Adaptive Graph Enhanced Spatio Temporal with Multi-Objective Reinforcement Learning (AGST-MOORL) for Dynamic Fertilizer Optimization

Abstract

To propose an advanced framework AGST-MOORL to improve fertilizer recommendation accuracy under dynamic agricultural conditions. The primary objective is to overcome the limitations of earlier models by incorporating adaptive learning of spatial relationships, temporal variations, and multi-objective decision optimization. The methodology integrates Adaptive Graph-Based Spatio-Temporal Learning (AGST) with Multi-Objective Reinforcement Optimization (MOORL). Crop Recommendation Dataset with 2480 samples consisting of 22 crop labels, Fertilizer Recommendation Dataset with 150 samples containing fertilizer class for 7 fertilizers and Crop Nutrient Database with 300 crops entries are three datasets used to build our consolidated multi-modal model. They collectively provide us with data about nutrients present in the soil (N, P, K, pH), weather data (taking temperature, humidity and rainfall into account) and nutrients required by each crop. Feature engineering techniques such as Nutrient Deficiency Index (NDI), Rainfall-Temperature Interaction (RTI), and Fertilizer Efficiency Coefficient (FEC) are applied to enrich the data representation. A graph-based structure encodes relationships among samples, while temporal attention mechanisms learn dynamic changes in agricultural conditions across time. The reinforcement learning component further refines recommendations by optimizing trade-offs among yield, cost, and environmental impact. The performance of the proposed AGST-MOORL is assessed using MATLAB and Python for detailed comparative analysis with two baseline models such as Lightweight-CNN, DAEN-MaskRCNN and STLF-MOAFDE. Experimental evaluation demonstrates significant improvement over the previous phase of works. The proposed model achieves 97.2% accuracy, 96.5% precision, 97.0% recall, 96.1% specificity, and 96.7% F1-score, with an enhanced AUC-ROC of 0.78, outperforming existing methods across all metrics. Additionally, convergence speed is improved, and decision stability is enhanced under varying agro-climatic conditions. The proposed work stands as a unified integration of graph-based learning, temporal modeling, and reinforcement-driven multi-objective optimization within a single framework.