Enhancing crop recommendation through graph attention and tabular deep learning on soil–climate datasets

Abstract

Crop recommendation is a critical challenge in modern agriculture of any day and age because of the enormous variation in the soil properties, climatic variations and different management methods. To solve this, the current research comes out with a new model of hybrid crop recommendations created and tested on the latest agricultural data sets (2023–2025) on Andhra Pradesh and Telangana, India. This paper proposes GeoTab-CRS, a hybrid deep learning framework for intelligent crop recommendation that combines Graph Attention Networks (GAT) for spatial agro-climatic modeling with a tabular transformer encoder for soil-nutrient feature interaction learning. The model combines the soil nutrient data (Nitrogen, Phosphorus, Potassium), climatic parameters (rainfall, temperature), and the past data of crop yields to come up with sound recommendations that are regional specific. In contrast to traditional methods of predicting yields, this model focuses on specific advice on suitable crops, allowing farmers to make a good decision before the sowing period. There were positive experimental results, as Top-1 accuracy was over 96 per cent and Top-3 was nearly 99 per cent, which provides flexibility in the choice of farmers. Robustness tests indicated that the system was accurate above 94.5 per cent even in noisy and imbalanced circumstances. Stability was initially made sure through cross-validation with a variation of less than 0.3% in inter-fold variations, and the analysis was also conducted through ROC and SHAP in a way that offered interpretation; they related the recommendations to agronomic factors. An example is that rice is most suited with high nitrogen younger than 800–1200 mm' with 'between 800 and 1200 mm and high temperatures exceeding 33 °C. Regionally validated decision support tool with demonstrated robustness under noise and class imbalance within the AP–TS agro-climatic context. The findings demonstrate that GeoTab-CRS provides a practical, deployable decision support tool for agricultural planners and state advisory systems, achieving real-time inference at 3.2 ms/sample with uncertainty quantification that enables risk-aware crop recommendations.