An improved Deeplab based Model for Extracting Cultivated Land Information from High Definition Remote Sensing Images

Despite the effectiveness in extracting information from images, the existing neural network model is not suitable for extracting cultivated land information from high resolution remote sensing images (HRRSIs). In this paper, three effective improvements are made to the existing Deeplab model, i.e., (1) A parameter to adjust dilated convolution kernel is introduced, (2) A linear activation function is selected, and (3) a more precise decoder group is added to the Deeplab model structure. With the above improvements, a HRI-Deeplab (High Resolution Image) model for extracting cultivated land information from HRRSIs is proposed. The HRI-Deeplab model can accurately extract the cultivated land and the advantage is to extract different objects with similar features. 10 large-scale remote sensing images of Feicheng city, China from 2016 to 2017 were selected for experiment. Through the above large-scale remote sensing images, we have cut and produced about 13,000 training images and 17,000 test images. The green land and bare land of cultivated land were selected as the object of model extraction. The results show that the extraction accuracy, recall rate and Kappa coefficients of the HRI-Deeplab are 90%, 89% and 91% respectively, which are obviously better than those of the Deeplab model and SegNet model.