Authors: Deepa Barethiya, Dhanashri Pannase, Gangasagar Kashyap
Abstract: Ensemble machine learning has pushed crop recommendation accuracy past 99% on standard soil-weather benchmarks — yet this milestone conceals a troubling gap. Systems built around Random Forest, XGBoost, and gradient boosting produce ranked crop labels while leaving the economic viability of each suggestion entirely unexamined. A farmer told "grow rice with 99% confidence" still does not know whether that choice will leave a positive margin after seed, fertiliser, and irrigation costs. This paper proposes a decision-oriented framework that moves beyond the accuracy plateau by coupling a soft-voting ensemble with per-crop yield regressors and a configurable economic layer that estimates expected profit. Where conventional pipelines terminate at a suitability label, the proposed architecture extends the output to a Risk-Adjusted Expected Profit, mathematically formulated as E[Π_c ]_(risk-adjusted)=P_ensemble (c│X) Π_c, where P_ensemble (c│X)is the Ensemble Suitability Probability and Π_c=((Y_c ) ̂(X)×P_(market,c)×1000)-Total Cost_cis the Nominal Net Profit. This coupling mathematically discounts the apparent value of high-risk crops by their probability of soil-weather failure — a correction absent from every reviewed system. To illustrate the theoretical decision dynamics of this framework, we construct a conceptual walkthrough across 200 hypothetical soil-weather scenarios derived from standard agricultural benchmarks. This analysis suggests that the agronomically top-ranked crop and the economically top-ranked crop diverge in roughly 46% of cases — a finding that, if borne out in empirical deployment, would have direct implications for farm-level income planning. A conceptual Streamlit dashboard design is also proposed, embedding real-time what-if sliders and SHAP-based feature attributions to make the system transparent to extension workers and farming cooperatives. The central argument of this paper is simple: a classifier that ignores profit is only half a tool. This framework proposes the other half.
