2.1.

Data Processing

For the preprocessing task, SPOT-6 image was corrected for radiometric and atmospheric effects. Also, speckle in the SAR data was removed using $5\times 5$ kernel local sigma. Subsequently, the image was enhanced texturally with $3\times 3$ kernel textural occurrence and resampled to 6-m spatial resolution to match SPOT-6 image. The two datasets were coregistered to bring them into correct alignment using image-to-image registration. Thereafter, the images were reprojected to Universal Transverse Mercator projection. Then, the two datasets were fused by layer stacking. Unlike other fusion methods that transform the image spectral, layer stacking combines dataset from different sensors without altering the spatial and spectral characteristics of the original data.¹ This ability to retain the image fidelity is an advantage for object-based image analysis because the algorithm exploits all image properties to detect physical object in the image.

2.2.

Optimization and Segmentation

Then, the segmentation process was optimized using integrated Taguchi-objective function optimization strategy.⁴¹ Based on the optimal parameter combinations, the image was segmented and the result classified using different algorithms, including data mining (DM) using a developed rule-set. Finally, the effect of the optimization was evaluated. Figure 2 presents the data processing and analysis workflow.

Fig. 2

Methodological workflow.

Image segmentation is the first and most fundamental step in GEOBIA.²⁷ The widely used region growing multiresolution segmentation (MRS) was employed. MRS starts with a single-pixel image object as seed, and then other neighboring image pixels are combined in several successive steps to produce larger ones till the predefined criteria are met.⁴⁴ However, the quality of this process depends on the proper selection of segmentation parameters: scale and homogeneity (shape and compactness) values. The scale value controls the size segment: high-scale value produces large image segment and small-scale value generates small segments.^10,45 The other parameters, homogeneity, combine color and shape properties. In this research, the robust Taguchi-objective function optimization technique was used to derive appropriate parameter combination for the scale, shape, and compactness.^{11,14,40,41,17,20} The technique integrates the statistical Taguchi and spatial objective function optimizations methods iteratively in a single processing workflow to produce an accurate result.

In GEOBIA, segmentation of image spectral relative to their spatial arrangement (autocorrelation) is fundamental to feature identification and grouping. The segmentation technique divides image into homogeneous contiguous regions that enclose identical pixels as objects within each segment based on the assumption that an image pixel most likely belongs to the same object as its neighboring pixels.⁴⁶ Accurate partitioning of image into distinct image objects is dependent on appropriate selection of segmentation parameters: scale, shape, and compactness. Over the time, the objective function,⁴⁷ which attempts to select appropriate parameters that can produce the best quality segmentation based on intrasegment homogeneity and intersegment reparability, has been widely used. However, this optimization technique relies on arbitrary selection of a range of parameters for experimentation. Not only that, it emphasizes on varying the scale factor while keeping the shape and compactness factors constant, even though the quality of the resulting segments depends on the correct combination of the three parameters. And this is perceived as bias for the objective function. Hence, the idea of incorporating Taguchi method into the optimization process by, first, optimizing the design of experiment using the Taguchi orthogonal array and, second, modeling a unique optimal parameter combination with the Taguchi SNR is presented. Comprehensive details of this approach can be found in literature.^41,48,49

For experimental design, five levels were defined for the three segmentation parameters (Table 2). The orthogonal array minimizes the number of experiments to only 25 number of experiment coded L25 ($3^5$) compared to 243 experiments using the standard factorial. The plateau objective function (POF)³⁵ was measured for each experiment to evaluate the quality of the segmentation for each experiment. POF is a combination of the weighted variance and spatial autocorrelation (Moran’s index) to evaluate both of the intersegment homogeneity and heterogeneity of image objects. To assess experimental results, SNR is calculated as a measure for the determination of the quality. The SNR values yield the optimum segmentation parameters, which were used for image segmentation, followed by classification to extract land-cover classes. Different land-cover features were correctly classified by exploiting the spectral, textural, and spatial relationship of the image objects (segments).

Table 2

Segmentation parameter level and variable definition.

2.3.

Classification and Accuracy Assessment

The outcome of segmentation is an unclassified image object with database of the layer values, such as spectral indices, backscattering values, and textural parameters (Table 3) that allows manipulating the feature characteristics for DM. The optimized image objects generated was classified into nine land-cover classes (palm oil, initial paddy stage, intermediate paddy stage, matured paddy stage, bare soil, flooded soil, built up areas, water bodies, and grass or other vegetation) using SVM, RF, and rule-based DM approach. Theoretical bases of SVM, RF, and DM can be found in literature.^{13,19,18,50–55}

Table 3

Description of the feature space included in SVM, RF, and rule-based classification to classify image objects derived from multisource data and optimized by Taguchi technique.

SVM and RF classifiers were applied by selecting training samples representative of the respective feature class. But for the rule-based classification, decision tree (Fig. 3) was constructed by implementing the C4.5 algorithm in Weka,⁵⁵ an open source software using the image object attributes and indices (Table 3). The defined relationships between the image attributes and land-cover classes were utilized in building rule-sets for classifying the image objects. Performance of the optimization process was evaluated by comparing its result with six different studies in which their segmentation parameters were arbitrarily combined and classified using the rule-based classification method. Evaluation of the classifiers was based on the classification accuracy using the traditional confusion matrix and its measures (overall accuracy, kappa coefficient, etc.).⁵⁷ Selection of training image objects for classification and accuracy assessment was done randomly. The GPS points collected during the site visit does not cover the entire range of classes defined in this study due to the coverage and accessibility. Therefore, the process of selecting training samples was guided by the GPS points, the land-use map provided by the Town and Country Planning Department and GoogleEarth image. In addition, quantitative assessment was carried out to know whether there is a significant change in the classification accuracy result before and after optimization using McNemar’s tests.⁵⁸

Fig. 3

Decision tree generated in Weka for developing russets used for the classification.

3.1.

Optimization Output

The first two stages of optimization derived from the 25 experiments produced a preliminary view of some level of optimization for the MRS and SVM parameters, where the POF and kappa values are indicators (Table 4). However, at this level, the operation still pose some challenges of the correct choice of optimal combination because there are a number of closely related POF and kappa from different combinations that seems optimal.

Table 4

L25 orthogonal array for MRS and SVM parameters and experiments responses.

But further iteration using SNR “larger is better” option provided refined optimal values (Fig. 4 and Table 5) that eliminates the ambiguity discussed above. Ultimately, the results yield optimum parameter combination 60:0.7:0.9 for the scale, shape, and compactness factors, respectively, for the MRS and 90 and 0.3 for $C$ and $\gamma$, respectively, for SVM.

Fig. 4

Main effect plot of the SNR “larger is better” for (a) MRS and (b) SVM parameters.

Table 5

Statistical SNR evaluations for MRS and SVM parameters.

3.2.

Classification Results and Accuracy

Classification results for the optimized (SVM, RF, and DT) and unoptimized (DT) segmentations processes are presented in Figs. 5 and 6. Each classified image contain nine classes: palm oil, initial paddy stage, intermediate paddy stage, matured paddy stage, bare soil, flooded soil, built up areas, water bodies, and grass or other vegetation types.

Fig. 5

LULC map using the optimized segments: (a) SVM, (b) RF, and (c) machine learning-based DT.

Fig. 6

Classification results of the unoptimized process using arbitrary segmentation parameter combinations (a) 100, 0.9, 0.3; (b) 90, 0.7, 0.3; (c) 80, 0.7, 0.5; (d) 70, 0.7, 0.7; (e) 60, 0.9, 0.5; and (f) 50, 0.9, 0.1.

Qualitatively, the use of multiattributes of the surface characteristics using the images and their derivatives enhanced the quality of the resulting map. Particularly, in tropical agricultural areas, integration of optical and radar sensor data has provided additional information to separate intraspecies vegetation and various land-cover classes.^3,5,59 The classification maps obtained from the optimized SVM, RF, and DM classifiers do not show much difference except for the flooded soil class in the southern part of the image, which is well classified in SVM and RF and partly misclassified as grass/other vegetation in DM. Aside this class, DM shows superiority in identifying subtle features as can be seen in the lot demarcation boundaries and road network, most of which are not detected in SVM and RF. The HH polarization is highly sensitive to moisture content and vegetation;⁵ these accounts for the well-classified vegetation types. Also, the rivers and their tributaries are well mapped in SVM and DM; but further away from the main river, the tributaries were not detected in RF. Visual analysis reveals that SVM and RF exhibit misclassification among built-up area, bare soil, and matured paddy classes and also between flooded soil and grass land, particularly within the paddy fields, all of which are distinctively separated into their respective classes with DM classifier.

Quantitative evaluation of the classification yielded overall classification accuracy of 87.25%, 88.69%, and 91.79% for SVM, RF, and DM, respectively (Table 6). The accuracy obtained is similar to the result obtained with closely related works. For example, Zhang and Xie¹⁹ obtained classification accuracy of 85% and 89% with SVM and RF, respectively, after fusion. Similarly, Ribeiro and Fonseca¹³ obtained an overall accuracy of 85.66% from pan-sharpened WorldView-2 using object-oriented techniques and DT classifier. On a scale of performance, the three maps produced acceptable results; however, DM using decision-based approach produces superior quality.

Table 6

Error matrix for SVM, RF, and DT classifiers with UA and PA.

Classification maps produced from the unoptimized segmentation show somewhat a similar result with obvious misclassification of the matured paddy class as built-up area. This phenomenon is much pronounced for the larger scale categories (i.e., 100, 90, and 80); but as the scale reduces, so also the degree of misclassification reduced. In all the outputs, the main water bodies are well represented, but most of the tributaries are misclassifies as built-up area, intermediate paddy, or bare-soil [Figs. 6(a)–6(e)], whereas at scale 50 [Fig. 6(f)], the tributaries are correctly classified as water. This indicates that using large scale could result in grouping two or more features in a segment (under segmentation), conversely, if the scale is too small it will lead to over segmentation, which could equally complicates the classification process.^32,33,60 The trend in the visual analysis is reflected in the quantitative evaluation results (Table 7).

Table 7

Classification accuracy of the unoptimized parameter combination.

In Table 7, it can also be observed that the classification accuracy of the unoptimized segmentation performs well, with the overall accuracy range between 84.44% and 88.08% for both SVM and RF, respectively. However, the optimized DT result out performed them all. On the basis of this, the optimized DT classification output was selected as the best for the LULC map (Fig. 7). Statistically, the McNemar’s computed (two-sided) $p<0.0001$ is less than the conventional 0.05 at 95% CI, we cannot accept the null hypothesis. Therefore, the conclusion is that there is a significant difference between the optimized and unoptimized classification processes. In general, optimization significantly improved the quality of the classification; however, careful selection of segmentation parameter combination can also produce acceptable result, as demonstrated in Table 7. Inability to obtain sufficient independent ground data for the evaluation process may affect the result; nonetheless, the use of the basic map provided and GoogleEarth image is adequate for this task. Decision tree based on DM provides better qualitative LULC classification map with the optimized segmentation.

Fig. 7

Final LULC map generated from the optimization process and DT classifier.