Crop classification using non-fixed length multi temporal images base on deep learning

Previous studies on crop classification methods based on deep learning for multi temporal images had already determined the number of inputs multi temporal images in the network structure design stage. However, in reality, due to satellite revisit cycles, weather, and other reasons, stable and clear remote sensing images (RSIs) cannot be continuously obtained. Once a period of image is missing from the multi temporal image sequence, the entire method cannot be used. Although methods such as interpolation and using other images instead can be used to address this issue, they greatly reduce the classification accuracy and stability of the methods, limiting their large-scale application. In response to the above issues, we first proposed a flexible multi temporal RSI dataset. For this dataset, an improved version UNet is constructed to train the model. Crop classification experiments shows that this model can be used without limiting the number of RSI periods and time inputs, and the classification accuracy gradually increases with the increase of image periods.