Enhanced Change Detection in Remote Sensing Using NAViT with UNet

Change detection (CD) in remote sensing (RS) images has become an essential tool for land resource planning, disaster monitoring and urban growth analysis. Despite significant progress with deep learning (DL) methods, traditional approaches struggle to effectively extract multi-scale features and balance local and global information while maintaining computational efficiency. This research introduces NAViT-UNet a novel framework leveraging Neighbourhood Attention Vision Transformers (NAViT) and Multi-Temporal Fusion (MTF) Networks for accurate and efficient CD in multi-temporal RS imagery. The proposed model integrates NAViT modules for local feature extraction, reducing computational overhead while enhancing performance. A robust Multi-Temporal (MT) Fusion Network dynamically aggregates temporal feature maps, optimizing the utilization of complementary MT information. The bottleneck layer incorporates Cascaded Atrous Spatial Pyramid Pooling (ASPP) to broaden the feature receptive field, enabling the precise segmentation of small and complex regions. To address imbalanced segmentation challenges, a Mixed Loss Function combining MS-SSIM Loss, Tversky Loss and Focal Loss ensures fine-grained boundary delineation and segmentation accuracy. The model also adopts the Swish activation function which enhances performance over traditional ReLU. Experiments demonstrated the proposed system achieved best performance in terms of 99.8 % of accuracy, 98.81 % of precision, 99.32 % of Recall and 99.07 % of F1 score. This framework establishes a new benchmark for multi-temporal CD, significantly improving the identification and segmentation of change regions with minimal computational costs. The proposed methodology offers a transformative solution for RS applications, unlocking new possibilities in land monitoring, disaster management and urban planning.