Hybrid deep learning with a random forest system for sustainable agricultural land cover classification using DEM in Najran, Saudi Arabia

Abstract Sustainable agriculture depends heavily on precise LULC classification to support soil conservation, water resource planning, and environmentally conscious land use. This study proposes a hybrid deep learning system integrating VGG16 and EfficientNetB7 models with a Random Forest (RF) classifier to classify agricultural and other LULC types in Najran, Saudi Arabia, utilizing digital elevation models (DEMs) and Indian Remote Sensing Advanced Wide Field Sensor satellite data from 2020. A stereo-derived DEM was used to extract topographical features, which, combined with multi-temporal imagery, were processed through VGG16 and EfficientNetB7 for spatial feature extraction. The Grasshopper Optimization Algorithm was applied to select the most essential features and remove the unimportant and redundant ones. The features were then fed into an RF classifier to classify the Najran terrain map efficiently. Evaluation of the hybrid system showed promising results for classifying the Najran terrain map, achieving an accuracy of 94.2%, precision of 79.88%, recall of 79.22%, F 1-score of 79.53%, and specificity of 96.01%. The system demonstrated robust performance in differentiating agricultural lands from urban and natural terrains, enabling efficient monitoring of land use patterns. This approach supports sustainable agricultural practices and environmental stewardship by providing decision-makers with high-resolution, automatically classified land maps for strategic planning in arid regions, such as Najran.