Interpretable deep learning models for independent fertilizer and crop recommendation

The integration of Artificial Intelligence (AI) and Internet of Things (IoT) technologies is fundamentally transforming agriculture. This integration enables the real-time collection and analysis of critical data such as soil nutrient levels, temperature, humidity, and climatic conditions. A data-driven approach allows for independent fertilizer and crop recommendations, maximizing yield and promoting resource efficiency. Soil properties play a crucial role in crop growth, facilitating the absorption of essential nutrients like Nitrogen (N), Phosphorus (P), and Potassium (K). However, traditional farming practices often lead to nutrient inefficiency and soil health degradation, negatively impacting productivity. While Machine Learning (ML) techniques have emerged for recommending fertilizers and crops, they frequently encounter issues such as inadequate feature selection and class imbalance. These limitations hinder their ability to accurately model the intricate relationships among environmental factors, soil conditions, and crop nutrient requirements. To address these challenges, this study presents a unified smart recommendation system that utilizes TabNet, a deep learning architecture specifically designed for tabular data. The novelty of this work lies in leveraging TabNet's attention-driven learning to directly discover important patterns from preprocessed IoT-enabled agricultural data for accurate and interpretable crop and fertilizer classifications, without relying on prior feature selection. To enhance transparency, SHapley Additive exPlanations (SHAP) is applied at the final stage to provide post hoc interpretability, allowing stakeholders to understand the model's reasoning behind each recommendation. Evaluated with the Crop and Fertilizer Dataset from Western Maharashtra, TabNet achieves impressive classification accuracies of 95.24% for fertilizer recommendations and 96.21% for crop recommendations, outperforming conventional classifier and existing approach. The study employs robust preprocessing techniques, such as iterative imputation and the Synthetic Minority Oversampling Technique (SMOTE), to ensure data quality and class balance. The model's performance and generalizability were rigorously assessed using fivefold cross-validation, consistently maintaining these high accuracy levels across folds.