Multi-temporal cultivated land change detection method and reliable pseudo label semi-supervised model training method

Abstract

The application discloses a multi-temporal cultivated land change detection method and a reliable pseudo-label semi-supervised model training method, and belongs to the technical field of remote sensing image processing. In order to solve the problems of irregular shape and fuzzy boundary of the changed land block in the farmland monitoring scene and the problem of the scarcity of high-precision labeled samples, the application adopts a multi-temporal cultivated land change detection model to perform cultivated land change detection. The model adopts an encoder-decoder architecture, and the input end receives double temporal images respectively. The encoder performs multi-level feature extraction and performs time-space-frequency domain fusion through a wavelet interaction module, and then inputs into the decoder. The decoder is decoded based on a hybrid double attention module, and the change prediction result is obtained based on the decoder output. The application adopts a reliable pseudo-label semi-supervised training method to train the model. The method differentiates the supervision of the unlabeled samples based on uncertainty maps and confidence, so as to obtain reliable pseudo-labels, and realizes the training of the model.

Images