UD-ConvoNet: Novel Architecture for Crop Yield Estimation Using MODIS Remote Sensing Multi-Spectral Data

The precise estimation of crop yield is essential to ensure food security and make informed decisions in agriculture. However, existing methods often struggle with the accurate interpretation of remote sensing (RS) data due to its high-dimensional and noisy nature, which hinders reliable yield prediction. In this paper, we introduce a novel architecture, UD-ConvoNet, for estimating crop yields using MODIS multi-spectral RS data. The UD-ConvoNet architecture consists of upsampling and downsampling layers, enabling the extraction of significant features and the reduction of noisy data for effective crop yield estimation. The effectiveness of the proposed UD-ConvoNet architecture in estimating soybean and corn yields has been evaluated, using 13 years of multi-spectral data across 31 states in the United States. We compared the performance of the UD-ConvoNet architecture with other deep learning (DL) models, including the convolutional neural network (CNN), the long-short-term memory (LSTM) network, and the state-of-the-art variant Vision Transformer (ViT). The results showed that UD-ConvoNet outperformed the other variants, with the R2 score ranging from 0.986 to 0.9892, the adjusted R2 score from 0.9664 to 0.9783, the Root Mean Square Error (RMSE) from 5.06 to 6.86, and the Root Mean Squared Logarithmic Error (RMSLE) in the range of 0.1101-0.1918 for the soybean crop. The UD-ConvoNet also performed well for the crop of corn, with the R2 score varying from 0.9719 to 0.9862, the adjusted R2 score from 0.9448 to 0.974, the RMSE from 17.5084 to 36.3605, and the RMSLE in the range of 0.128-0.3583.