Coastal Forest Cover Change Detection Using Satellite Images and Convolutional Neural Networks in Vietnam
Coastal Forest Cover Change Detection Using Satellite Images and Convolutional Neural Networks in Vietnam
Coastal Forest Cover Change Detection Using Satellite Images and Convolutional Neural Networks in Vietnam
Corresponding Author:
Khanh Nguyen-Trong
Faculty of Information Technology, Posts and Telecommunications Institute of Technology
Km10 Nguyen Trai, Hanoi, Vietnam
Email: khanhnt@ptit.edu.vn
1. INTRODUCTION
Coastal forests are an important part of tropical biodiversity. They provide a lot of important
services for our ecosystem, such as extreme weather protection, erosion prevention, environments for
different species, and storage of blue carbon, which allow to mitigate climate change [1]. However, these
forests are increasingly vulnerable to degradation as a result of climate change, sea- level rise or
anthropogenic processes such as deforestation [2]. To address these issues, accurate and automated forest
cover monitoring is crucial [3]. In this context, high-resolution remote-sensing images collected from
satellites, such as the European Sentinel-2A, -2B, or LandSat-8, offer potential and cost-efficient sources for
an automatic solution [4].
Most previous studies focused on traditional methods using hand-crafted features, such as
multi-variant change vector analysis (MVCA), normalized difference vegetation index (NVDI), and so on
[1], [2], [5]–[9]. They have drawbacks that prevent their wide application, especially for non-domain experts
in forestry and remote sensing technology. On the one hand, they require more effort and time due to the
excessive dependence on handcrafted features. On the other hand, they are ad-hoc solutions that are suitable
only for specific regions. Therefore, these methods are time-consuming and inefficient.
Recently, with the development of deep learning technology, the field of object detection in remote
sensing images has made significant progress. Deep neural networks allow an automatic feature extraction,
avoid feature selection and reduce manual steps in monitoring forest cover change [10]–[12]. Convolutional
neural networks (CNNs) are one of the well-known deep learning algorithms that have been widely used in
remote sensing image classification. They allow us to extract more meaningful features, the classification of
these images usually results in higher performance [13].
For example, de Bem et al. [10] presented a method that used CNN and Landsat data for
deforestation detection in the Brazilian Amazon. The authors applied three CNN architectures including
U-Net, ResUnet, and SharpMask to classify the change between the years of 2017-2018 and 2018-2019. The
experiment results show that the network achieved a high accuracy, without any post-processing for noise
cancelling. Stoian et al. [11] also proposed an application of CNN to build land cover maps using high-
resolution satellite image time series. Based on data from Sentinel-2 L2A, the U-Net network was applied in
this study to deal with sparse annotation data while maintaining high-resolution output. Such networks are
even applicable with incomplete satellite imagery in similar problems. For instance, Khan et al. [14] detected
forest cover changes over 29 years (1987–2015), in which the authors faced issues of incomplete and noisy
data. By using a deep CNN network, they mapped the raw data to more separable features. These features
were employed to detect the changes. Many similar applications can be found in the literature, such as the
works of [12], [15]–[20] and so on.
We are interested in monitoring forest cover change using deep learning. Numerous works, which
applied deep neural networks, such as CNN, U-Net, and satellite images to detect forest loss areas, have been
proposed worldwide [10], [11], [14], [21], [22]. However, in Vietnam, traditional machine learning is still
widely used. In this paper, we proposed a method for coastal forest cover change detection in Vietnam. Based
on sensing images from the European Sentinel-2A and -B, we trained a U-Net model to detect forest and
non- forest areas. We then combined geographic information systems (GIS) information to compute the
forest cover changes and evaluated results with available information from the national forest monitoring
system. The proposed method is capable of applying to different areas, with less effort from domain experts.
The paper is organized as: section 2 introduces our research method. Section 3 presents the
experimental results and discussion. Lastly, section 4 concludes the work conducted and proposes some
future works.
2. RESEARCH METHOD
2.1. Method overview
The main objective of this study is to automatically detect and calculate coastal forest cover changes
of Hai Phong city, Northern Vietnam. We performed pixel-level semantic segmentation on Sentinel-2A and
2B images, to classify forest and non-forest areas. These images were chosen from the same areas between
two periods times. Therefore, combining with GIS information, we can detect and calculate forest cover
changes. To do so, we realized three big steps, as presented in Figure 1, including:
− Model training: in this first step, we trained a semantic segmentation model, which was based on U-Net
neural network. The training dataset came from the Sentinel-2. We also evaluated the trained model
using the forest cover layers extracted from the national forest monitoring system (FRMS) of Vietnam.
− Forest cover detection: after training, we applied the model to classify forest and non-forest areas of
images of the same location taken in different times.
− Cover change analysis: lastly, based on forest covers at the two different times, we detected and
calculated changes.
The model training consists of data preparation, model training, and testing. While the last two steps
are composed of model using and GIS analysis. At each of these steps, we applied related techniques in deep
learning, satellite image processing, GIS, and so on. The following sections will detail these steps.
Coastal forest cover change detection using satellite images and … (Khanh Nguyen-Trong)
932 ISSN: 2252-8938
The model was based on U-Net network architecture [24]. We used satellite images for model
training. While, the information extracted from FMRS (forest cover layers), combined with GIS, was
employed for model evaluation. The following section will detail the data preparation step.
− Scene detection: we selected only image scenes at the coastal and mangrove forest of Hai Phong city.
Then, we filtered and kept only images captured in 2018 and 2019. Lastly, the images with cloud rate
greater than 30% were eliminated. After this step, we obtained 26 and 32 images captured in 2018 and
2019 respectively.
− Band selection: sentinel 2 have 13 spectral bands, with different bandwidth and spatial resolution. In
this study, we directly used ten bands for input features, including the bands from 2 to 8, 8A, 11 and 12
with wavelength of 0.490 µm, 0.560 µm, 0.665 µm, 0.705 µm, 0.740 µm, 0.783 µm, 0.842 µm,
0.865 µm, 1.610 µm, and 2.190 µm. The bands 1, 9, and 10 were ignored because they are not relevant
to vegetation [26]. Moreover, we also computed three indices: normalized difference vegetation index
(NDVI), normalized difference snow index (NDSI), normalized difference water index (NDWI), which
are widely applied in similar problems. They are computed as in (1).
𝑁𝐼𝑅−𝑅𝑒𝑑 𝐺𝑟𝑒𝑒𝑛−𝑆𝑊𝐼𝑅 𝑁𝐼𝑅−𝑆𝑊𝐼𝑅
𝑁𝐷𝑉𝐼 = ; 𝑁𝐷𝑆𝐼 = ; 𝑁𝐷𝑊𝐼 = (1)
𝑁𝐼𝑅+𝑅𝑒𝑑 𝐺𝑟𝑒𝑒𝑛+𝑆𝑊𝐼𝑅 𝑁𝐼𝑅+𝑆𝑊𝐼𝑅
where near infrared reflectance (NIR) is band 8, Red is band 4, Green is band 3, and short-wave infrared
(SWIR) is band 11.
− Cloud free: we removed the cloud using the QA60 band, which is a bitmask band with cloud mask
information [27]. Since bits 10 and 11 specify clouds and cirrus, we could filter all cloudy pixels.
Figure 4(a) and Figure 4(b) show an example of selected images before and after cloud-free.
− Median value calculation: to improve the quality of images, we applied a median filter that moves
through the image pixel by pixel, and replaced each value with the median value of neighboring ones.
− Image cropping: at this step, we cropped images to focus only on studied areas.
(a) (b)
Figure 4. Satellite images before (a) and after (b) cloud free
These scene images were then combined with forest cover layers extracted from FRMS to build a
labeled pixel-level dataset for model training, as shown in the bottom process in Figure 3. We extracted four
important pieces of information from FRMS, including administrative information, coordinates, forest
observations (0 for non-forest, 1 for forest), detailed plot information. Since the forest cover layers were
manually entered to FRMS by local rangers, therefore we conducted several field trips to verify the ground-
truth labels. Based on the available resources, we selected a number of sample points to manually check if the
information is correct (forest or non-forest). After verification, we got 1,500 sample points with correct
labels. Centering on these points, two corresponding neighborhood patches were created, including i) image
patches of size 256×256×13 from cropped satellite images and ii) forest cover layer of size 256×256, as
presented in Figure 3. Finally, we obtained a dataset of size 256×256×14 for model training.
(a) (b)
Figure 7. Progress of accuracy (a) and loss (b) on the training and validation set
To detect forest cover changes, we applied the trained model to images of the same location taken in
2018 and 2019. The obtained results were then used to detect and calculate forest cover changes, as shown in
Figure 8. The model accurately detected forest areas at the beginning and ending period (2018 - Figure 8(a),
and 2019 - Figure 8(b)). Then, we mapped the two results and performed several GIS operations to get the
forest cover changes, as detailed in Figure 8(c). According to the policy of the Vietnam government, an
increase or decrease of forest covers, which is greater than 0.3 ha, will be considered to be a change.
Therefore, we calculated and detected five forest loss areas, as presented in Figure 8(d) (red parts). The
results were similar to those reported by local rangers in 2019. Therefore, our model is capable of accurately
detecting forest cover changes.
Compared with existing methods that are widely applied in Vietnam, our proposed method is more
robust and more accurate in forest cover detection. Experimental results show that our method outperforms
MVCA by 3.8% (91.6% on the testing set). It allows detecting a higher level of forest disturbances, as shown
in Figure 9. Figure 9(a) and Figure 9(b) shows forest cover changes (the white part and yellow part),
predicted by MVCA and our proposed method, respectively. Our method produced results that are closer to
the real data reported by local rangers.
Coastal forest cover change detection using satellite images and … (Khanh Nguyen-Trong)
936 ISSN: 2252-8938
Moreover, the proposed method requires less expert knowledge than methods based on NVDI,
NVSI as in [1], [2], [7], [9], [29]. For these methods, domain experts are highly required to determine
threshold values that are applicable only for a specific area. In contrast, our method does not require these
thresholds. The model automatically learns useful features from input data to detect forest cover.
Furthermore, as a deep learning model, our model can be incrementally trained with the new target areas. It
means that the model can be gradually provided with new samples to update its weights and thus improve its
classifications with time.
Figure 8. Comparing forest cover (dark green part) in (a) 2018 and (b) 2019 to compute (c) all forest changes
(yellow part) and (d) the ones greater than 0.3 ha (red part)
Figure 9. Forest and forest plot cover change prediction by (a) MVCA (green: forest covers; white: forest
cover change) and (b) U-Net (green: forest covers; yellow: forest cover change)
Despite their advantages, the proposed method is not as easy to implement as MVCA and similar
methods based on thresholds. It requires, on the one hand, a relatively large quantity of samples, and on the
other hand, ground truth masks that can be challenging and time-consuming. Whereas the MCVA and similar
methods work with simpler sampling schemes and can produce reasonably acceptable results. However, in
Vietnam thanks to FRMS, we already have ground-truth labels that are regularly entered by local rangers.
Therefore, the proposed method is able to be widely applied for automatically monitoring forest covers.
4. CONCLUSION
In this study, a deep learning based-method for coastal forest cover change detection has been pro-
posed. We used multi-temporal Sentinel-2 imagery to train a segmentation model based on U-Net neural net-
work. Furthermore, we evaluated the model with forest cover information extracted from the national forest
resource monitoring system of Vietnam. The results shown that our method achieved a good performance on
remote sensing images. The trained model achieved a high accuracy of 95.4% on the testing set and
outperformed the popular methods based on thresholds in Vietnam. Future works will focus on tree species
classification by improving the network architecture, increasing our dataset and proposing augmentation
methods for forest cover images.
REFERENCES
[1] H. T. T. Nguyen et al., “Mangrove forest landcover changes in coastal vietnam: a case study from 1973 to 2020 in Thanh Hoa and
Nghe An provinces,” Forests, vol. 12, no. 5, May 2021, doi: 10.3390/f12050637.
[2] H.-H. Nguyen et al., “Monitoring changes in coastal mangrove extents using multi-temporal satellite data in selected communes,
Hai Phong City, Vietnam,” Forest and Society, vol. 4, no. 1, Apr. 2020, doi: 10.24259/fs.v4i1.8486.
[3] B. K. Kogo, L. Kumar, and R. Koech, “Forest cover dynamics and underlying driving forces affecting ecosystem services in
western Kenya,” Remote Sensing Applications: Society and Environment, vol. 14, pp. 75–83, Apr. 2019,
doi: 10.1016/j.rsase.2019.02.007.
[4] M. Jaskulak and A. Grobelak, “Forest degradation prevention,” in Handbook of Ecological and Ecosystem Engineering, Wiley,
2021, pp. 377–390., doi: 10.1002/9781119678595.ch20.
[5] E. Gyamfi-Ampadu, M. Gebreslasie, and A. Mendoza-Ponce, “Multi-decadal spatial and temporal forest cover change analysis of
Nkandla Natural Reserve, South Africa,” Journal of Sustainable Forestry, pp. 1–24, Feb. 2021,
doi: 10.1080/10549811.2021.1891441.
[6] T. V. Ramachandra, B. Setturu, and S. Vinay, “Assessment of forest transitions and regions of conservation importance in Udupi
district, Karnataka,” Indian Forester, vol. 147, no. 9, Oct. 2021, doi: 10.36808/if/2021/v147i9/164166.
[7] T. H. Nguyen, V. D. Pham, S. D. Le, and V. D. Nguyen, “Determining the locations of deforestation using multi-variant change
vector analysis (MCVA) on Landsat-8 satellite data,” Journal of Forestry Science and Technology, vol. 4, no. 96–105, 2017
[8] D. Tien Bui, K.-T. T. Le, V. C. Nguyen, H. D. Le, and I. Revhaug, “Tropical forest fire susceptibility mapping at the cat ba
national park area, hai phong city, vietnam, using gis-based kernel logistic regression,” Remote Sensing, vol. 8, no. 4, 2016.
doi: 10.3390/rs8040347.
[9] N. Van Thi, T. Q. Bao, L. S. Doanh, P. Van Duan, N. N. Hai, and T. X. Hoa, “Study on deforestation detection in Gia Lai
province using sentinel-2 optical satellite image and sentinel-1 radar data,” Journal of Vietnam Agricultural Science and
Technology, vol. 5, pp. 105–112, 2020
[10] P. P. de Bem, O. A. de Carvalho, R. F. Guimarães, and R. A. T. Gomes, “Change detection of deforestation in the brazilian
amazon using landsat data and convolutional neural networks,” Remote Sensing, vol. 12, no. 6, Mar. 2020,
doi: 10.3390/rs12060901.
[11] A. Stoian, V. Poulain, J. Inglada, V. Poughon, and D. Derksen, “Land cover maps production with high resolution satellite image
time series and convolutional neural networks: adaptations and limits for operational systems,” Remote Sensing, vol. 11, no. 17,
Aug. 2019, doi: 10.3390/rs11171986.
[12] T. Chang, B. Rasmussen, B. Dickson, and L. Zachmann, “Chimera: a multi-task recurrent convolutional neural network for forest
classification and structural estimation,” Remote Sensing, vol. 11, no. 7, Mar. 2019, doi: 10.3390/rs11070768.
[13] C. Shi, T. Wang, and L. Wang, “Branch feature fusion convolution network for remote sensing scene classification,” IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 13, pp. 5194–5210, 2020,
doi: 10.1109/JSTARS.2020.3018307.
[14] S. H. Khan, X. He, F. Porikli, and M. Bennamoun, “Forest change detection in incomplete satellite images with deep neural
networks,” IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 9, pp. 5407–5423, Sep. 2017,
doi: 10.1109/TGRS.2017.2707528.
[15] M. Pritt and G. Chern, “Satellite image classification with deep learning,” in 2017 IEEE Applied Imagery Pattern Recognition
Workshop (AIPR), Oct. 2017, pp. 1–7., doi: 10.1109/AIPR.2017.8457969.
[16] K. Hufkens et al., “Historical aerial surveys map long-term changes of forest cover and structure in the Central Congo Basin,”
Remote Sensing, vol. 12, no. 4, Feb. 2020, doi: 10.3390/rs12040638.
[17] L. Bragagnolo, R. V. da Silva, and J. M. V. Grzybowski, “Amazon forest cover change mapping based on semantic segmentation
by u-nets,” Ecological Informatics, vol. 62, May 2021, doi: 10.1016/j.ecoinf.2021.101279.
[18] R. Senthilkumar, V. Srinidhi, S. Neelavathi, and S. Renuga Devi, “Forest change detection using an optimized convolution neural
network,” IETE Technical Review, pp. 1–8, Oct. 2020, doi: 10.1080/02564602.2020.1827987.
[19] Z. Mohammed, C. Hanae, and S. Larbi, “Comparative study on machine learning algorithms for early fire forest detection system
using geodata,” International Journal of Electrical and Computer Engineering (IJECE), vol. 10, no. 5, pp. 5507–5513, Oct. 2020,
doi: 10.11591/ijece.v10i5.pp5507-5513.
[20] A. AlDabbas and Z. Gal, “Cassini-Huygens mission images classification framework by deep learning advanced approach,”
International Journal of Electrical and Computer Engineering (IJECE), vol. 11, no. 3, pp. 2457–2466, Jun. 2021,
doi: 10.11591/ijece.v11i3.pp2457-2466.
Coastal forest cover change detection using satellite images and … (Khanh Nguyen-Trong)
938 ISSN: 2252-8938
[21] D. Peng, Y. Zhang, and H. Guan, “End-to-end change detection for high resolution satellite images using improved UNet++,”
Remote Sensing, vol. 11, no. 11, Jun. 2019, doi: 10.3390/rs11111382.
[22] J. V. Solórzano, J. F. Mas, Y. Gao, and J. A. Gallardo-Cruz, “Land use land cover classification with u-net: advantages of
combining sentinel-1 and sentinel-2 imagery,” Remote Sensing, vol. 13, no. 18, Sep. 2021, doi: 10.3390/rs13183600.
[23] “Forest management information system in Viet Nam ensures sustainability by providing up-to-date information,” Niras, 2018.
https://www.niras.com/projects/formis-ii/ (accessed Mar. 14, 2022)
[24] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: vonvolutional networks for biomedical image segmentation,” in Lecture Notes
in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 9351,
Springer International Publishing, 2015, pp. 234–241., doi: 10.1007/978-3-319-24574-4_28.
[25] “Sentinel-2 msi: multispectral instrument, level-2a,” Google. https://developers.google.com/earth-engine/
datasets/catalog/COPERNICUS_S2_SR (accessed Mar. 14, 2022)
[26] X. Zhou et al., “Discriminating urban forest types from sentinel-2A image data through linear spectral mixture analysis: a case
study of Xuzhou, East China,” Forests, vol. 10, no. 6, May 2019, doi: 10.3390/f10060478.
[27] K. Soudani et al., “Potential of C-band synthetic aperture radar sentinel-1 time-series for the monitoring of phenological cycles in
a deciduous forest,” International Journal of Applied Earth Observation and Geoinformation, vol. 104, Dec. 2021,
doi: 10.1016/j.jag.2021.102505.
[28] F. H. Wagner et al., “Using the U‐net convolutional network to map forest types and disturbance in the Atlantic rainforest with
very high resolution images,” Remote Sensing in Ecology and Conservation, vol. 5, no. 4, pp. 360–375, Dec. 2019,
doi: 10.1002/rse2.111.
[29] N. T. Hoan, H. T. Quynh, L. M. Hang, N. M. Ha, H. T. H. Ngoc, and D. X. Phong, “Estimation of of land use changes in tan rai
bauxite mine by multi-variants change vector analysis (MCVA) on multi-temporal remote sensing data,” Journal of Geoscience
and Environment Protection, vol. 08, no. 03, pp. 70–84, 2020, doi: 10.4236/gep.2020.83006.
BIOGRAPHIES OF AUTHORS
Dr. Khanh Nguyen-Trong holds a Doctor of Computer Science from University of Paris
VI, France, in 2013. He received his M.Sc. (System and Networking) from University of Lyon I in 2008,
his B.S (Information Technology) at Hanoi University of Science and Technology in 2005. He is now a
lecturer at Faculty of Information Technology, Posts and Telecommunications Institute of Technology,
Hanoi, Vietnam. He is a member of Naver AI lab (PTIT and Korean Naver cooperation) since 2020. His
research includes Machine Learning, Deep Learning, Distributed Systems, Agent based Modelling and
Simulation, Collaborative and Participatory Simulation and Modelling, and Computer Support
Collaborative Work. He can be contacted at email: khanhnt@ptit.edu.vn.
Hoa Tran-Xuan received his B.S. in Information System at Vietnam Forestry University in
2012, and MSc in Information System at Posts and Telecommunications Institute of Technology, Hanoi,
Vietnam in 2020. His main research interests are the application of information technology in forestry,
Geographic Information Systems, Forest Cover Change Detection, and Forest Species Classification.
Currently, Hoa Tran Xuan is a lecturer of Computer Science at Vietnam National University of Forestry.
He can be contacted at email: hoatx@vnuf.edu.vn.