National‐Scale Orchard Monitoring in Pakistan through Machine Learning and Remote Sensing

Accurate geospatial inventories of fruit orchards are essential for precision horticulture and food security, yet Pakistan lacks consistent datasets at district and tehsil levels. This study develops the first national-scale, ML-enabled framework for orchard delineation and yield estimation by integrating multi-temporal Sentinel-2 imagery on Google Earth Engine (GEE) with high-resolution Pakistan Remote Sensing Satellite-1(PRSS-1) data, advanced machine learning, and object-based image analysis (OBIA). Among tested classifiers, Random Forest (RF) achieved the highest performance on pixel-based Senti-nel-2 data (OA = 79.0%, κ = 0.78) outperforming Support Vector Machines (74.5%, κ = 0.74) and Gradient Boosting Decision Trees (73.8%, κ = 0.73), with improvements confirmed by McNemar’s test (p < 0.01). Integrating RF with OBIA on PRSS-1 imagery further enhanced delineation, increasing OA to 92.6% (κ = 0.89), with producer’s and user’s accuracies of 90.4% and 91.5%, and IoU improving from 0.71 to 0.86 (p < 0.01) in orchards. Yield modeling based on field-observed data showed mean- and median-based vegetation index aggregation performed best (R² = 0.77–0.79; error = 72–105 kg/tree), while extreme-value models performed poorly (R² = 0.46–0.56; error >560 kg/tree). The resulting validated geospatial inventory of citrus (Central Punjab) and mango (South Punjab) supports productivity forecasting, export planning, and resource allocation, demonstrating the transformative role of national space assets in advancing precision horticulture and strengthening food security in data-scarce regions