Spatial Location-Guided Global Context Modeling Network for Cropland Dynamic Monitoring in Multitemporal Remote Sensing Images

Monitoring the time series change of cropland is an important measure to protect cropland resources. In recent years, many studies have applied convolutional neural networks (CNN) to cropland dynamic monitoring tasks and achieved good results. However, the problem that CNN is difficult to capture long-range dependencies limiting network performance still needs to be solved. For this purpose, a spatial location guided global context modeling network (SLGNet) is proposed for cropland dynamic monitoring. Specifically, SLGNet accurately locates cropland change and long-range modeling by constructing a novel cropland change predictor. In the cropland change predictor, the spatial location module adds spatial coordinates to guide the network to learn spatial location information about the feature map and improve the perception ability of cropland change at different locations. In addition, the global context modeling module in the cropland changes predictor aggregates the difference information from all locations to strengthen the identification of cropland change at different scales by SLGNet. The experimental results show that SLGNet obtains comprehensive optimal results on four datasets, including CLCD(F1:76.46%), JL-1(F1:89.07%), PX-CLCD(F1:95.95%), and Hi-CNA(F1:77.48%).