Deep Learning Based Spatio Temporal Learning with Bio-Inspired Optimization for Smart Fertilizer Decision Making

Objectives: To propose a new model for robust smart fertilizer recommendation by integrating deep learning based spatiotemporal and bio-inspired optimization. The system addresses the limitations of existing models in capturing features in diverse farming zones to maximize crop yield and minimize nutrient loss. Methods: Deep Spatio Temporal learning ST-FusionFormer (STLF) is employed to capture spatial and temporal dependencies in agricultural data with the help of transformer-based attention modules. Multi-Objective Adaptive Firefly Differentiated Evolution (MO-AFDE) is used to optimize fertilizer type, learning rate, and timing of delivery subject to various constraints, i.e., crop yield enhancement, cost reduction, and environmental sustainability. Three benchmarked datasets (Crop Recommendation with 2480 samples for 22 crop labels, Fertilizer Prediction with 150 entries for 7 fertilizer class, and Crop Nutrient Database for 300 crops extracted from Kaggle Repository) are utilized in this model to frame a multi-modal structure combines soil nutrients (N, P, K, pH), meteorological features (Temperature, Rainfall, Humidity) and crop-specific nutrient demands. The performance of the proposed STLF-MOAFDE is evaluated using MATLAB and Python, and the results are compared with two baseline models, Lightweight-CNN and DAEN-MaskRCNN. Findings: The proposed STLF-MOAFDE is trained using a 5-fold CV approach. The promising results shows the superior performance of 94.5% accuracy, 93.4% precision, 94.8% recall, 93.2% specificity, 93.8 F1-Score, and 0.67 AUC-ROC. The model outperforms in addressing spatio-temporal dependencies and fertilizer optimization under dynamic agro-climatic conditions. Novelty: This hybrid integration of STLF-MOAFDE provides a novel and robust self-adaptive fertilizer recommendation system with high environmental impact and sustainability considerations. Unlike traditional single-objective systems, the proposed system delivers balanced and effective optimization suitable for real-time precision agriculture and smart farming systems. Keywords: Smart Agriculture, Spatio-Temporal Deep Learning, FusionFormer, Bio-Inspired Optimization, Precision Farming, Environmental Sustainability