Pixel-level regression for UAV hyperspectral images: Deep learning-based quantitative inverse of wheat stripe rust disease index

Previous research on utilizing unmanned aerial vehicle (UAV) remote sensing imagery for plant disease detection has predominantly focused on the qualitative identification of healthy and infected plants. Notably, pixel-level regression analysis for the quantification of the wheat stripe rust disease index (DI) using hyperspectral imaging data and deep learning methods is still lacking. Traditionally, quantitative inversion has been achieved by employing radiative transfer model inversion techniques or a combination of vegetation indices and machine learning methodologies. This investigation presents an end-to-end, pixel-level quantitative regression methodology, underpinned by deep learning techniques. This methodology carries substantial importance not only for the accurate assessment of disease distribution maps, but also for an array of common quantitative regression challenges within agricultural systems. For example, the approach can be utilized for the regression inversion of continuous phenotypes, including crop yield and plant height. In this study, 1,560 local wheat varieties (lines) from Henan Province were selected as experimental subjects, resulting in a wide-ranging gradient of DI. Hyperspectral images at a height of 100 m were obtained based on UAVs at different stages of infection. This work utilized a deep learning semantic segmentation method with continuous loss functions such as Laplacian loss to achieve pixel-level regression and end-to-end quantitative inversion of the DI. The performance of models with different loss functions, model architectures and datasets was compared. The optimal results were achieved using the Laplacian + MSE loss function combined with the HRNet_W18 algorithm model, yielding a test set R2 value of 0.875 and an MSE of 0.0129. Incorporating a PSA module further improved the outcomes, resulting in an R2 value of 0.880 and an MSE of 0.0123. Modeling with a limited number of feature indices (e.g., six feature indices) reduced the model recognition performance to 0.829 compared to full-band modeling. These findings suggest that full-band, end-to-end modeling based on deep learning algorithms can lead to superior inversion outcomes and streamline data analysis steps. The insights from this research hold relevance for high-throughput crop phenotyping, plant disease and pest monitoring, and quantitative yield assessment.