DAFT-Net: A Hybrid Deep Learning Framework for Multi-Modal Agro-Environmental Data Fusion in Smart Farming

Smart farming as the central element of Agriculture 4.0, in combination with intelligent models, can provide the capacity to incorporate multi-source agro-environmental data to support an optimized and data-based decision-making process. Three significant contributions of the proposed research are as follows: To develop DAFT-Net. This new hybrid deep learning framework applies to multi-modal information fusion of agro-environmental data in smart farming. First, the DAFT-Net architecture combines the Dense Convolutional Networks (DenseNet), Bidirectional Gated Recurrent Units (Bi-GRU), and Cross-Attention Transformers to achieve good results in terms of capturing spatial, temporal, and cross-modal correlations through heterogeneous data sources, such as Sentinel-2 multispectral images, Soil Grids v2.0 soil data, ERA5 climate reanalysis, and on-field IoT sensor data. Second, the framework shows promising results in three important tasks of smart farming, that is, predicting crop yield (R2 = 0.957, RMSE = 0.173 t / ha), soil moisture (MAPE = 5.1%, NSE = 0.89), and stress classification (F1-score = 96.3%). These findings are significant in comparison with baseline models Convolutional Neural Networks-Long Term Short Memory (CNN-LSTM and LSTM-Transformer). Therefore, these findings attest to the soundness and adaptability of DAFT-Net over various climatic conditions. Third, a dual-level Explainable AI (XAI) module based on Shapely Additive explanations (SHAP) and Layer-wise Relevance Propagation (LRP) is also integrated into the model to increase the level of interpretability and agronomic trust. SHAP determined that Normalized Difference Vegetation Index (NDVI) and red-edge reflectance contributed most heavily to yield and stress models, whereas LRP identified areas of images with crop stress. All these contributions form DAFT-Net as a scalable and explainable data-driven and climate-resistant framework of agricultural decision-making.