Hybrid deep learning and optimization-based land use and land cover classification for advancing sustainable agriculture in Najran city, Saudi Arabia

In arid regions, land-use/land-cover (LULC) mapping, in central ways, plays a significant role in the sustainability of agriculture. The paper builds on a streamlined hybrid learning system that can categorize the terrain in the Najran, Saudi Arabia, based on 2023 Landsat-8 images to identify indicators of sustainable land use and to guide decisions on the issue. Ten CNN-Random Forest variants were tested; to highlight agronomically informative features, the redundancy of features was minimized with the help of the Ant Colony Optimization. The top models were the ones with high, measured accuracy: VGG19-RF (97.56 overall accuracy, 9726), GoogleNet-RF (96.15), DenseNet121-RF (92.39) and ResNet152-RF (92.26). Class-based area statistics show the presence of built-up area at approximately 29-33%, vegetation area at approximately 14-25%, bare ground at approximately 9-22%, and water area at approximately 9-22%, which reflect the urban growth and development and pressures on developed and irrigated lands. Best models also had a precision/recall/F1 9699% showing dependable separation of agronomic classes. These measured outputs can be converted into operational sustainability indicators: vegetation and bare-soil area to inform crop rotation, soil-cover, and erosion management; built-up encroachment measures to safeguard agricultural buffers; water-body delineation measures to prioritize irrigation efficiency and groundwater recharge areas. The proposed hybrid model is more accurate and interpretable than single-architecture baselines, which provides an extendable avenue toward commonplace LULC monitoring, agricultural risk screening, and policy tracking within the framework of Saudi Vision 2030. The framework can easily be applied to other semi-arid regions in which sustainable production relies on accurate and periodically updated spatial information on the land.