Deep Learning Models for Change Detection in Multispectral Satellite Imagery

Identification of changes utilizing multispectral satellite imagery has become a necessity for monitoring urban and deforestation and alteration of environments due to climate changes. Traditional methods use a static and simplistic approach and are especially poor at capturing the fine-grained spatial-temporal changes which are better suited for models that derive features hierarchically. This work proposes a modified convolutional and recurrent deep learning architecture that is specially tailored for change detection and integrates spectro-temporal, spatial, and spectral features simultaneously. The proposed method yields a 97.3% overall accuracy, 96.8% F1 score, and Kappa of 0.95 which is a score of 0.95 concerning the three deep learning models of utmost exposure, suggesting a dominance over conventional CNN, Siamese networks, and U-Net based architectures. Analysis of some multispectral datasets that are open for the public demonstrates the effectiveness of the method for change detection and the controlling of false positive rates. Such outcomes highlight the positive impact of deep learning on remote sensing analytics, emphasizing the effectiveness of the proposed method for extensive environmental surveillance and decision-making support.