Agentic AI for Autonomous Soil and Fertilization Management for Agriculture Sustainability

Abstract

Soil fertility loss and excessive chemical fertilization are major environmental and economic issues in developing regions such as Punjab, Pakistan. This paper proposes an Agentic AI framework for autonomous soil and fertilization management that combines (i) IoT soil sensing and drone-based crop monitoring for real-time perception, (ii) predictive modelling for short-horizon nutrient and moisture forecasting, and (iii) multi-agent reinforcement learning (MARL) for adaptive decision-making. The system operates with operational autonomy, executing daily management decisions without routine human-in-the-loop control. Agronomic expert knowledge is incorporated only offline as safety constraints and initialization priors (e.g., allowable nutrient ranges and stress-avoidance rules) to bound the action space and prevent unsafe behavior, rather than to prescribe actions. Experiments were conducted across two seasons at two sites (Sheikhupura and Multan) under four treatments: Farmer Practice (FP), Rule-Based Control (RBC), Machine Learning Predict (ML-Predict), and the proposed Agentic AI. Results show that Agentic AI reduces nitrogen fertilizer use while maintaining/improving yield proxy and improving soil indicators (including residual nitrate reduction and improved Soil Health Index). We also analyze irrigation outcomes as a sustainability objective and show how water usage must be treated as a constrained or multi-objective term in the reward function to avoid over-irrigation. Overall, the framework supports scalable, data-driven soil management with bounded autonomy, preserving expert-defined agronomic safety.