Using Convolutional Neural Networks for Scene Classification of Plastic-Mulched Landcover from High-resolution Remote Sensing Imagery

This study applied deep learning methods for scene classification of plastic-mulched landcover (PML) from high-resolution remote sensing (HRRS) images. Firstly, 1,555 and 900 HRRS sample images with a size of $256\times 256$ pixels were labeled into three types (PML, farmland and impervious land) in agricultural regions in Guyuan and Shouguang, China. Secondly, nine commonly used convolutional neural networks (CNNs) and four adjusted AlexNet models were chosen and trained using the labeled image sample set. Finally, the trained 13 CNN models were used to testify the effectiveness of PML scene classification. The results show that ResNet34 has both obtained the highest overall accuracies of 93.85% in Guyuan, and 85.22% in Shouguang.