Wide-area Crop Yield Prediction Based on Multi-Source Remote Sensing Data Fusion and Attention Net

The development of remote sensing technology has facilitated the prediction of crop yield over large areas, and more researchers are using multi-spectral geographic data of the growing season to mine the characteristics of crop growth and harvest. However, most of the existing methods do not make full use of dimensionality reduction embedding technology and spatial heterogeneity. In this paper, we use multi-spectral time-series images of crop growth process, combined with non-time-series regional images such as elevation and slope, and learn from better data embedding methods to predict yield by convolutional neural network and self-attention network. We evaluated the method in the county soybean yield forecasting task in the United States, and the experimental results showed the effectiveness of the method.