Abstract: Accurately predicting cotton futures prices is challenging due to the difficulties in fusing multi-source heterogeneous data and the inefficiency in feature extraction. To address this, we propose a multimodal data fusion framework driven by a feature co-selection mechanism and a Bidirectional Long Short-Term Memory (BLSTM) network. This framework integrates multi-source information, including futures market indicators, remote sensing image features, and investor sentiment derived from textual data. The feature co-selection mechanism facilitates hierarchical dimensionality reduction, while the BLSTM captures nonlinear temporal dependencies. Our experiments validate the effectiveness of multi-source data fusion in elucidating the complex drivers of price fluctuations. Results demonstrate that our proposed method significantly enhances prediction accuracy, reducing the root mean square error (RMSE), mean absolute error (MAE), Mean absolute percentage error (MAPE), and normalized root mean square error (NRMSE) by 49.1%, 56.1%, 11.2%, and 14.5%, respectively, compared to a BLSTM model trained on non-integrated data. Further feature analysis reveals that historical futures prices, market sentiment indices, and remote sensing vegetation features all contribute to the prediction. This study offers novel insights for analyzing agricultural financial derivatives and provides an empirical reference for applying multimodal data in financial modeling.