Change Detection for Land-Cover Classification Using Deep Learning

This work presents a deep learning-based strategy to recognize changes and classify six fundamental groups for land-cover: non-vegetated ground surface, tree, low vegetation, water, buildings, and playgrounds. This method effectively detects semantic changes in multispectral satellite images by means of a deep Siamese network architecture. The results show an average Intersection over Union (mIoU) of 89.3% over all classes together with an overall accuracy of 92.5%. With 91.2% accuracy in building detection, strong urban environment adaptation, and efficient identification of water coverage at 95.1% and tree coverage at 93.7%, the technique shows Confirming the results were tabular analysis, confusion matrices, and a thorough review of performance criteria. The suggested method's strength and efficiency imply its possible relevance in many fields needing radar imaging. The results show how well our remote sensing system detects and labels significant changes.