Framework for Adaptive Multimodal Learning to Provide Interpretable and Climate-Aware Crop Recommendations in Precision Agriculture

Agriculture increasingly needs adaptive, transparent, and data-driven decision-support systems to sustain productivity in the face of intensifying climate variability and soil decline. Conventional crop recommendation methods (Decision Trees, Random Forest, SVM, XGBoost) are based on static descriptions of soils and thus do not capture temporal climate dynamics and multimodal interactions; similarly, earlier deep learning solutions (Bi-LSTM, GRU, XAI-CROP) enhance pattern recognition but lack robustness in handling heterogeneous inputs and actionable transparency. In order to address these limitations, this work introduces a Perceiver IO–based multimodal crop recommendation framework that combines soil chemistry and agronomic features, long-term climate time series data, and Sentinel-2 image time-series into a single latent representation. Perceiver IO encoder uses cross-attention to synchronize heterogeneous modalities, stacked latent transformer layers retain higher-order temporal and cross-modal relations, and a decoder generates probabilistic crop suitability scores. In order to ensure future adaptation to changes in season and climate, Elastic Weight Consolidation (EWC) is added to avoid catastrophic forgetting in the incremental updates. Transparency is realised by means of SHAP-based feature attributions and counterfactual analyses that explain how soil nutrients, rainfall, temperature and spectral indices contribute to the recommendations. Implemented in TensorFlow and PyTorch and evaluated on the Crop Recommendation Dataset, the Indian Agriculture and Climate Dataset (1961–2018), and the Sentinel-2 Image Time Series for Crop Mapping, the framework attains R² = 0.992, RMSE = 0.028, and MAE = 0.019, outperforming unimodal baselines. Ablation and continual-learning experiments demonstrate the benefits of multimodal fusion and EWC for resilient, interpretable, and field-deployable crop recommendation, advancing climate-aware precision agriculture. Proposes a Perceiver IO–based multimodal framework that integrates soil, climate, and Sentinel-2 imagery for comprehensive crop recommendation. Employs cross-attention fusion and continual learning (EWC) to ensure adaptability across seasonal and climatic variations without catastrophic forgetting. Demonstrates high interpretability and transparency through SHAP-based feature attribution and counterfactual visual analytics. Achieves state-of-the-art performance with R² = 0.992, RMSE = 0.028, and MAE = 0.019, outperforming conventional and deep learning baselines. Advances climate-resilient, explainable, and data-driven decision support for sustainable and intelligent precision agriculture.