Crop Yield Prediction Based on Reconstructive Independent Component Analysis and Bag-of-Visual-Words Model

Crop yield prediction over large areas plays an important role in agricultural planning and management. With the advancement of remote sensing technology, an increasing number of researchers have started utilizing multispectral geospatial and temporal data for crop yield prediction. However, existing methods often fail to fully utilize dimensionality reduction embedding techniques and the spectral characteristics of the data. In this paper, we propose a new approach for crop yield prediction that leverages multispectral remote sensing images and incorporates advanced data embedding methods along with convolutional neural networks. We evaluate the effectiveness of our proposed method on predicting soybean yield at county level, and experimental results demonstrate its superiority over existing approaches, e.g., on United States dataset the proposed method outperformed the C3D+LSTM by 3.8%, and has higher prediction accuracy compared with LSTM+GP, CNN+GP and other methods.