Extracting patches of arable land in hilly areas based on GF-2 remote sensing images and an improved PSPNet model
Traditional methods of classification encounter challenges in extracting accurate and timely information on cultivated land in hilly areas owing to blurred boundaries,diverse forms of land cover,and irregular spatial distribution in the relevant images.In this study,we used Zhugao Town and Gaoban Town in Jintang County of China as the objects of research,and proposed an improved model of the PSPNet semantic segmentation network to extract cultivated land patches from high-resolution satellite images of hilly areas taken by the GF-2 satellite.The CBAM attention module was introduced to the network to enhance its capabilities of feature extraction and expression during model training,and the cosine annealing learning rate was used to accelerate the convergence of the model.The results of tests showed that the improved PSPNet model could extract cultivated land from images of hilly areas with an overall accuracy of 95.69％,1.07％higher than that of the standard PSPNet model,and 1.32％,1.75％and 6.33％higher than the those of the models of semantic segmentation UNet++,DeepLabv3+,and the support vector machine,respectively.This showed that the improved PSPNet model had strong capabilities of feature extraction and expression that could be used to accurately identify cultivated land in hilly areas.This provided important support for decision-making in agriculture,promotes intelligence and precision in the field,and helped improve the yield and quality of crops.
cultivated land in hilly areaPSPNet modelCBAM attention modulecosine annealing learning rateGF-2 remote sensing images
万方数据
维普
