The Study Of Cultivated Land Extraction And Change Detection Based On Deep Learning Of Gaofen Satellite Images

With the implementation of the geographic national conditions monitoring program,relevant departments in China have put forward a higher demand for geographic information mapping,which includes the demand for mapping of cultivated land.The extraction of cultivated land is crucial to obtain timely information on the change of cultivated land and the utilization of cultivated land,so it is imperative to further improve intelligent remote sensing surveying and mapping of cultivated land.Deep learning technology has become an effective tool for remote sensing target extraction and change detection.In this thesis,the extraction and change detection of cultivated land is achieved with deep learning based on the data source of Gaofen-2(GF-2)images.The main research contents are as follows:1)In order to address the problem of lack of datasets for cultivated land extraction and change detection in China,the special cultivated land extraction datasets and a small sample cultivated land change detection dataset were built according to the actual research area and research objectives,which can prepare for the subsequent extraction and change detection of cultivated land.The cultivated land extraction datasets and a small sample cultivated land change detection dataset were built based on GF-2 images in Shijiazhuang City,Hebei Province.2)In order to solve the problems of poor integrity and rough edge localization of large scale cultivated land extraction in high-resolution remote sensing images,a multiscale bilateral spatial direction-aware network to extract cultivated land was proposed in this thesis by combining multi-attention mechanism.The extraction of attention feature maps was achieved by using spatial contextual direction perception to retain detailed direction-aware information to determine the features of spatial relationships and locations of cultivated land;the extraction of local-to-global information of cultivated land was achieved by capturing dense multi-scale detailed features with pyramid pooling and attention awareness.Finally,the refined extraction of cultivated land was achieved by fusing different levels of feature information and combining with boundary-aware supervision.The proposed method was verified to complete the refined extraction of cultivated land with high efficiency through comparative experiments as well as quantitative and qualitative analysis on self-made cultivated land datasets.The proposed method achieved an accuracy of 94.81% and 90.85% in extracting cultivated land between urban and rural areas and complex cultivated land in mountainous areas,respectively.3)In order to address the current problems of depending heavily on cultivated land labels and to improve the precision of semi-supervised change detection of cultivated land,a multi-task feature perturbation semi-supervised cultivated land change detection method was proposed in this thesis,which further improved the precision of cultivated land change detection by combining with a multi-task detection network after improving the precision of cultivated land extraction by multi-scale bilateral spatial direction-aware network.In addition,feature perturbation was added to the network branches and a triangular cross-consistency constraint was imposed on the generated pseudo-labels,so that the semi-supervised cultivated land change detection can be achieved under the three-way task branch consistency constraint.Experiments and analyses on a small sample cultivated land change detection dataset were performed to verify that the proposed method can effectively improve the precision of semi-supervised cultivated land change detection.Compared to other methods,the accuracy obtained by the proposed method at different proportions of labeled small samples is improved by about over 1.62%,0.77%,0.14%,and 1.29%,respectively.