DeepFarm-YieldNet: A Novel Integrated CNN-BILSTM-Attention Framework for Real-Time Soil Moisture Prediction and Crop Yield Forecasting in Smart Farming

Abstract

Smart farming has gained increasing attention recently as one of the most efficient approaches towards achieving sustainability in meeting the challenges to global food security brought about by factors such as climate change, water scarcity, and population growth. Real-time prediction of soil moisture status and its impact on the yield of the crops grown in the field is crucial for precision irrigation, effective utilization of resources, and climate-resilient farming operations. In this research study, DeepFarm-YieldNet is proposed as a novel deep-learning framework which is able to perform simultaneous prediction of multi-layer soil moisture dynamics and seasonal crop yield based on multi-source data inputs. The architecture integrates 1D-CNN for spatial feature representation of the data gathered through sensors and satellite observations along with BiLSTM and self-attention networks. The proposed model was carefully validated against a real-world data set derived from 12 farms located in the Pune and Nashik regions of Maharashtra, India, during the 2023–2025 cropping seasons and consisting of sugar cane, soybean, and cotton crops. The outcomes show the best results, where soil moisture prediction is conducted with an accuracy of R² = 0.95 and RMSE = 0.014, while crop yield prediction gives R² = 0.93 and RMSE = 1.8 q/ha, which is significantly better than the benchmark CNN-LSTM and Transformer-based algorithms by 12–18%. In addition, the suggested approach helps to achieve up to 28% savings in water consumption by optimizing the process of irrigation management. Using explainable AI methods (SHAP and LIME), the developed application can provide explanations to the farmers via the mobile app.