KRISHI.AI: an explainable machine learning framework for data-driven crop recommendation

Abstract

Abstract Accurate crop selection is a fundamental determinant of farm productivity under varying soil characteristics and climatic conditions. Most current AI-based advisory systems are confined to predictive outputs or generic recommendations, providing little transparency into the reasoning behind a particular crop suggestion. This paper presents KRISHI.AI (derived from the Hindi word “krishi,”meaning“ agriculture,” and“ AI”for“ artificial intelligence), an agriculture-focused explainable artificial intelligence framework for data-driven crop recommendation. The framework recommends suitable crops from seven agronomic and environmental inputs, namely nitrogen (N), phosphorus (P), potassium (K), soil pH, temperature, humidity, and rainfall. It integrates a LightGBM-based multiclass prediction model with SHAP-based feature attribution and agronomically constrained counterfactual “what-if” analysis to explain why a crop is recommended and how feasible changes in the input conditions may alter the recommendation. The system uses a gradient-boosted tree ensemble (LightGBM) to produce high-confidence crop recommendations across 22 crop classes, achieving 99.39% test accuracy and a macro-averaged F1-score of 0.9900. These predictions are supplemented by SHAP (SHapley Additive exPlanations)-based feature-attribution visualizations and counterfactual “what-if” analysis to explain how input factors influence recommendations and how manipulating field conditions modifies crop suitability. Quantitative evaluation of the counterfactual module across 100 test instances exhibits a validity rate of 94.60%, a mean L1 feature-change distance of 0.18, and a mean of 2.10 features modified per scenario-confirming the practical credibility and sparsity of the generated explanations. These capabilities are delivered through an interactive Streamlit web dashboard that provides intuitive real-time feedback, enabling farmers and agricultural practitioners to explore scenario-based decisions without technical expertise. By integrating high-accuracy prediction with human-focused explainability in a single lightweight framework (4.3 MB model, 187 ms response latency), KRISHI.AI advances accessible and trustworthy decision support for technology-driven agriculture, with particular relevance to agricultural decision-support applications serving India’s smallholder farming community. The 189.3 ms value corresponds to the dedicated ablation profiling configuration, whereas 187 ms represents the representative end-to-end latency of the deployed pipeline reported in the main computational profiling experiment. The reported accuracy should be understood as a benchmark performance rather than as evidence of field-ready predictive reliability, because the uniformly balanced dataset does not capture the measurement noise, missing values, spatial heterogeneity, and seasonal variability present in real-world agricultural environments. However, the explainability pipeline and architectural contributions provide a reproducible basis for transparent agricultural AI.