographic location. When the natural areas are converted into the managed land use types due to human activities, such as the transformation of grassland into irrigated agriculture, the evapotranspiration will be determined by both anthropogenic and natural factors. Therefore, in this study, the ETt of areas that have transitioned from natural to anthropogenic-managed agricultural land use types for a certain period can be expressed as follows:
ET t = ET n + ET h
where ETt represents the total evapotranspiration over the region; ETn is the fraction contributed by natural factors; ETh is the fraction resulting from various human activities, such as farming and irrigation, within the managed land cover of the watershed.
In our study, we specifically examine the changes in ET in areas where land cover has transitioned from natural to agricultural types between 1980 and 2015. Consequently, we designate 1980 as the starting year and 2015 as the ending year for calculating the change in ET (ΔET). It is assumed that natural land cover solely produces ET due to natural factors. Since neither the Topographic data nor the geographic location changed during the study period, the changes in ET from natural factors between 1980 and 2015 were primarily driven by climate change. Thus, ET can be defined for the years 1980 and 2015, respectively.
ET 1980 = ET n 1980
ET 2015 = ET n 2015 + ET h
So ΔET is calculated as follows:
Δ ET = ET 2015 − ET 1980 = ET n 2015 − ET n 1980 + ET h = ET m + ET h
where ETm denotes the change in ET due to shifts in meteorological factors. When ETm (or ETh) is greater than 0, it indicates that climate change (or agricultural activities) has contributed to an increase in ET, and vice versa.
Then, ETh can be expressed as follows:
ET h = ET 2015 − ET n 2015
In order to predict the ETn2015 for managed agricultural land cover and subsequently quantify ETm and ETh, we conducted a preliminary classification of land cover types from 1980 and 2015 based on land cover data (Figure 3). Using transfer matrix analysis, the distributions of unchanged natural areas, unchanged human-managed areas, areas converted from natural to human-managed, and areas changed from human management back to natural degradation in the Songhua River Basin between 1980 and 2015 were identified. Subsequently, the ET, Met, Top, and Geo datasets were categorized by land cover type. First, ET and environmental variables were extracted from surface cover grid points of four natural land cover regions (i.e., forest, grassland, marshland, and other natural areas). Then, Random Forest Regression (RFR) models were employed to explore the intrinsic relationship between evapotranspiration (ET) and annual meteorological factors (P, Shum, T, Srad, Lrad, Wind, Pres), topographic factors (Ele, Slo, Asp), and geographic factors (Lon, Lat) for each natural land cover region in 1980. Ultimately, these four ET prediction models were utilized to simulate the ETn for the regions transformed from natural to agricultural covers. This study posits that ETn is determined by meteorological, topographic, and geographic factors with human interference absent.
ETn in 2015 is characterized as follows:
ET n 2015 = f ( Pre ,   Shum ,   Temp ,   Srad ,   Lrad ,   Wind ,   Pres ,   Lon ,   Lat ,   Dem ,   Slo ,   Asp ) 2015

2.3.5. Attribution Analysis for the Spatial Heterogeneity of ETh and ETm

To explore the relationship between ETn, ETh and meteorological elements, this study employs the SOM, a clustering method based on Self-Organizing Mapping Neural Networks that utilizes unsupervised learning to organize and classify high-dimensional data. SOM is particularly effective for complex datasets, as it reduces the dimensionality while maintaining the topology of the data. The MATLAB 2021b software was used to build an SOM model for spatial environment variable separation. Considering both quantitative and structural errors, the final neural matrix comprises 3 × 3 neurons. The SOM model output usually includes U-matrices (Uniformity Matrix) and Component Planes, aiding in understanding and visualizing the structure of the high-dimensional data. The U-matrix visually represents the distances between SOM nodes, with colors indicating similarity or distance between each node and its neighboring nodes on the map [37]. Darker colors indicate greater differences between neighboring nodes, and lighter colors indicate more similarity between nodes. The component planes map each input feature of the SOM to a 2D space and demonstrate the distribution of features in the dataset [38].

3. Results

3.1. Changes from Natural Land Use to Cropland in the SRB

Land use patterns in the SRB for the years 1980 and 2015 are illustrated in Figure 4. It can be found that land use types in the SRB significantly changed from 1980 to 2015. The area of natural land use regions experienced an overall reduction, except for the saline-alkali land. Specifically, grassland, forest and marshland in the SRB decreased by 2.11%, 2%, and 0.8%, respectively, while saline-alkali land increased by 0.13% of the total basin area.
Anthropogenic activities in the SRB have led to an increase in both cultivated land and settlement areas, with rainfed agriculture experiencing the most significant increase from 29.08% to 31.7%. Irrigated agriculture and settlement also increased by 2.05% and 0.33%, respectively. The primary characteristic of land use conversion in the SRB was the transformation of natural areas into agricultural land. The results of land use transfer matrix indicated that grassland and forest were the primary natural land use types affected by agricultural activities (Table 2), with grassland mainly converting to rainfed agriculture (64.03%) and a smaller portion to irrigated agriculture (~8.23%). Forest area was mainly transferred to rainfed agriculture and grassland, accounting for 66.28% and 27.1% of the total forest transfers, respectively. In addition, 68.58% of the rainfed agriculture was transferred to irrigated agriculture, becoming one of the main sources of irrigated agricultural land. The details of land use transfer within the watershed are illustrated in Figure 4c.

3.2. ET Changes in the Natural Land Use Regions in the SRB

Figure 5 shows the trends and spatial distribution of the ET in natural areas within the Songhua River region from 1980 to 2015. There is considerable interannual variation in ET, accompanied by notable spatial differences across the basin.
The ET of forested areas within the Songhua River Basin showed an increasing trend over the study period, with annual ET ranging from 403 to 696 mm. The peak value of ET appeared in 2013, exceeding the multi-year average by 81 mm, while the lowest happened in 2007, falling 52 mm below the average. The cumulative ET from 1980 to 2001 was positive, followed by a sharp decline from 2002 to 2010, and a substantial increase from 2011 to 2015 (Figure 5a,e,i). The trend in ET for grassland region was similar with that of forested areas, with average annual ET ranging from 404 to 696 mm. The maximum ET in grassland also occurred in 2013, exceeding the average value by 103 mm, while the minimum ET was observed in 2001, which was 76 mm below the average. The change patterns of ET in marshland areas and saline-alkali lands were analogous, both exhibiting a distinct negative trend over the period (Figure 5c,g,k,d,h,l). The lowest ET values in the marshland and other natural regions were observed in 2001, being 83 mm and 93 mm lower than the average values, respectively.
The interannual trend of the Songhua River Basin was assessed using the Theil–Sen median method, with the significance testing conducted via the Mann–Kendall method. Forested areas were predominantly characterized by decreasing ET. The regions showing a decreasing trend accounted for a proportion of 69.61% (Table 3), yet an increasing trend was observed in 30.71% of the total areas, primary located in the main Songhua River Basin and the second Songhua River Basin. Grassland and marshland showed a clear increasing trend, with an increasing area share of 51.92% and 62.03%, respectively. Saline-alkaline land showed a particularly pronounced increasing trend, with 94.12% of the area experiencing an increase in ET.

3.3. Performance of the Random Forest Regression Model in ET Simulation

This study developed four ETn prediction models on a scale of 1 km × 1 km driven by 12 environmental predictors, respectively, for the four types of natural regions in the Songhua River Basin, specifically including forest, grassland, marshland, and saline-alkali land. The optimal parameters for the models corresponding to each type are listed in Table 4. The predicted results of ETn exhibited a strong correlation with the ET acquired from remote sensing products. The R2 values of the forest model, grassland model, marshland model, and saline-alkali land model were all 0.99. The error metrics for these four natural models are shown in Figure 6. Moreover, the scatter points closely align with the y = x line, indicating the robust performance of the ETn models developed for these four natural areas. The excellent performance of the four ETn models confirms the feasibility of the ETn prediction in this study, and the annual evapotranspiration changes driven by natural factors can be effectively predicted using the annual meteorological dataset, Top dataset and Geo dataset.
RFR based on Impurity was used to assess the importance of 12 environmental predictors in the ETn prediction model, respectively, for forest areas, grassland areas, marshland areas, and saline-alkali land areas (Figure 7). The Impurity method determines variable importance by calculating the total reduction in Impurity resulting from each variable during node splits across all trees. This method is also referred to as “Incremental Impurity Importance” (IncNodePurity). It specifically assesses variable impact by calculating the residual sum of squares at each node, which reflects the contribution of each variable to reducing heterogeneity in the observed values at decision tree nodes, thereby determining the relative importance of each variable. As presented in Figure 7, the importance of the influencing factors varied across different natural regions. For instance, in forest areas, the top 5 environmental predictors were latitude, wind, Shum, longitude, and precipitation. In grassland areas, longitude, temperature, latitude, Lrad and DEM emerged as the primary predictors. Similarly, in marshland areas, Temp, Srad, Lon, DEM and Lat ranked as the top predictors. Conversely, in saline-alkali land, Lon, Srad, Pre, pressure and Temp were identified as the most important variables. Overall, the location, particularly latitude and longitude, along with the energy conditions are the key factors influencing regional evapotranspiration in the Songhua River Basin. Based on sensitivity analysis, the factors Asp and Slo, which were not sensitive in all four models, were excluded to enhance the reliability of the models.

3.4. Separation of Climate and Anthropogenic Contributions to ET Changes in the Natural–Agricultural Areas

3.4.1. Separation in the Natural to Rainfed Agriculture Areas

Based on the framework proposed in Figure 3, this study generated ETm and ETh for the areas transferred from natural ecosystems to human-managed agricultural areas (i.e., rainfed agriculture and irrigated agriculture areas) in 2015 (Figure 8 and Figure 9). During 1980–2015, there was a notable increase in evapotranspiration in regions converted from natural areas to rainfed agriculture (Figure 10), driven by both climate change and anthropogenic activities. Among the four types of natural land cover, forests contributed the largest area converted to rainfed agriculture in the Songhua River Basin (Figure 4). The ETt in the transition area from forest to rainfed agriculture (i.e., forest–rainfed agriculture) in 2015 was 548.0 mm, of which ETn was 547.1 mm and ETh was 0.9 mm. Additionally, the transformation from grassland to rainfed agriculture was also extensive. This region exhibited the most significant increase in evapotranspiration attributed to anthropogenic influences compared to other natural transition areas. The converted region from grassland to rainfed agriculture (i.e., grassland–rainfed agriculture) achieved an ETt of 535.9 mm in 2015, comprising 482.5 mm from ETn and 53.4 mm from human activities (ETh). Similarly, the transition region from marshland to rainfed agriculture (i.e., marshland–rainfed agriculture), recorded an ETt of 555.1 mm in the same year, with contributions of 532.2 mm from ETn and 22.8 mm from ETh. Human activities also significantly enhanced evapotranspiration in this region, ranking second only to the grassland–rainfed agriculture region. The saline-alkali land–rainfed agriculture region, which covered the smallest area, recorded an ETt of 510.1 mm, with ETn at 496.6 mm and ETh at 13.2 mm. As illustrated in Figure 8, the spatial distribution pattern of ETh can be clearly observed, showing that human activities have had a positive impact in almost all saline-alkali land–rainfed agriculture areas. Moreover, ETh demonstrated consistent distribution patterns across the transition regions from grassland and marshland to rainfed agriculture, with calculations indicating higher values in the southern regions and lower in the eastern regions.
The impact of climate change on ET across four transition regions was significant and predominantly positive. In the marshland–rainfed agriculture region, climate change led to a substantial increase in annual total evapotranspiration. The ETn in this region was 502.6 mm in 1980. By 2015, the increase in ETm amounted to 29.7 mm, that equals an increase of 60.9 million cubic meters of water consumption driven by climate change. Similarly, in the saline-alkali land–rainfed agriculture region, climate change had a positive effect on evapotranspiration. In 1980, the ETn in this region was 476 mm, and by 2015, the additional ETm had increased by 20.9 mm. In the forest–rainfed agriculture region, ETn was 528.3 mm in 1980, resulting in an ETm increase of 18.8 mm by 2015. Conversely, the grassland transition region exhibited a negative response to climate change. In this region, ETn was 485.8 mm in 1980, and climate change resulted in a decrease in evapotranspiration of 3.4 mm by 2015. Figure 8 illustrates the spatial distribution of ETm in 2015, identifying common characteristics among the four transition regions: areas where evapotranspiration has declined due to climate change are primarily concentrated in the main basin of the Songhua River and the Second Songhua River Basin, exhibiting higher values in the north and lower values in the south.

3.4.2. Separation in the Natural to Irrigated Agriculture Areas

Among the four natural regions, the marshland contributed the largest area converted to irrigated agriculture. In 2015, the ETt in the marshland–irrigated agriculture transition region was 572.5 mm, of which ETn accounted for 543.4 mm and ETh for 29.1 mm. Significant transformation was also observed from grassland to irrigated agriculture, and the grassland–irrigated agriculture region experienced the most pronounced increase in evapotranspiration attributed to human activities. In this region, ETt in 2015 was 565.2 mm, comprising ETn of 509.3 mm and ETh of 55.9 mm. Conversely, the increases in evapotranspiration due to human activities in the forest–irrigated and saline-alkali land–irrigated agriculture regions were relatively minor, recorded at 2.9 mm and 30.3 mm, respectively. The spatial distribution of ETh within the watershed indicates that the additional soil water recharge from irrigated agriculture generally elevates evapotranspiration across all conversion areas.
Climate change also had a significant impact on evapotranspiration across the four regions transitioning from natural to irrigated agriculture. Specifically, climate change positively influenced ET changes in the marshland–irrigated agriculture and saline-alkali land–irrigated agriculture regions. In the saline-alkali land–irrigated agriculture region, climate change led to an increase of 24.9 mm (ETm) in evapotranspiration by 2015 from the baseline ETn value of 461.4 mm in 1980. Similarly, climate change caused the ET in the marshland–irrigated agriculture region increased by 21.8 mm from 1980 to 2015. Conversely, the forest–irrigated agriculture and grassland–irrigated agriculture regions exhibited a negative response to climate change with respect to evapotranspiration. In the forest–irrigated agriculture region, ETn was 617.8 mm in 1980, but decreased by 8.9 mm from 1980 to 2015 due to climate change. The negative impact of climate change was more pronounced in the grassland–irrigated agriculture region, where ETn was 519.5 mm in 1980 and climate change led to a 10.2 mm reduction in evapotranspiration. Figure 9 illustrates the spatial distribution of ETm in 2015, which reflects patterns observed in regions transitioning to rainfed agriculture, characterized by higher values in the north and lower in the south.

4. Discussions

4.1. Spatial Heterogeneity of Anthropogenic Contributions to Evapotranspiration Changes

Agricultural production, as a significant aspect of human activity, is the primary driver of increased evapotranspiration [39]. A comparison of regional ET separation results for the four types of natural-to-farmland conversion in the study area in 2015 (Figure 10) indicates that all conversions, except for forest-to-farmland areas, led to a significant increase in regional ET. As detailed in Table 5, human activities have resulted in an increase in ETt in natural–irrigated agricultural areas by 2.9 mm, 55.9 mm, 29.1 mm, and 30.3 mm, corresponding to growth rates of 0.47%, 10.76%, 5.58%, and 8.41%, respectively. In natural–rainfed agricultural areas, the increases are noted at growth rates of 0.17%, 10.99%, 4.54%, and 2.77%, respectively. Similar results have also been reported in the Heihe River Basin, revealing an increase of 13.6% in ET caused by human activities [39]. The smaller impact in the Songhua River Basin may be attributed to its relatively humid climate [40], in contrast to the arid and semi-arid conditions of the Heihe River Basin, where human activities exert a more pronounced effect on ET [41].
Generally, changes in vegetation types caused by human activities significantly influence evapotranspiration through alterations in canopy structure and root water retention capacity. For example, forests typically have taller and denser canopies, with deep and extensive root systems that not only intercept large amounts of rainfall and prolong water residence time within the system [42] but also enhance the soil’s ability to obtain and retain moisture [43]. These structural characteristics give forests a strong transpiration capacity under favorable water and energy conditions, resulting in high ET levels. In contrast, grasslands have lower canopies and shallower roots, with limited water interception and weaker inherent transpiration capacity [44], leading to lower ET levels [45]. When forests and grassland are converted to farmland, the reduction in canopy height and density in forests, along with root degradation, leads to a decrease in soil water retention and transpiration capacity, making it difficult for farmland ET to reach the original level of forests. In the case of grassland, however, conversion to farmland, along with improved soil water retention and the increased transpiration demand of crops through agricultural management practices, leads to a relative increase in ET [46].
Overall, the impact of vegetation type conversion on evapotranspiration is related not only to meteorological conditions but also to the combined effects of canopy structure and root water retention capacity [47]. This integrated effect reveals significant differences across various vegetation type conversions. Additionally, our study observed instances where ET decreased due to human activities in regions with abundant water and energy, particularly in areas with originally high ET, such as forests and marshlands. To further investigate the reasons for this phenomenon, we used the SOM algorithm to cluster meteorological data from the eight natural-to-agricultural transition regions into nine sub-regions. We analyzed the ET of each sub-region to assess spatial variability in Eth and to explicitly study its influencing factors. Figure 11 illustrates the components of seven meteorological factors across the eight natural-to-agricultural regions. By examining the similarity in the color patterns of each variable, a clear correlation between these parameters can be inferred [48]. For instance, Pres and Shum exhibit a relatively strong similarity across different regions. Additionally, the intensity of each color block can also reflect the weights of variables within this category [49].
In this paper, colors ranging from light yellow to black represent the zonal means of meteorological factors, indicating high to low values. On the right side, we present the sub-region means of ETh results, extracted from each region after clustering and partitioning at a 1 km ×1 km grid scale. Despite the varying distribution of meteorological factors, patterns influencing the distribution of ETh become apparent when comparing regions with very high and very low ETh values. For grassland–rainfed agricultural areas, clusters one and eight belong to the regions with high ETh values, and cluster three corresponds to regions with low ETh values. It is evident that regions with maximum ETh values are characterized by high Wind and Srad, while regions with minimum ETh value are characterized by low Srad and P. For saline-alkali land–rainfed agricultural areas, regions with maximum ETh values are characterized by high Srad and Temp, while regions with the lowest ETh values have significantly lower Wind and Lrad compared to all other sub-regions.
Similar patterns are observed in grassland and saline-alkali land transitioning to irrigated agriculture. In these contexts, areas of maximum value are characterized by high values of P and Wind, while areas of minimum value areas are marked by low P and Wind. This analysis highlights that in agricultural development regions of grassland and saline-alkali land, human activities exert a more pronounced positive impact on evapotranspiration in regions with greater moisture or energy availability [50].
Conversely, in areas converting from forest and marshland to both irrigated and rainfed agriculture, contrasting characteristics emerge. In forest–irrigated agriculture conversion areas, regions with significantly higher ETh were mainly concentrated in areas with low Srad and low Wind. For marshland–irrigated agriculture conversion areas, regions with significantly higher ETh were concentrated in areas with low Srad and low Temp. Notably, regions with high ETh for both forest–rainfed agriculture and marshland–rainfed agriculture contexts are predominantly found in areas with low P and Wind. This phenomenon may be attributed to variations in evapotranspiration across different land use types [51], where forest and marshland exhibit higher ET compared to agricultural land, while grassland and saline-alkali lands demonstrate lower ET than agricultural land [52]. Under conditions of sufficient moisture and energy, the actual ET of vegetation approaches the potential evaporation [38,53]. Therefore, agricultural land has a lower ET compared to the original vegetation type [54], resulting in a decrease in regional ET in the converted areas from natural to agricultural land covers. Consequently, under unfavorable meteorological conditions for evapotranspiration, additional disturbances from human activities may lead to an increase in evapotranspiration beyond that of the original vegetation.
The impact of human activities on evapotranspiration is influenced not only by climate and vegetation type but also by temporal changes and spatial variations in soil properties, particularly when considering different combinations of soil, vegetation, and climate. However, this study primarily focuses on the impact of vegetation type conversion on evapotranspiration, without an in-depth analysis of soil properties. Future research could further explore the effects of soil properties on changes in evapotranspiration.

4.2. Spatial Heterogeneity of Climate Contributions to Evapotranspiration Changes

Understanding the impact of climate change on evapotranspiration is crucial for improving crop management, as it can significantly influence agricultural production [55,56]. Significant climate contributions on ET changes have been reported in previous studies. For instance, [57] quantified the impact of climate and vegetation change on ET change in 110 catchments in southern China. They indicated that climate contributes more than 90% to ET variations in humid regions, whereas vegetation changes account for less than 10%. This study indicates that climate change led to a general increase in ET in 2015 compared to 1980 in the natural areas that transitioned to farmland. In all the studied regions, the climate-driven ET n accounted for 89.7% or higher of the total ET in 2015, which is consistent with the above findings. A cluster analysis of meteorological differences from 1980 to 2015 revealed that ETm was highest in region 1 due to significant increases in Pre and Temp, in the grassland–irrigated agriculture transition region. Conversely, region 3 exhibited the lowest ETm, due to the most significant decrease in both Pre and Temp of this area (Figure 12). In the marshland–irrigated agriculture transition region, the area with the most significant increase in ETm was associated with the largest increases in Srad and Wind, along with the smallest decrease in Pres. Similar characteristics were observed in the areas that transitioned from forest and saline-alkali land to irrigated agriculture. In these regions, the area with the highest increase in ETm was also marked by a significant increase in Temp and Wind. Thus, we can summarize that across the four regions transitioning from natural to irrigated agriculture, energy factors (Lrad, Srad, Temp, Wind and Pres) generally exert a greater influence on the spatial heterogeneity of the ETm than moisture factors (Pre, Shum) in irrigated agriculture; however, the opposite trend was observed in rainfed agriculture.
Regarding both the forest and grassland areas, where land covers were converted to rainfed agriculture, the meteorological elements contributing most significantly to ETm were marked by a significant increase in Pre and a non-significant increase in radiation and Wind. In saline-alkali land–rainfed agriculture regions, significant factors included a marked increase in both Pre and Temp. However, in marshland–rainfed agriculture regions, areas where climate change led to a significant increase in ET were distinguished by the highest increases in Srad and Wind.
These findings suggest that moisture factors are the primary drivers of evapotranspiration in rainfed agriculture regions, whereas energy factors prevail in irrigated agriculture regions. This is because in irrigated agriculture, anthropogenic regulation of moisture conditions ensures a consistent water supply [58], thereby diminishing the variability in the impact of moisture factors (Pre, Shum) on ETm and amplifying the significance of energy factors [59]. Conversely, rainfed agriculture depends on natural moisture conditions that directly influence soil moisture availability, resulting in a substantial effect on ETm [60]. Without artificial moisture management, the variability in moisture factors is heightened, thereby playing a major role in the spatial heterogeneity of ETm. Overall moisture limitation is mitigated only when soil water storage is relatively high [61], shifting the primary control of ETm from moisture factors to energy factors [62].

4.3. Uncertainty in the ET Separation Model

In the early stages of our research, we conducted an investigation of existing ET products. Many researchers are actively comparing different ET products. In this study, we adopted the multi-source ensembled ET product (REA). Cai et al. (2024) conducted a comprehensive comparison of over 90 advanced ET data products, including satellite-based estimates, land surface models, climate models, reanalysis data, machine learning approaches, and ensemble-based estimates, covering the period from 1980 to 2014. Their analysis employed statistical methods to quantify the uncertainty associated with various types of ET products globally, revealing that ensemble ET products exhibited the lowest uncertainty at 4.39%, followed by machine learning ET products at 5.2%, land surface models at 7.07%, and reanalysis products at 13.11% [63]. Additionally, by utilizing 98 ground truth measurements to assess the error of each ET product, they found that both machine learning and multi-source ensemble ET products demonstrated the lowest errors. Concurrently, Wang et al. (2024) and Xiao et al. (2024) compared three long-term series of ET data products in the Yellow River and Tarim River basins, specifically the Global Land Surface Satellite (GLASS) ET, Penman–Monteith–Leuning (PML)-V2 ET, and the Reliability-Economic Average (REA) ET. Their findings indicated that REA exhibited the least uncertainty and error across multiple regions in China [64,65]. Data uncertainty arises from the limitations in spatial and temporal resolution inherent to remote sensing data. While this study utilized a spatial resolution of 1000 m and a temporal resolution of one year, these may be insufficient to fully capture localized meteorological or geographical variations [66]. Additionally, the interpolation of evapotranspiration data from a 0.25 degree grid to a 1000 m grid may have introduced interpolation errors [67].
Additionally, the model presented in this study demonstrates a high simulation accuracy of approximately 0.99, reflecting its exceptional performance in predictive tasks. We also employed the XGBoost algorithm using a gradient boosting approach to validate the results following the same procedure. As shown in Figure 13, the model evaluation metrics indicate that the differences in results produced by different algorithms are not significant (as summarized in Table 6). Furthermore, compared to the evapotranspiration separation method that integrates GRACE data with hydrological models, this study significantly enhances spatial resolution and refines the distinction between human and climatic contributions to total evapotranspiration [24]. Despite the model’s high prediction accuracy, uncertainties persist, particularly in the integration of meteorological and geodemographically remote sensing data for predictive analysis [68]. The uncertainty inherent in the Random Forest model primarily stems from the variability in feature importance and prediction volatility [69]. To assess this, we employed cross-validation and repeated sampling to train the Random Forest model multiple times under varying parameter settings, analyzing the fluctuations in feature importance to ultimately determine the optimal parameter configuration (Table 4). In future work, we may consider integrating algorithms and coupling multiple models to reduce the uncertainty associated with any single model. Additionally, incorporating higher resolution remote sensing data or integrating various data sources in subsequent studies could enhance the accuracy of model predictions.
Finally, our current model is constructed using annual data from different grid points within the study area to simulate evapotranspiration. The focus of our research is on the impacts of long-term climate change and land use on evapotranspiration rates; thus, the evapotranspiration simulation model discussed herein is suitable for long-term scale simulations. However, if we wish to conduct studies with higher temporal resolution, such as examining seasonal variations or different growth stages, our proposed methodology is theoretically applicable. In such cases, it will be necessary to select data from different seasons to build the model.

5. Conclusions

In this study, we propose a framework for the quantitative assessment of climate (ETm) and anthropogenic (ETh) contributions to evapotranspiration changes in the conversion areas of natural to agriculture land use/covers based on the Random Forest algorithm. The impacts of climate change and agricultural development on ET in the Songhua River Basin were evaluated, and SOM clustering analysis was conducted to further explore the spatial heterogeneity of ETm and ETh. The main findings can be summarized as follows:
(1)
The developed models for natural areas (i.e., forest, grassland, marshland, and saline-alkali land) exhibited robust performance in fitting regional ET data, with R2 values reaching approximately 0.99.
(2)
There was a substantial conversion from natural land covers to agricultural covers from 1980 to 2015, with 24,972.32 km2 converted to rainfed agriculture and 4552.1 km2 to irrigated agriculture. In the regions converted to farmland, climate change caused a variation in ET from −3.4 mm to 29.7 mm in rainfed agriculture areas and −8.9 mm to 24.9 mm in irrigated agriculture areas.
(3)
Both the development of rainfed and irrigated agriculture led to an increase in evapotranspiration (ET) in the converted regions, respectively, ranging from 0.9 mm to 53.4 mm, and 2.9 mm to 55.9 mm. It is evident that agricultural development has a greater influence on ET changes than climate change.
(4)
Spatially, significant changes in ET m for irrigated agriculture areas were predominantly located in regions with substantial shifts in energy-related factors (i.e., Wind, Temp, Srad, Lrad, and Pres). Conversely, pronounced changes in ET m in rainfed agriculture areas were concentrated in areas with notable variations in moisture-related factors (i.e., Pre and Shum). This observation highlights the pivotal role of moisture supply in determining ET, which shows shifts to energy factors under adequate moisture conditions.
(5)
Evapotranspiration of farmland falls between that of forest/marshland and grassland/saline-alkali land. In regions with relatively low evapotranspiration capacity, such as grassland and saline-alkali land, agricultural development tends to significantly enhance evapotranspiration. These findings suggest that anthropogenic contributions to changes in ET are closely related to the original vegetation type and prevailing climatic conditions.
The results of this study and the framework proposed can provide mechanistic insights on hydrologic responses to the complex natural and anthropogenic impacts and pave the way for a better understanding of land and water management to better adapt to a changing climate and the growing demand for food.

Author Contributions

Conceptualization, Z.L. and F.L.; methodology, validation, formal analysis, Z.L., F.L. and H.L.; investigation, software, resources, data curation, Z.L., G.Z. and P.Q.; writing—original draft preparation, Z.L.; writing—review and editing, Z.L., F.L., H.L., G.Z. and P.Q.; visualization, Z.L., G.Z. and P.Q.; supervision, F.L.; project administration, G.Z. and P.Q.; funding acquisition, G.Z. and P.Q. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Strategic Priority Research Program of the Chinese Academy of Sciences, China (XDA28020501).

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Acknowledgments

The authors would like to acknowledge the funding support from the Strategic Priority Research Program of the Chinese Academy of Sciences, China (XDA28020501).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Fisher, J.; Melton, F.; Middleton, E.; Hain, C.; Anderson, M.; Allen, R.; McCabe, M.; Hook, S.; Baldocchi, D.; Townsend, P.; et al. The Future of Evapotranspiration: Global Requirements for Ecosystem Functioning, Carbon and Climate Feedbacks, Agricultural Management, and Water Resources. Water Resour. Res. 2017, 53, 2618–2626. [Google Scholar] [CrossRef]
  2. Jin, Z.; Liang, W.; Yang, Y.; Zhang, W.; Yan, J.; Chen, X.; Li, S.; Mo, X. Separating Vegetation Greening and Climate Change Controls on Evapotranspiration Trend over the Loess Plateau. Sci. Rep. 2017, 7, 8191. [Google Scholar] [CrossRef] [PubMed]
  3. Zhang, Y.; Kong, D.; Gan, R.; Chiew, F.; McVicar, T.; Zhang, Q.; Yang, Y. Coupled Estimation of 500 m and 8-Day Resolution Global Evapotranspiration and Gross Primary Production in 2002–2017. Remote Sens. Environ. 2019, 222, 165–182. [Google Scholar] [CrossRef]
  4. Donat, M.G.; Lowry, A.L.; Alexander, L.V.; O’Gorman, P.A.; Maher, N. More Extreme Precipitation in the World’s Dry and Wet Regions. Nat. Clim. Change 2016, 6, 508–513. [Google Scholar] [CrossRef]
  5. Wang, Y.; Ao, Y.; Li, Z. Evapotranspiration Characteristics of Different Oases and Effects of Human Activities on Evapotranspiration in Heihe River Basin. Remote Sens. 2022, 14, 6283. [Google Scholar] [CrossRef]
  6. Yang, Y.; Roderick, M.L.; Guo, H.; Miralles, D.G.; Zhang, L.; Fatichi, S.; Luo, X.; Zhang, Y.; McVicar, T.R.; Tu, Z.; et al. Evapotranspiration on a Greening Earth. Nat. Rev. Earth Environ. 2023, 4, 626–641. [Google Scholar] [CrossRef]
  7. Yuan, X.; Wang, Y.; Ji, P.; Wu, P.; Sheffield, J.; Otkin, J.A. A Global Transition to Flash Droughts under Climate Change. Science 2023, 380, 187–191. [Google Scholar] [CrossRef] [PubMed]
  8. Forootan, E.; Khaki, M.; Schumacher, M.; Wulfmeyer, V.; Mehrnegar, N.; van Dijk, A.; Brocca, L.; Farzaneh, S.; Akinluyi, F.; Ramillien, G.; et al. Understanding the Global Hydrological Droughts of 2003-2016 and Their Relationships with Teleconnections. Sci. Total Environ. 2019, 650, 2587–2604. [Google Scholar] [CrossRef]
  9. Gharbia, S.; Smullen, T.; Gill, L.; Johnston, P.; Pilla, F. Spatially Distributed Potential Evapotranspiration Modeling and Climate Projections. Sci. Total Environ. 2018, 633, 571–592. [Google Scholar] [CrossRef]
  10. Jin, L.; Chen, S.; Yang, H.; Zhang, C. Evaluation and Drivers of Four Evapotranspiration Products in the Yellow River Basin. Remote Sens. 2024, 16, 1829. [Google Scholar] [CrossRef]
  11. Ma, N.; Zhang, Y. Increasing Tibetan Plateau Terrestrial Evapotranspiration Primarily Driven by Precipitation. Agric. For. Meteorol. 2022, 317, 108887. [Google Scholar] [CrossRef]
  12. Tang, Y.; Wang, Z. Derivation of the Relative Contributions of the Climate Change and Human Activities to Mean Annual Streamflow Change. J. Hydrol. 2021, 595, 125740. [Google Scholar] [CrossRef]
  13. Wang, Q.; Jiang, S.; Zhai, J.; He, G.; Zhao, Y.; Zhu, Y.; He, X.; Li, H.; Wang, L.; He, F.; et al. Effects of Vegetation Restoration on Evapotranspiration Water Consumption in Mountainous Areas and Assessment of Its Remaining Restoration Space. J. Hydrol. 2022, 605, 127259. [Google Scholar] [CrossRef]
  14. Zhao, S.; Huang, Y.; Liu, Z.; Liu, T.; Tang, X. Estimation of Actual Evapotranspiration and Water Stress in Typical Irrigation Areas in Xinjiang, Northwest China. Remote Sens. 2024, 16, 2676. [Google Scholar] [CrossRef]
  15. Boé, J. Modulation of the Summer Hydrological Cycle Evolution over Western Europe by Anthropogenic Aerosols and Soil-Atmosphere Interactions. Geophys. Res. Lett. 2016, 43, 7678–7685. [Google Scholar] [CrossRef]
  16. Malhi, G.S.; Kaur, M.; Kaushik, P. Impact of Climate Change on Agriculture and Its Mitigation Strategies: A Review. Sustainability 2021, 13, 1318. [Google Scholar] [CrossRef]
  17. Huang, Y.; Yang, S.; Zhao, H. Distinct Contributions of Climate Change and Anthropogenic Activities to Evapotranspiration and Gross Primary Production Variations over Mainland China. Remote Sens. 2024, 16, 475. [Google Scholar] [CrossRef]
  18. Zeng, R.; Cai, X. Climatic and Terrestrial Storage Control on Evapotranspiration Temporal Variability: Analysis of River Basins around the World. Geophys. Res. Lett. 2016, 43, 185–195. [Google Scholar] [CrossRef]
  19. Shi, S.; Yu, J.; Wang, F.; Wang, P.; Zhang, Y.; Jin, K. Quantitative Contributions of Climate Change and Human Activities to Vegetation Changes over Multiple Time Scales on the Loess Plateau. Sci. Total Environ. 2021, 755, 142419. [Google Scholar] [CrossRef]
  20. Cong, Z.; Shen, Q.; Zhou, L.; Sun, T.; Liu, J. Evapotranspiration Estimation Considering Anthropogenic Heat Based on Remote Sensing in Urban Area. Sci. China-Earth Sci. 2017, 60, 659–671. [Google Scholar] [CrossRef]
  21. Pan, Y.; Zhang, C.; Gong, H.; Yeh, P.; Shen, Y.; Guo, Y.; Huang, Z.; Li, X. Detection of Human-Induced Evapotranspiration Using GRACE Satellite Observations in the Haihe River Basin of China. Geophys. Res. Lett. 2017, 44, 190–199. [Google Scholar] [CrossRef]
  22. Zheng, Y.; Wang, L.; Chen, C.; Fu, Z.; Peng, Z. Using Satellite Gravity and Hydrological Data to Estimate Changes in Evapotranspiration Induced by Water Storage Fluctuations in the Three Gorges Reservoir of China. Remote Sens. 2020, 12, 2143. [Google Scholar] [CrossRef]
  23. Liu, Y.; Mo, X.; Hu, S.; Chen, X.; Liu, S. Assessment of Human-Induced Evapotranspiration with GRACE Satellites in the Ziya-Daqing Basins, China. Hydrol. Sci. J. 2020, 65, 2577–2589. [Google Scholar] [CrossRef]
  24. Zeng, H.; Elnashar, A.; Wu, B.; Zhang, M.; Zhu, W.; Tian, F.; Ma, Z. A Framework for Separating Natural and Anthropogenic Contributions to Evapotranspiration of Human-Managed Land Covers in Watersheds Based on Machine Learning. Sci. Total Environ. 2022, 823, 153726. [Google Scholar] [CrossRef] [PubMed]
  25. Wu, B.; Zeng, H.; Yan, N.; Zhang, M. Approach for Estimating Available Consumable Water for Human Activities in a River Basin. Water Resour. Manag. 2018, 32, 2353–2368. [Google Scholar] [CrossRef]
  26. Shen, C. A Transdisciplinary Review of Deep Learning Research and Its Relevance for Water Resources Scientists. Water Resour. Res. 2018, 54, 8558–8593. [Google Scholar] [CrossRef]
  27. Atiquzzaman, M.; Kandasamy, J. Prediction of Hydrological Time-Series Using Extreme Learning Machine. J. Hydroinformatics 2016, 18, 345–353. [Google Scholar] [CrossRef]
  28. Yan, J.; Jia, S.; Lv, A.; Zhu, W. Water Resources Assessment of China’s Transboundary River Basins Using a Machine Learning Approach. Water Resour. Res. 2019, 55, 632–655. [Google Scholar] [CrossRef]
  29. Liang, Y.; Zhao, P. A Machine Learning Analysis Based on Big Data for Eagle Ford Shale Formation. In Proceedings of the SPE Annual Technical Conference and Exhibition, Calgary, AB, Canada, 30 September–2 October 2019. [Google Scholar]
  30. Alhashem, M. Machine Learning Classification Model for Multiphase Flow Regimes in Horizontal Pipes. In Proceedings of the International Petroleum Technology Conference, Dhahran, Saudi Arabia, 13–15 January 2020. [Google Scholar] [CrossRef]
  31. Allam, M.M.; Figueroa, A.J.; McLaughlin, D.B.; Eltahir, E.A.B. Estimation of Evaporation over the Upper Blue Nile Basin by Combining Observations from Satellites and River Flow Gauges. Water Resour. Res. 2016, 52, 644–659. [Google Scholar] [CrossRef]
  32. Lu, J.; Wang, G.; Chen, T.; Li, S.; Hagan, D.F.T.; Kattel, G.; Peng, J.; Jiang, T.; Su, B. A Harmonized Global Land Evaporation Dataset from Model-Based Products Covering 1980–2017. Earth Syst. Sci. Data 2021, 13, 5879–5898. [Google Scholar] [CrossRef]
  33. Peng, S.; Ding, Y.; Liu, W.; Li, Z. 1 Km Monthly Temperature and Precipitation Dataset for China from 1901 to 2017. Earth Syst. Sci. Data 2019, 11, 1931–1946. [Google Scholar] [CrossRef]
  34. He, J.; Yang, K.; Tang, W.; Lu, H.; Qin, J.; Chen, Y.; Li, X. The First High-Resolution Meteorological Forcing Dataset for Land Process Studies over China. Sci. Data 2020, 7, 25. [Google Scholar] [CrossRef] [PubMed]
  35. Sen, P.K. Estimates of the Regression Coefficient Based on Kendall’s Tau. Publ. Am. Stat. Assoc. 1968, 63, 1379–1389. [Google Scholar] [CrossRef]
  36. Fan, L.; Wang, H.; Wang, C.; Lai, W.; Zhao, Y. Exploration of Use of Copulas in Analysing the Relationship between Precipitation and Meteorological Drought in Beijing, China. Adv. Meteorol. 2017, 2017, 4650284. [Google Scholar] [CrossRef]
  37. Kohonen, T. Essentials of the Self-Organizing Map. Neural Netw. Off. J. Int. Neural Netw. Soc. 2012, 37, 52–65. [Google Scholar] [CrossRef] [PubMed]
  38. Chen, I.T.; Chang, L.C.; Chang, F.J. Exploring the Spatio-Temporal Interrelation between Groundwater and Surface Water by Using the Self-Organizing Maps. J. Hydrol. 2017, 556, 131–142. [Google Scholar] [CrossRef]
  39. Zou, M.; Niu, J.; Kang, S.; Li, X.; Lu, H. The Contribution of Human Agricultural Activities to Increasing Evapotranspiration Is Significantly Greater than Climate Change Effect over Heihe Agricultural Region. Sci. Rep. 2017, 7, 8805. [Google Scholar] [CrossRef]
  40. Shen, Q.; Cong, Z.; Lei, H. Evaluating the Impact of Climate and Underlying Surface Change on Runoff within the Budyko Framework: A Study across 224 Catchments in China. J. Hydrol. 2017, 554, 251–262. [Google Scholar] [CrossRef]
  41. Zhou, Y.; Li, Y.; Li, W.; Li, F.; Xin, Q. Ecological Responses to Climate Change and Human Activities in the Arid and Semi-Arid Regions of Xinjiang in China. Remote Sens. 2022, 14, 3911. [Google Scholar] [CrossRef]
  42. Martin, D.A.; Osen, K.; Grass, I.; Hoelscher, D.; Tscharntke, T.; Wurz, A.; Kreft, H. Land-Use History Determines Ecosystem Services and Conservation Value in Tropical Agroforestry. Conserv. Lett. 2020, 13, e12740. [Google Scholar] [CrossRef]
  43. Zhang, L.; Du, H.; Song, T.; Yang, Z.; Peng, W.; Gong, J.; Huang, G.; Li, Y. Conversion of Farmland to Forest or Grassland Improves Soil Carbon, Nitrogen, and Ecosystem Multi-Functionality in a Subtropical Karst Region of Southwest China. Sci. Rep. 2024, 14, 17745. [Google Scholar] [CrossRef]
  44. Wang, Y.; Shao, M.; Shao, H. A Preliminary Investigation of the Dynamic Characteristics of Dried Soil Layers on the Loess Plateau of China. J. Hydrol. 2010, 381, 9–17. [Google Scholar] [CrossRef]
  45. Zhang, Q.; Lv, X.; Yu, X.; Ni, Y.; Ma, L.; Liu, Z. Species and Spatial Differences in Vegetation Rainfall Interception Capacity: A Synthesis and Meta-Analysis in China. Catena 2022, 213, 106223. [Google Scholar] [CrossRef]
  46. Zhang, Y.; Wang, K.; Wang, J.; Liu, C.; Shangguan, Z. Changes in Soil Water Holding Capacity and Water Availability Following Vegetation Restoration on the Chinese Loess Plateau. Sci. Rep. 2021, 11, 9692. [Google Scholar] [CrossRef] [PubMed]
  47. Zeng, J.; Zhang, Q.; Zhang, Y.; Yue, P.; Yang, Z.; Wang, S.; Zhang, L.; Li, H. Enhanced Impact of Vegetation on Evapotranspiration in the Northern Drought-Prone Belt of China. Remote Sens. 2023, 15, 221. [Google Scholar] [CrossRef]
  48. Zhu, G.; Wu, X.; Ge, J.; Liu, F.; Zhao, W.; Wu, C. Influence of Mining Activities on Groundwater Hydrochemistry and Heavy Metal Migration Using a Self-Organizing Map (SOM). J. Clean. Prod. 2020, 257, 120664. [Google Scholar] [CrossRef]
  49. Haselbeck, V.; Kordilla, J.; Krause, F.; Sauter, M. Self-Organizing Maps for the Identification of Groundwater Salinity Sources Based on Hydrochemical Data. J. Hydrol. 2019, 576, 610–619. [Google Scholar] [CrossRef]
  50. Ma, L.; Yu, G.; Chen, Z.; Yang, M.; Hao, T.; Zhu, X.; Zhang, W.; Lin, Q.; Liu, Z.; Han, L.; et al. Cascade Effects of Climate and Vegetation Influencing the Spatial Variation of Evapotranspiration in China. Agric. For. Meteorol. 2024, 344, 109826. [Google Scholar] [CrossRef]
  51. Shuai, Y.; Tian, Y.; Shao, C.; Huang, J.; Gu, L.; Zhang, Q.; Zhao, R. Potential Variation of Evapotranspiration Induced by Typical Vegetation Changes in Northwest China. Land 2022, 11, 808. [Google Scholar] [CrossRef]
  52. Caballero, C.B.; Ruhoff, A.; Biggs, T. Land Use and Land Cover Changes and Their Impacts on Surface-Atmosphere Interactions in Brazil: A Systematic Review. Sci. Total Environ. 2022, 808, 152134. [Google Scholar] [CrossRef] [PubMed]
  53. Thornthwaite, C.W. An Approach Toward a Rational Classification of Climate. Geogr. Rev. 1948, 38, 55–94. [Google Scholar] [CrossRef]
  54. Liu, J.; You, Y.; Li, J.; Sitch, S.; Gu, X.; Nabel, J.E.M.S.; Lombardozzi, D.; Luo, M.; Feng, X.; Arneth, A.; et al. Response of Global Land Evapotranspiration to Climate Change, Elevated CO2, and Land Use Change. Agric. For. Meteorol. 2021, 311, 108663. [Google Scholar] [CrossRef]
  55. Benitez-Alfonso, Y.; Soanes, B.K.; Zimba, S.; Sinanaj, B.; German, L.; Sharma, V.; Bohra, A.; Kolesnikova, A.; Dunn, J.A.; Martin, A.C.; et al. Enhancing Climate Change Resilience in Agricultural Crops. Curr. Biol. 2023, 33, R1246–R1261. [Google Scholar] [CrossRef]
  56. Nam, W.-H.; Hong, E.-M.; Choi, J.-Y. Has Climate Change Already Affected the Spatial Distribution and Temporal Trends of Reference Evapotranspiration in South Korea? Agric. Water Manag. 2015, 150, 129–138. [Google Scholar] [CrossRef]
  57. Zhang, D.; Liu, X.; Zhang, L.; Zhang, Q.; Gan, R.; Li, X. Attribution of Evapotranspiration Changes in Humid Regions of China from 1982 to 2016. JGR Atmos. 2020, 125, e2020JD032404. [Google Scholar] [CrossRef]
  58. Rosa, L. Adapting Agriculture to Climate Change via Sustainable Irrigation: Biophysical Potentials and Feedbacks. Environ. Res. Lett. 2022, 17, 063008. [Google Scholar] [CrossRef]
  59. Yu, H.; Yang, T.; Li, S.; Kang, S.; Du, T.; Wang, Y.; Chen, H.; Guo, H. Surface Energy Fluxes in a Drip-Irrigated Agroecosystem: Unique Advection Effect of Oasis. Agric. For. Meteorol. 2024, 357, 110204. [Google Scholar] [CrossRef]
  60. Chen, X.; Yu, Y.; Chen, J.; Zhang, T.; Li, Z. Seasonal and Interannual Variation of Radiation and Energy Fluxes over a Rain-Fed Cropland in the Semi-Arid Area of Loess Plateau, Northwestern China. Atmos. Res. 2016, 176–177, 240–253. [Google Scholar] [CrossRef]
  61. Denissen, J.M.C.; Teuling, A.J.; Pitman, A.J.; Koirala, S.; Migliavacca, M.; Li, W.; Reichstein, M.; Winkler, A.J.; Zhan, C.; Orth, R. Widespread Shift from Ecosystem Energy to Water Limitation with Climate Change. Nat. Clim. Change 2022, 12, 677–684. [Google Scholar] [CrossRef]
  62. Chen, Q.-W.; Liu, M.-J.; Lyu, J.; Li, G.; Otsuki, K.; Yamanaka, N.; Du, S. Characterization of Dominant Factors on Evapotranspiration with Seasonal Soil Water Changes in Two Adjacent Forests in the Semiarid Loess Plateau. J. Hydrol. 2022, 613, 128427. [Google Scholar] [CrossRef]
  63. Cai, Y.; Xu, Q.; Bai, F.; Cao, X.; Wei, Z.; Lu, X.; Wei, N.; Yuan, H.; Zhang, S.; Liu, S.; et al. Reconciling Global Terrestrial Evapotranspiration Estimates from Multi-Product Intercomparison and Evaluation. Water Resour. Res. 2024, 60, e2024WR037608. [Google Scholar] [CrossRef]
  64. Xiao, J.; Sun, F.; Wang, T.; Wang, H. Estimation and Validation of High-Resolution Evapotranspiration Products for an Arid River Basin Using Multi-Source Remote Sensing Data. Agric. Water Manag. 2024, 298, 108864. [Google Scholar] [CrossRef]
  65. Wang, R.; You, X.; Shi, Y.; Wu, C. Enhancing Evapotranspiration Estimations through Multi-Source Product Fusion in the Yellow River Basin, China. Water 2024, 16, 2603. [Google Scholar] [CrossRef]
  66. Sillmann, J.; Thorarinsdottir, T.; Keenlyside, N.; Schaller, N.; Alexander, L.V.; Hegerl, G.; Seneviratne, S.I.; Vautard, R.; Zhang, X.; Zwiers, F.W. Understanding, Modeling and Predicting Weather and Climate Extremes: Challenges and Opportunities. Weather Clim. Extrem. 2017, 18, 65–74. [Google Scholar] [CrossRef]
  67. Băicoianu, A.; Gavrilă, C.G.; Păcurar, C.M.; Păcurar, V.D. Fractal Interpolation in the Context of Prediction Accuracy Optimization. Eng. Appl. Artif. Intell. 2024, 133, 108380. [Google Scholar] [CrossRef]
  68. Xiong, R.; Shi, Y.; Jing, H.; Liang, W.; Nakahira, Y.; Tang, P. Calibrating Subjective Data Biases and Model Predictive Uncertainties in Machine Learning-Based Thermal Perception Predictions. Build. Environ. 2024, 247, 111053. [Google Scholar] [CrossRef]
  69. Coulston, J.W.; Blinn, C.E.; Thomas, V.A.; Wynne, R.H. Approximating Prediction Uncertainty for Random Forest Regression Models. Photogramm. Eng. Remote Sens. 2016, 82, 189–197. [Google Scholar] [CrossRef]
Remotesensing 16 04408 g001
Figure 1. The geographical location of the Songhua River Basin.
Figure 1. The geographical location of the Songhua River Basin.
Remotesensing 16 04408 g001
Remotesensing 16 04408 g002
Figure 2. The flowchart of Random Forest Regression.
Figure 2. The flowchart of Random Forest Regression.
Remotesensing 16 04408 g002
Remotesensing 16 04408 g003
Figure 3. Conceptual framework of the method for quantifying evapotranspiration influenced by climate change and that influenced by human activities of human-managed land cover types (taking the rainfed agricultural transition areas in saline-alkali land as an example).
Figure 3. Conceptual framework of the method for quantifying evapotranspiration influenced by climate change and that influenced by human activities of human-managed land cover types (taking the rainfed agricultural transition areas in saline-alkali land as an example).
Remotesensing 16 04408 g003
Remotesensing 16 04408 g004
Figure 4. Land cover change in the study area: (a) Songhua River Basin in the 1980s, (b) Songhua River Basin in the 2015s and (c) Land use transfer contribution. F, forest area; G, grassland area; M, marshland area; SA, saline-alkali land; R, rainfed agriculture; I, irrigated agriculture; S, settlement; W, water.
Figure 4. Land cover change in the study area: (a) Songhua River Basin in the 1980s, (b) Songhua River Basin in the 2015s and (c) Land use transfer contribution. F, forest area; G, grassland area; M, marshland area; SA, saline-alkali land; R, rainfed agriculture; I, irrigated agriculture; S, settlement; W, water.
Remotesensing 16 04408 g004
Remotesensing 16 04408 g005
Figure 5. Annual anomaly and cumulative anomaly of evapotranspiration (ET) for forest (a), grassland (b), marshland (c), and saline-alkali land (d) in the Songhua River Basin from 1980 to 2015, along with the spatial distribution of both the average annual ET (e–h) and its changing trends (i–l) across the basin.
Figure 5. Annual anomaly and cumulative anomaly of evapotranspiration (ET) for forest (a), grassland (b), marshland (c), and saline-alkali land (d) in the Songhua River Basin from 1980 to 2015, along with the spatial distribution of both the average annual ET (e–h) and its changing trends (i–l) across the basin.
Remotesensing 16 04408 g005
Remotesensing 16 04408 g006
Figure 6. Cross-validation of ETn prediction for the four types of natural areas.
Figure 6. Cross-validation of ETn prediction for the four types of natural areas.
Remotesensing 16 04408 g006
Remotesensing 16 04408 g007
Figure 7. The importance of the variables for four regional ETn prediction models.
Figure 7. The importance of the variables for four regional ETn prediction models.
Remotesensing 16 04408 g007
Remotesensing 16 04408 g008
Figure 8. Spatial distribution of ETm and ETh in the natural (forest, grassland, marshland, and SA) to rainfed agriculture areas from 1980 to 2015. SA is short for saline-alkali land.
Figure 8. Spatial distribution of ETm and ETh in the natural (forest, grassland, marshland, and SA) to rainfed agriculture areas from 1980 to 2015. SA is short for saline-alkali land.
Remotesensing 16 04408 g008
Remotesensing 16 04408 g009
Figure 9. Spatial distribution of ETm and ETh in the natural (forest, grassland, marshland, and SA) to irrigated agriculture areas from 1980 to 2015. SA is short for saline-alkali land.
Figure 9. Spatial distribution of ETm and ETh in the natural (forest, grassland, marshland, and SA) to irrigated agriculture areas from 1980 to 2015. SA is short for saline-alkali land.
Remotesensing 16 04408 g009
Remotesensing 16 04408 g010
Figure 10. Climate and anthropogenic contributions to evapotranspiration changes from 1980 to 2015 in the natural to agricultural areas.
Figure 10. Climate and anthropogenic contributions to evapotranspiration changes from 1980 to 2015 in the natural to agricultural areas.
Remotesensing 16 04408 g010
Remotesensing 16 04408 g011
Figure 11. Component planes of the seven training parameters in the SOM of the ecological–agricultural transformation region and average ET h in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).
Figure 11. Component planes of the seven training parameters in the SOM of the ecological–agricultural transformation region and average ET h in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).
Remotesensing 16 04408 g011
Remotesensing 16 04408 g012
Figure 12. Component planes of the seven training parameters in the SOM of the natural–agricultural transformation region and average ET m in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).
Figure 12. Component planes of the seven training parameters in the SOM of the natural–agricultural transformation region and average ET m in the nine allocated areas. (IA: irrigated agriculture; RA: rainfed agriculture; SA: saline-alkali land; Numbers 1 to 9 indicate the sub-regions with different meteorological conditions obtained through clustering using the SOM algorithm in each natural–agricultural transformation region).
Remotesensing 16 04408 g012
Remotesensing 16 04408 g013
Figure 13. Cross-validation of ETn prediction for the four types of natural areas based on XGBoost algorithm.
Figure 13. Cross-validation of ETn prediction for the four types of natural areas based on XGBoost algorithm.
Remotesensing 16 04408 g013
Table 1. Criterion for trend analysis.
Table 1. Criterion for trend analysis.
β ValueZ ValueTrend
β > 0Z > 2.58Highly significant increase
1.96 < Z ≤ 2.58Significant increase
Z ≤ 1.96Slight increase
β = 0Z∈Q Unchanged
β < 0Z ≤ 1.96Slight reduction
1.96 < Z ≤ 2.58Significant reduction
Z > 2.58Highly significant reduction
Note: Q stands for rational number.
Table 2. The transferred area among different land use types in the Songhua River Basin from 1980 to 2015 (km2).
Table 2. The transferred area among different land use types in the Songhua River Basin from 1980 to 2015 (km2).
2015ForestGrasslandMarshlandSaline-Alkali LandIrrigated
Agriculture		Rainfed
Agriculture		SettlementWater
1980
Forest214,836.80 4013.69 207.02 62.60 423.94 9818.00 160.42 126.90
Grassland2295.84 59,138.95 754.68 1573.33 1553.00 12,078.19 335.18 272.34
Marshland184.11 907.27 20,960.36 352.62 2495.31 2047.56 83.79 341.01
Saline-alkali land85.341009.46182.289475.6479.85428.5772.8744.02
Irrigated
agriculture		17.89 50.58 143.55 1.45 15,637.93 1771.06 85.81 59.89
Rainfed
agriculture		1125.04 785.28 269.10 97.49 8471.83 150,241.77 1312.20 291.33
Settlement4.39 9.09 1.82 0.97 62.02 83.66 12,535.78 14.55
Water42.88 363.54 439.18 534.21 400.38 548.47 19.46 13,789.24
Table 3. Percentage of area that have different changing trends in ET for the four natural regions in the Songhua River Basin.
Table 3. Percentage of area that have different changing trends in ET for the four natural regions in the Songhua River Basin.
Land Use TypesPercentage of the Total Area (%)
Significant ReductionSlight ReductionSlight IncreaseSignificant IncreaseHighly Significant Increase
Forest0.368.9929.6810.03
Grassland0.0448.0445.75.011.21
Marshland037.9751.258.652.13
Saline-alkali land0.035.8572.0618.563.5
Table 4. Parameters of the ETn models for forest, grassland, marshland, and saline-alkali land, respectively.
Table 4. Parameters of the ETn models for forest, grassland, marshland, and saline-alkali land, respectively.
ParametersETn Model
ForestGrasslandMarshlandSaline-Alkali Land
n_tree300300300300
m_try7776
max. depth1010107
min. node. size30303030
Table 5. Evapotranspiration in the natural to agricultural areas in 1980 and 2015 and the climate and anthropogenic contributions to the changes in ET.
Table 5. Evapotranspiration in the natural to agricultural areas in 1980 and 2015 and the climate and anthropogenic contributions to the changes in ET.
LULCAreaET1980ET2015 ET m ET n ET h
km2mmmmmm106 m3mm106 m3mm106 m3
Forest–irrigated agriculture423.94617.8611.8−8.93.8608.9257.62.91.2
Forest–rainfed agriculture9818528.3548.018.8184.6547.15471.40.98.8
Grassland–irrigated agriculture1553519.5565.2−10.215.8509.3790.955.986.8
Grassland–rainfed agriculture12,078.19485.8535.9−3.441.1482.48255826.553.4645
Marshland–irrigated agriculture2495.31521.6572.521.823.9543.41355.829.172.6
Marshland–rainfed agriculture2047.56502.6555.129.760.9532.31090.122.846.7
SA–irrigated agriculture79.85461.4516.724.92486.338.830.32.4
SA–rainfed agriculture428.57476510.120.99496.921313.25.6
Table 6. Comparison of predicted evapotranspiration (ETn) Results from RFR and XGBoost algorithms.
Table 6. Comparison of predicted evapotranspiration (ETn) Results from RFR and XGBoost algorithms.
LULCETn (mm)Relative Error (%)
RFRXGBoost
Forest–irrigated agriculture608.9607.10.3
Forest–rainfed agriculture547.1544.90.4
Grassland–irrigated agriculture509.3511.4−0.41
Grassland–rainfed agriculture482.5483.1−0.12
Marshland–irrigated agriculture543.4540.20.59
Marshland–rainfed agriculture532.3531.20.2
SA–irrigated agriculture486.3483.90.49
SA–rainfed agriculture496.9493.40.7
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Liang, Z.; Li, F.; Li, H.; Zhang, G.; Qi, P. A Framework for Separating Climate and Anthropogenic Contributions to Evapotranspiration Changes in Natural to Agricultural Regions of Watersheds Based on Machine Learning. Remote Sens. 2024, 16, 4408. https://doi.org/10.3390/rs16234408

AMA Style

Liang Z, Li F, Li H, Zhang G, Qi P. A Framework for Separating Climate and Anthropogenic Contributions to Evapotranspiration Changes in Natural to Agricultural Regions of Watersheds Based on Machine Learning. Remote Sensing. 2024; 16(23):4408. https://doi.org/10.3390/rs16234408

Chicago/Turabian Style

Liang, Zixin, Fengping Li, Hongyan Li, Guangxin Zhang, and Peng Qi. 2024. "A Framework for Separating Climate and Anthropogenic Contributions to Evapotranspiration Changes in Natural to Agricultural Regions of Watersheds Based on Machine Learning" Remote Sensing 16, no. 23: 4408. https://doi.org/10.3390/rs16234408

APA Style

Liang, Z., Li, F., Li, H., Zhang, G., & Qi, P. (2024). A Framework for Separating Climate and Anthropogenic Contributions to Evapotranspiration Changes in Natural to Agricultural Regions of Watersheds Based on Machine Learning. Remote Sensing, 16(23), 4408. https://doi.org/10.3390/rs16234408

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop