An Optimized Deep Belief Network for Land Cover Classification Using Synthetic-Aperture Radar Images and Landsat Images.

This article presents an automated deep learning-based land cover (LC) classification model for remote sensing images, integrating synthetic-aperture radar (SAR) and Landsat-8 data. The model involves three stages: pre-processing via Gabor filtering, multi-feature extraction (including gray-level-co-occurrence matrix (GLCM) texture features from SAR images and temperature vegetation index (TVX), normalized difference vegetation index (NDVI), normalized difference index (NDI), and coloration index (CI) features from Landsat-8 images), followed by classification using an optimized deep belief network (DBN). The DBN weights are fine-tuned using a novel hybrid optimization algorithm called Sunflower adopted Red Deer (SARD), which combines the Red Deer algorithm and Sunflower optimization to improve convergence and global optimum determination. Experimental results on data from Vijayawada, India, demonstrate that the proposed SARD+DBN approach outperforms several existing optimization and classification methods in accuracy, precision, sensitivity, specificity, and error rates across multiple test cases. The study highlights the model’s effectiveness and robustness in classifying heterogeneous land cover types using fused SAR and optical satellite imagery.