Precision Agriculture System

A two-stage chained machine learning pipeline that dynamically recommends the optimal crop and matching fertilizer based on environmental and soil metrics.

• End-to-End Precision Agriculture System: Developed a two-stage chained machine learning pipeline that dynamically recommends the optimal crop and matching fertilizer based on environmental and soil metrics.
• State-of-the-Art Deep Learning: Built and optimized tabular deep learning models using PyTorch TabNet (TabNetClassifier), achieving a 99.81% accuracy on crop prediction (Stage 1) and a 93.61% accuracy on fertilizer recommendation (Stage 2).
• Advanced Preprocessing & Data Imbalance Handling: Applied SMOTE (Synthetic Minority Over-sampling Technique) to resolve target class imbalances and employed MICE Imputation (IterativeImputer) to robustly handle missing features.
• Chained Inference Pipeline: Engineered a unified inference function that scales, imputes, and encodes raw data on-the-fly, passing predicted crop parameters directly into the fertilizer recommendation stage for unified decision making.
• Comprehensive Performance Evaluation: Evaluated models using multi-class ROC-AUC/Precision-Recall curves, confusion matrices, and early-stopping criteria to guarantee high prediction accuracy and strong generalization.