NSNUNet: A Neighbourhood Attention-Based Siamese Nested Unet for Change Detection using Satellite Images

Detecting the difference between two images of same area over different time is change detection, which plays a crucial role in the domain of remote sensing. Properly detecting the changes can help in various sectors, such as urbanization, deforestation, land use, river erosion, etc. Various deep learning approaches have been developed and implemented for this purpose. But conventional approaches don’t properly capture fine-grained details and maintain high-resolution features. Considering all this, in this methodology, a novel architecture termed Neighbourhood Attention-Based Siamese Nested UNet was proposed, which combines the strengths of Nested UNet and the Neighborhood Attention mechanism. The core component of the model is a Siamese encoder-decoder architecture, which is a skip connection-based UNet++. This design facilitates the reconstruction of fine-grained details and enhances the gradient flow. The neighborhood attention is a more efficient sliding window attention, which, for each query, makes the area of processing focused to a small window, maintaining translational equivariance and improving computational efficiency. NSNUNet is evaluated by high and medium-resolution dataset. The performance of NSNUNet is better than other existing methods when compared, achieving F1-scores of 88.43% and 92.14% on the Levir-CD and agricultural land change detection datasets, respectively.