Abstract
Crop productivity is heavily impacted by inefficient fertilizer usage, improper fertilizer handling, and inappropriately chosen crops. To address these issues, this research work proposes an AI-powered Smart Agriculture Prediction System utilizing intelligent agents for carrying out soil classification, estimating soil parameters, crop suggestion, and fertilizer suggestion. The soil classifier module is trained with 1,563 images of black soils, red soils, clay soils, and alluvial soils using MobileNet-V2, ResNet, and Custom CNN, with Custom CNN resulting in a higher accuracy of 92.88%, which performed better in classifying soils based on textures. A soil parameter estimation agent utilizes regression models for estimating pH and NPK content of soils using images. For crop suggestion, a crop dataset with 2,200 samples with parameters such as N, P, K, T, H, pH, and rainfall is used, in which Random Forest model performed better with an accuracy of 92.4% when compared with CNN and DNN models. For fertilizer suggestion, XGBoost performed better with an accuracy of 94.7% in estimating fertilizers such as Urea, DAP, NPK, Potash, and Compost. Real-time climatic parameters are obtained using API in order to make dynamic updates for climatic parameters. Real-time weather data obtained through APIs enables dynamic updates of climatic parameters, while Explainable AI techniques such as SHAP and LIME enhance model transparency and user trust. Additionally, the system incorporates an interactive agent-based framework that processes user inputs, including location, soil images, and nutrient levels, to generate adaptive outputs such as weather alerts, yield potential, and personalized recommendations. The experimental results demonstrate that the proposed system effectively integrates deep learning, ensemble learning, and explainability to deliver a scalable, efficient, and sustainable decision-support solution for precision agriculture, promoting optimized resource utilization and environmental stewardship.