The breeding of high-yield wheat varieties is needed to ensure food security. Accurately and rapidly predicting wheat yield at the plot level via UAVs would enable breeders to identify meaningful genotypic variations and select superior lines, thus accelerating the selection of climate-adapted high-yield varieties. Although current prediction models have already utilized multivariate time series data, these models usually adopt a simple concatenation operation to embed all the raw data, resulting in low prediction accuracy. To address these limitations, we propose an improved transformer-based wheat yield prediction model with a variate-independent tokenization approach. The proposed variate-independent tokenization approach facilitates the embedding of 14 vegetation indices and 28 morphological traits via the feature dimension, enabling the learning of variate-centric representations. We also apply a multivariate attention mechanism to evaluate the contribution of each variate and capture the multivariate correlation. Extensive experiments are conducted to verify the effectiveness of our model, including comparisons across 3 nitrogen treatments, 2 years, and 56 wheat varieties. We also compare our model with state-of-the-art approaches. The experimental results indicate that our model achieves the optimal prediction performance, with an R² of 0.862, surpassing those of the classical recurrent neural network and transformer variants. We also confirm that combining both the vegetation indices and morphological traits is advantageous over using single-source data for the prediction task, achieving an approximately 4 ​% prediction performance gain. In conclusion, this study provides a novel approach for utilizing an improved transformer model and multivariate time series data to quantitatively predict plot-level wheat yield, thus enabling the rapid selection of high-yield varieties for breeding.

Asian soybean rust (ASR) is one of the major diseases that causes serious yield loss worldwide, even up to 80%. Early and accurate detection of ASR is critical to reduce economic losses. Hyperspectral imaging, combined with deep learning, has already been proved as a powerful tool to detect crop diseases. However, current deep learning models are limited to extract both spatial and spectral features in hyperspectral images due to the use of fixed geometric structure of the convolutional kernels, leading to the fact that the detection accuracy of current models remains further improvement. In this study, we proposed a deformable convolution and dilated convolution neural network (DC²Net) for the ASR detection. The deformable convolution module was used to extract the spatial features, while the dilated convolution module was applied to extract features from the spectral dimension. We also adopted the Shapley value and the channel attention methods to evaluate the importance of each wavelength during decision-making, thereby identifying the most contributing ones. The proposed DC²Net can realize early asymptomatic detection of ASR even when visual symptoms have not appeared. The results of the experiment showed that the detection performance of DC²Net dominated state-of-the-art methods, reaching an overall accuracy at 96.73%. Meanwhile, the experimental result suggested that the Shapley Additive exPlanations method was able to extract feature wavelengths correctly, thereby helping DC²Net achieve reasonable performance with less input data. The research result of this study could provide early warning of ASR outbreak in advance, even at the asymptomatic period.