Optimized deep learning based classification and prediction of land use/land cover changes from satellite images

Prediction of Land Use/Land Cover (LU/LC) changes from satellite images is crucial for effective environmental monitoring, yet existing methods often struggle with accurately distinguishing small segments and boundaries in high-resolution imagery. To overcome this, an Adaptive multi-scale dual attention network with Optimized Dual interactive Wasserstein generative adversarial network based Classification and Prediction of Land Use/Land Cover Changes from satellite images (AMSDAN-ODIWGA-LU/LC) is proposed. Initially, the input imageries are obtained from DeepGlobe Land Cover Classification Dataset. It covers images captured over Thailand, Indonesia, and India between the years 2013 and 2017 with 50 cm pixel resolution. Then the images are pre-processed using Altered Phase Preserving Dynamic Range Compression technique. The pre-processed image is given to Adaptive multi-scale dual attention network (AMSDAN) for classifying the image as urban land and Forest land. Then, the LU/LC changes are predicted using Dual interactive Wasserstein generative adversarial network (DIWGAN) method. The DIWGAN is optimized using Sand Cat Swarm Optimization, as it does not have any optimization strategies. The proposed AMSDAN-ODIWGA-LU/LC technique is implemented in Python and examined under performance metrics, like accuracy, specificity, sensitivity, f-score, precision and Kappa coefficient. The proposed technique provides 12.34%, 26.37% and 24.39% higher accuracy, 11.23%, 42.17% and 22.39% higher precision compared with existing methods like, Multiscale Context-Aware Feature Fusion Network for Land Cover Categorization of Urban Scene Imagery, Long-range correlation supervision for LC classification from remote sensing images, An effectual network and Sobel boundary loss for LC categorization of satellite remote sensing imagery respectively.