4.1. Constructed Decision Tree and Selected Polarimetric Parameters
In this study, 2567 training samples for 10 land cover types (wetland vegetation, rivers, sea, sand, paddy rice, dry land, irrigable land, fish ponds, grasslands and roads) were used to build the decision tree (see
Supplemental Table S1). During the tree development, the depth was set to five to avoid over-fitting. To generate a simple tree, the method of “cost-complexity” pruning was employed based on 10-fold cross-validation. The structure of the final decision tree is shown in
Figure 5. As presented in the figure, the layer mean values for the Shannon entropy, Krogager_Kd, Barnes2_T33, Alpha, HAAlpha_T11, polarization fraction, VanZyl3_Vol, Derd, Barnes1_T33, Neuman_delta_mod, the standard deviation of entropy, the shape index and the distance to neighbor objects were selected for the decision tree. The optimal polarimetric parameters selected by the QUEST algorithm are displayed in
Figure 6.
Figure 5.
Structure of the decision tree. SE, Shannon entropy; KG, Krogager_Kd; B2, Barnes2_T33; HAA, HAAlpha_T11; PF, Polarization Fraction; VZ, VanZyl3_Vol; B1, Barnes1_T33; NDM, Neuman_delta_mod; SD, the standard deviation of entropy; DS, the distance to the neighbor objects.
Figure 5.
Structure of the decision tree. SE, Shannon entropy; KG, Krogager_Kd; B2, Barnes2_T33; HAA, HAAlpha_T11; PF, Polarization Fraction; VZ, VanZyl3_Vol; B1, Barnes1_T33; NDM, Neuman_delta_mod; SD, the standard deviation of entropy; DS, the distance to the neighbor objects.
The radar signals returned from the vegetation-covered area, such as paddy rice, irrigable land, dry land and wetland vegetation, include the vegetation canopy backscatter (volume scattering) and trunk-water or trunk-ground backscatter (double-bounce scattering). The backscatter from grasslands and sand is mainly single-bounce scattering, which is greatly influenced by the soil surface. Water bodies, such as fish ponds, rivers and the sea, generally exhibit low return scattering; these areas are mostly dark on the PolSAR image. The scattering on roads is mainly double-bounced due to the interaction with telegraph poles and trees on either side, so they are shown as bright lines on radar imagery. Overall, the land cover types in this study area are complex, and some of them exhibit similar physical scattering mechanisms; thus, it is difficult to distinguish land cover classes accurately by using polarimetric scattering information alone.
Figure 6.
Polarimetric parameters selected in the decision tree.
Figure 6.
Polarimetric parameters selected in the decision tree.
Shannon entropy is a measurement that was introduced by Morio
et al. [
45]; it is the sum of two contributions related to intensity and polarimetry. The intensity contribution depends on the total backscattering power, and the polarimetric contribution depends on the Barakat degree of polarization. The training samples could be classified into two groups by the mean value of the Shannon entropy with a threshold of −3.01. Crop fields that are near water bodies could not be differentiated from water. Corresponding to the contribution of the di-plane component in the Krogager decomposition [
24], the mean value of Krogager_Kd was helpful in distinguishing water bodies and vegetation. In the right branch of the mean value of Krogager_Kd, vegetation could be identified, as it had stronger double-bounce scattering (trunk-water or trunk-ground backscatter) than the water bodies in the left branch. Representing the averaged scattering mechanisms from surface scattering to double-bounce scattering, the mean value of Alpha with the threshold of 31.511 was used to divide the water bodies into two groups. The left branch included grasslands and water, the dominant scattering of which should be single bounce. However, the trails across fish ponds presented a slight volume scattering. Thus, the mean value of VanZyl_Vol could be used to distinguish water bodies from nearby grasslands. The entropy (H) represents the randomness of a scattering medium, from isotropic scattering (H = 0) to totally random scattering (H = 1) [
44]. For smoother surfaces, such as the ocean, surface scattering dominates and H is near zero. As mentioned, fish ponds presented a slight volume scattering in this study area, so compared with the sea, fish ponds had higher entropy values. The standard deviation of entropy is calculated from the entropy and is more sensitive to the variation in surface roughness. Thus, the standard deviation of entropy was helpful in further classifying the sea and fish ponds. With a relatively compact shape, a fish pond could be easily distinguished from a river by using the shape index. The mean value of Alpha and the distance to the sea could successfully differentiate herbaceous wetland vegetation from artificial grasslands. According to knowledge, wetland vegetation grows on the coast near the sea, and the artificial grasslands are located on farms far from the sea, so the distance to the sea can be regarded as an important spatial indicator to reduce the confusion between artificial grasslands and wetland vegetation. When the distance between an object and the sea was less than 70 pixels, the object was regarded as wetland vegetation; otherwise, the object was grassland. The double-bounce eigenvalue relative difference (Derd), which was introduced by Allain
et al. [
46], is very sensitive to surface roughness. In the study, the mean value of Derd was used twice to classify paddy rice, irrigable land and dry land, all of which had similar dominant scattering mechanisms, but different surface roughness.
The classes in the right branch of the Shannon entropy node, such as roads, sand and paddy rice, could be divided into two groups according to the mean value of Barnes2_T33. The parameter, polarization fraction (PF), ranged between zero and one [
47]. When the third eigenvalue (
) of the coherency matrix was zero, the entire radar return was polarized. When
was greater than zero, the value of PF dropped. As shown in
Figure 6 (PF), roads exhibited lower PF than paddy rice and dry land. As a result, the mean value of PF could be used to reduce the confusion between farmland and roads. With similar scattering mechanisms, but different surface roughness and soil moisture, paddy rice and dry land could be distinguished by employing the mean value of Derd.
Figure 6 shows that the sand and paddy rice were relatively homogeneous in the HAAlpha_T11 image, so the mean value of HAAlpha_T11 could be used to separate roads from paddy rice and sand. Corresponding to pure targets, Barnes1_T33 was useful in further distinguishing sand from paddy rice. The Neumann decomposition, which is a simple vegetation model for polarimetric covariance or coherency matrix elements, was originally intended for volumetric distributed targets in terms of second-order statistics. As a general scattering mechanism indicator, Neumann_delta_mod was helpful in identifying some paddy fields, which typically had stronger volume scattering than roads.
4.2. LULC Classification Results
The LULC classification results from the proposed method and five other methods are displayed in
Figure 7 and
Table 2 (and see
Supplemental Tables S2–S7 for detailed results from individual methods). A total of 2043 field samples were collected to calculate the classification accuracy (
Tables S1), including overall accuracy (OA), producer’s accuracy (PA) and user’s accuracy (UA). The analysis and discussion are based on visual interpretations and quantitative comparisons. As seen from the comparison between the proposed method and WSC method, the overall accuracy for the latter method was 66.6%, which was lower than that of the proposed method (87.3%). In addition, the producer’s and user’s accuracies for most classes decreased with the WSC method. In particular, fish ponds, the sea and rivers were grouped into one class due to their similar backscatter values (
Figure 7b). Given that soil moisture is a crucial factor that generally affects the land cover scattering type, grasslands have lower soil moisture than paddy rice. In
Figure 7b, however, some artificial grasslands could not be differentiated from paddy rice by only using scattering mechanisms, perhaps because the PolSAR data were acquired in April when little water was present in the paddy fields. Furthermore, some bare lands that exhibited dominant single-bounce scattering were observed in the paddy rice during the fieldwork. As a result, some of the paddy rice showed similar scattering to artificial grasslands. This comparison demonstrated the effects of integrating the polarimetric decomposition, object-based analysis and decision tree algorithm.
Figure 7.
Classification results from: (a) the proposed method; (b) the Wishart supervised classification; (c) the proposed method without polarimetric parameters; (d) the proposed method without an object-based segmentation; (e) the proposed method without textural and geometric information; and (f) the proposed method using the nearest neighbor classifier instead of the decision tree algorithm.
Figure 7.
Classification results from: (a) the proposed method; (b) the Wishart supervised classification; (c) the proposed method without polarimetric parameters; (d) the proposed method without an object-based segmentation; (e) the proposed method without textural and geometric information; and (f) the proposed method using the nearest neighbor classifier instead of the decision tree algorithm.
The results of the proposed method and PWPP method were compared to verify the improvement after polarimetric parameters were included in the LULC classification. From
Table 2, the producer’s and user’s accuracies for almost all land cover types decreased when the polarimetric parameters were not used in the classification. Moreover, the overall accuracy decreased by 13.3% compared with the proposed method. Specifically, as observed in
Figure 7c and
Table 2, fish ponds achieved rather low accuracies due to their confusion with the sea. The user’s accuracy of fish ponds with the PWPP method was 24.6%, and the producer’s accuracy of the sea was 49.2%; in the proposed method, these accuracies increased to 89.5% and 88.8%, respectively. Entropy, which is a polarimetric parameter derived from the Cloude–Pottier decomposition, was useful in identifying fish ponds and the sea (
Figure 5). The mean value of Derd helped distinguish paddy rice and dry land. In the proposed method, the producer’s and user’s accuracies of dry land were 88.8% and 87.1%, respectively, while in the PWPP method, the two indicators decreased to 79.4% and 84.3%, respectively. For paddy rice, the two indicators were 84.9% and 90.9%, respectively, in the proposed method, but decreased to 75.7% and 71.6%, respectively, with the PWPP method. The improvement in the accuracies for roads and paddy rice indicated that the mean value of Neumann_delta_mod was useful in separating these classes. Overall, the polarimetric parameters selected by the QUEST algorithm were highly important in reducing land cover confusion and improving the classification performance in the coastal wetlands.
Table 2.
Classification accuracy. SA, sand; DL, dry land; FP, fish pond; GL, grassland; IL, irrigable land; PR, paddy rice; RI, river; RO, road; S, sea; W, wetland vegetation; WSC, Wishart supervised classification; PWPP, proposed method without polarimetric parameters; PWOS, proposed method without object-based segmentation; PWTG, proposed method without textural and geometric information; PNNC, proposed method with the nearest-neighbor classifier.
Table 2.
Classification accuracy. SA, sand; DL, dry land; FP, fish pond; GL, grassland; IL, irrigable land; PR, paddy rice; RI, river; RO, road; S, sea; W, wetland vegetation; WSC, Wishart supervised classification; PWPP, proposed method without polarimetric parameters; PWOS, proposed method without object-based segmentation; PWTG, proposed method without textural and geometric information; PNNC, proposed method with the nearest-neighbor classifier.
Method | Accuracy | Class |
---|
SA | DL | FP | GL | IL | PR | RI | RO | S | W |
---|
Proposed method | PA (%) | 83.2 | 88.8 | 86.0 | 84.6 | 80.3 | 84.9 | 95.3 | 92.0 | 88.8 | 93.5 |
UA (%) | 89.2 | 87.1 | 89.5 | 86.4 | 87.4 | 90.9 | 85.5 | 85.2 | 90.0 | 80.7 |
OA (%) | 87.3 | |
WSC | PA (%) | 84.3 | 77.9 | 42.8 | 75.4 | 76.1 | 74.1 | 30.3 | 89.3 | 52.3 | 94.6 |
UA (%) | 83.4 | 87.2 | 6.6 | 78.8 | 87.4 | 71.6 | 80.7 | 87.9 | 17.3 | 72.5 |
OA (%) | 66.6 | |
PWPP | PA (%) | 88.6 | 79.4 | 73.6 | 74.4 | 67.9 | 75.7 | 73.4 | 91.1 | 49.2 | 89.3 |
UA (%) | 83.0 | 84.3 | 24.6 | 67.6 | 90.6 | 71.6 | 80.7 | 92.1 | 79.6 | 72.5 |
OA (%) | 74.0 | |
PWOS | PA (%) | 90.6 | 78.8 | 92.6 | 72.8 | 69.4 | 82.4 | 51.6 | 90.7 | 55.6 | 92.1 |
UA (%) | 79.6 | 84.8 | 82.5 | 67.6 | 90.6 | 78.8 | 23.4 | 93.2 | 89.0 | 77.6 |
OA (%) | 77.1 | |
PWTG | PA (%) | 87.8 | 79.9 | 61.6 | 71.8 | 71.2 | 75.7 | 75.6 | 90.1 | 47.6 | 84.5 |
UA (%) | 84.4 | 87.1 | 14.0 | 65.6 | 87.4 | 71.7 | 80.8 | 92.1 | 84.1 | 72.5 |
OA (%) | 73.2 | |
PNNC | PA (%) | 87.6 | 80.5 | 94.6 | 70.7 | 65.5 | 89.5 | 65.7 | 90.3 | 82.6 | 94.4 |
UA (%) | 79.5 | 83.8 | 69.7 | 67.1 | 91.7 | 81.2 | 81.8 | 88.4 | 84.6 | 77.8 |
OA (%) | 80.5 | |
The PWOS method was used for the LULC classification by applying similar operations to the proposed method, including polarimetric decomposition and the decision tree algorithm; however, the object-based segmentation was not employed. Regarding the comparison between the object-based (the proposed method) and pixel-based (the PWOS method) classifications,
Figure 7d and the accuracy evaluation in
Table 2 show that the performance of the proposed method was generally enhanced. The overall accuracy of the PWOS method had an accuracy of 77.1%, which was 10.2% lower than the proposed method. Specifically, the PWOS method failed to correctly distinguish rivers, roads and grasslands due to the limitation of using only single-pixel information. More importantly, there was a mass of isolated points in
Figure 7d compared with
Figure 7a, which provided powerful evidence that the object-based method was less affected by noise in the SAR image.
Figure 7a had lower spatial heterogeneity compared with
Figure 7d, and the proposed method represented reality more accurately than the pixel-based method. These results proved the importance of the object-based method in PolSAR LULC classification.
When textural and geometric information were not employed in the classification, the overall accuracy of the PWTG method was 73.2%, which was 14.1% lower than the proposed method. As displayed in
Figure 7e, some rivers could not be separated from fish ponds by merely using polarimetric information, because both were dominated by single-bounce scattering. Because the shapes of these two land cover types are significantly different, the shape index could be used to distinguish them. The producer’s and user’s accuracies of fish ponds were 61.6% and 14.0% with PWTG, respectively; after the shape index was added to the classification, the two indicators increased to 86.0% and 89.5%, respectively. Some wetland vegetation areas were confused with grasslands when using the PWTG method, because their scattering mechanisms were similar. Based on knowledge and reference maps, wetland vegetation grows on the beach near the sea and grasslands are located on farms far from the sea, so the distance to the sea can be regarded as an important criterion for discriminating herbaceous vegetation-covered wetlands and grasslands (
Figure 5). The accuracies of the two land cover types displayed in
Table 2 also demonstrate this point. This comparison noted that the spatial relationships of the segmented objects could be used to improve the classification accuracies.
A comparison was also made between the proposed method and the PNNC method. In the latter method, the nearest-neighbor classifier replaced the decision tree algorithm, and the feature selection method was the feature space optimization function provided in eCognition. From
Table 2, the overall accuracy of the PNNC method was 80.5%, which decreased by 6.8% compared with the proposed method. The classification results of irrigable land and roads using the PNNC method were slightly higher than those using the proposed method. However, for most of the land cover types, the proposed method obtained higher producer’s and user’s accuracies than the PNNC method.