Predicting crop yields accurately is crucial for maintaining food security and promoting sustainable farming methods, especially in areas with highly variable climates like India. Conventional yield prediction models are less accurate under changing climatic circumstances because they frequently rely on static historical data and miss dynamic environmental changes. Using time-series deep learning models that incorporate multi source data, such as weather factors, soil properties, and historical crop yield records, this paper suggests a climate-adaptive crop yield prediction framework. The suggested method uses Transformer-based architectures and Long Short-Term Memory (LSTM) to identify long-term patterns and temporal connections in agricultural data. In order to improve prediction resilience under uncertain weather conditions, a dynamic adaptation mechanism is integrated to update the model with real-time climatic inputs.According to experimental results, the suggested model performs better in terms of RMSE and MAE metrics than conventional machine learning methods like random forest and linear regression. The study emphasizes how deep learning may help farmers make data-driven decisions and enable precision agriculture. The suggested framework is a viable option for climate-resilient agriculture since it is scalable and adaptable to various crops and geographical areas.