Abstract: Aiming at the problems in change detection,such as difficultly implement of prototype algorithm and poor applicability of sample set,this paper adopts SNUNet as the prototype algorithm.A comparison is made with similar algorithms on the public data set CDD,such as FC-EF,FC-Siam,UNet++＿SOF,etc.,and the results show that the SNUNet can fully utilize the semantic information of surface coverage,and has higher and more stable change detection accuracy.In order to improve applicability of the algorithm model additionally,more than 26 000 sets of change detection sample datasets are produced by using domestic multi-source satellite remote sensing images in Anhui province for localized training and optimization of the algorithm,and the accuracy reaches 84.3%.An application of cultivated land change detection is conducted in the experimental area.The algorithm and sample sets have good application prospects in the protection of cultivated land in Anhui province,and provide sufficient technical support for change detection in other places and different fields.

Key words: domestic satellite image, sample set, deep learning, cultivated land change detection