Machine learning and geoprocessing in the modeling of agricultural suitability in a River Basin for sustainable agricultural management

Agricultural expansion faces environmental and socioeconomic challenges. Consequently, the implementation of sustainable soil management practices is essential. This study employed the Random Forest algorithm to map soil chemical parameters and identify areas within the Sorocabuçu River Basin (SRB) that are suitable for growing vegetable crops. Soil samples were collected from 27 points, distributed according to land use and topographical characteristics, and analyzed for the macronutrients Calcium (Ca), Magnesium (Mg), Potassium (K), P-resin, Sulphur (S), as well as pH and Cation Exchange Capacity (CEC). The data were interpolated using the Inverse Distance Weighted (IDW) method. Considering the financial and logistical constraints of sampling, the IDW method was adopted for data interpolation, and it was subsequently validated through statistical analysis. In the supervised classification, we assumed that the soil must possess an adequate pH or CEC, along with favorable macronutrient levels, to be considered suitable for agricultural use. The results indicated that 61.06% of the SRB is highly suitable for growing vegetable crops, characterized by optimal Ca distribution, low Mg concentrations, and uniform K levels. However, P-resin was found to be insufficient in 67.86% of the area. The average pH of 4.94 indicated the acidic nature of the soil, while the average CEC of 88.73 mmolc/dm³ reflected the predominance of acidic cations (H + Al). The Random Forest model demonstrated high performance in classifying agricultural suitability, with a Kappa coefficient of 0.94, sensitivity of 0.97, and specificity of 0.92. The model highlighted pH and Ca as the most influential factors in the algorithm’s decision-making process, emphasizing their significance in predicting soil fertility. Field validation confirmed the reliability of the model, further supporting its potential for application in sustainable agricultural planning. The use of the Random Forest algorithm in this modeling process proved to be effective in facilitating decision-making in sustainable soil management, allowing for targeted interventions to address edaphic limitations. Furthermore, the results contribute to sustainable management practices that support SDG 2 (Zero Hunger and Sustainable Agriculture) and SDG 15 (Life on Land), promoting soil conservation and agricultural productivity. Future research could incorporate additional environmental variables and refine the modeling approach to enhance its applicability across diverse agricultural scenarios.