Advance Change Detection in Remote Sensing with a CNN-Transformer enhanced model by Spatial-Temporal Attention Mechanism
Abstract
Remote sensing change detection is a very essential task since it is used to monitor and analyze the dynamics of changes in the surface of the earth. Such as urbanization, deforestation, agricultural development, and impact of disaster. In this paper, a new deep learning-based is presented change detection methodology based on the combination of Convolutional Neural Networks (CNNs) and Transformer models, augmented with Convolutional Block Attention Modules (CBAM). The suggested model exploits the powers of CNNs to extract local features, Long-range dependency capturing transformers, and attention which concentrates on the most significant aspect of the image, enhancing performance in multifaceted settings. Our model is trained on high resolution remote sensing images and proves to be much better in important performance measurements, accuracy, precision, recall and IoU (Intersection over Union) and the accuracy is highest at 94.8%. The model also is highly generalized, as it is shown by stable improvements in validation curves, which means that it can operate with unseen data effectively. Further integration of multimodal data, including optical and Synthetic Aperture Radar (SAR) imagery is also possible increases the robustness of the model to diverse environmental conditions as well as sensor modalities. The findings indicate that the proposed deep learning hybrid model can be successfully used in various change detection uses, such as in the environment, the growth of cities, and reaction to disasters. This work provides a solid base of future progress in remote sensing change detection, which will provide a possibility of enhanced decision-making processes pertaining to land use and management of resources.