Extracting aquaculture pond using multi-feature optimization of GF-3 PolSAR Imagery

Abstract: Artificial aquaculture ponds have been mainly distributed in the coastal and inland areas of China in recent years, particularly with the continuous growth of the population and the ever-increasing demand for food. Aquaculture can inevitably pose a series of environmental degradation on a large number of wetlands, soil, and water bodies in ecosystems. Therefore, it is necessary to monitor the distribution and spatiotemporal evolution of aquaculture land, especially for the decision-making on the resource management and the expanding reproduction of fishery. A traditional survey cannot fully meet the high requirements of monitoring in large-scale aquaculture, due to the time-consuming and labor-intensive tasks. In this study, a feature-optimized extraction was proposed for the culture pond using the full-polarization SAR images. Two study areas were selected in the Gucheng Lake area of Nanjing and the offshore area of Yancheng Dongtai in Jiangsu Province of China. The fully polarimetric GF-3 data was used as the data source. 1) 47 polarization features of GF-3 data were obtained using a variety of polarimetric target decomposition, and 8 texture features were extracted using the gray level co-occurrence matrix, according to the Shannon Entropy (SE) polarization feature. All features were merged into one multi-channel image. 2) The multi-channel image was divided into numerous image objects by multiresolution segmentation. 3) The Random Forest-Recursive Feature Elimination (RF-RFE) was applied for the feature selection. 4) An object-oriented random forest was used to extract the aquaculture ponds. Three combined strategies and two classifications were selected to evaluate the new model, including the Freeman polarization decomposition, H/A/α polarization decomposition, and the 47-dimensional optimal polarization without texture features, as well as the Maximum Likelihood (MLC), and Support Vector Machine (SVM). Four strategies of feature combination and three classifications were designed to finally obtain 12 sets of classification. The experimental results show that the accuracy of the random forest was higher than that of MLC and SVM, in terms of the classifier. The feature combination presented the highest accuracy, where the three-dimensional features were selected as SE, SE_I, and Variance, indicating a better performance than that for the feature combination of single-polarization decomposition without texture introduction. Therefore, the features after polarization decomposition could greatly contributed to representing the difference between ground objects, particularly with the higher classification accuracy than before. The texture feature with the Variance was greatly improved the classification accuracy of polarization features. In the Gucheng Lake area, the overall accuracy was 96.85%, and the Kappa coefficient was 0.95. In the coastal area of Dongtai offshore, the overall accuracy was 94.60%, the Kappa coefficient was 0.92. Consequently, the user and mapping accuracies were 94.44%, and 95.07%, respectively, indicating the better universality and stability of the new model for the identification of aquaculture ponds in various breeding modes. More importantly, the GF-3 satellite imagery can be widely expected to apply for the extraction of aquaculture ponds.