Integrating random forest and irrigation management in geographic information systems-based land suitability and rice productivity modeling in tropical landscapes

This study was conducted in Barru Regency, Indonesia, a region characterized by diverse topography, high ag-ricultural potential, and environmental constraints. It evaluated the predictive performance of the random forest (RF) machine learning algorithm in FAO-based land suitability classification and rice productivity estimation, in- tegrating geographic information systems (GIS) and technical irrigation as a novel managerial variable. Using 12 GIS-derived soil parameters, the RF model achieved high accuracy (0.95 ± 0.01) for land suitability classification via cross-validation. However, its productivity prediction yielded a low R2 (0.32 ± 0.04) with 530 filtered samples, likely due to data complexity. The low R2 value indicates that unmodeled factors, such as agricultural management practices significantly impact productivity beyond the environmental and irrigation variables captured in the model. Including irrigation data improved R2 by 6%, raising it to 0.38 ± 0.06, highlighting the importance of managerial factors in tropical regions with limited infrastructure. Sensitivity analysis identified slope, cation exchange capacity (CEC), and soil depth as key for land suitability, while slope, potassium, and CEC influenced productivity predic- tion. The land suitability map showed that class S3 (marginally suitable) dominates (34,836.62 ha), followed by not suitable (N: 8,303.68 ha) and moderately suitable (S2: 2,722.62 ha), with Barru Sub-district having the largest S2 area (1,181.16 ha). These findings suggest that enhancing irrigation infrastructure and improving soil conditions can support better land management strategies. Despite limitations in productivity prediction, the integration of managerial variables represents an innovative approach, improving model performance and practical applications.