ISANet: An Interactive Self-attention Network for Cropland Image Change Detection

The change detection task is a vital research topic in the remote sensing area, particularly significant in the context of agricultural land monitoring. Currently, many researchers perform change detection tasks based on the self-attention mechanism. A common method involves applying a self-attention mechanism within single-temporal images, then fusing the feature maps of bitemporal images to decode and obtain the prediction result. We propose a bitemporal interactive self-attention (ISA) mechanism for change detection tasks, with a special focus on cropland areas where monitoring changes can lead to enhanced agricultural productivity and resource management. In the ISA module, we construct pixel dependency relationships both within single-temporal images and between bitemporal images. This allows the network to learn the feature relationship between bitemporal images through the ISA module, thus optimizing the effect of the change detection task. Moreover, we propose a hybrid convolutional feature fusion module that strengthens the learning of change features in the decoding part. Ultimately, the ISANet network we proposed surpassed other advanced algorithms on the Cropland Change Detection (CLCD) dataset. The experimental results show that the proposed ISANet network can deliver robust performance in remote sensing cropland image change detection tasks.