An Explainable Transformer-Based Deep Learning Framework for Crop Sustainability Prediction in Sustainable Agriculture

Sustainable agriculture necessitates data-intensive decision support systems capable of recommending crops aligned with local soil and climatic conditions. Predicting crop sustainability presents a significant challenge due to the complex, non-linear interactions among critical agronomic variables, including nitrogen, phosphorus, potassium, temperature, humidity, rainfall, and soil pH. Traditional fixed-rule guidelines and shallow statistical models often fail to capture these intricate relationships and typically offer limited interpretability, hindering their practical utility in agricultural decision-making. This paper introduces the Crop Sustainability Prediction Framework (CSPF), a transformer-based deep learning model engineered for accurate and calibrated multi-class crop sustainability forecasting. The CSPF integrates a transformer classification network designed to model complex dependencies between environmental variables. It

also incorporates an explainable feature attribution module, utilizing techniques such as SHAP and Integrated Gradients, to quantify the influence of each environmental factor on predictive outcomes. Furthermore, the framework includes a calibration-based reliability assessment, employing metrics like Expected Calibration Error (ECE) and Brier score, to evaluate prediction confidence. Evaluated on a crop recommendation dataset comprising 2,200 samples across 22 crop types, CSPF demonstrated high performance, achieving a classification accuracy of 0.982, a Macro-F1 score of 0.980, and a One-vs-Rest ROC-AUC of 0.996. The calibration results further indicated a low ECE, affirming the model's reliability.

Message: Hi FE, I need help with “An Explainable Transformer-Based Deep Learning Framework for Crop Sustainability Prediction in Sustainable Agriculture” in “Agricultural Engineering”