FHBNet: A severity level evaluation model for wheat Fusarium head blight based on image-level annotated aerial RGB images

Shuxin Zhu¹Huayong Li²Shun Zou¹Huanliang Xu¹Zhaoyu Zhai^1*

• ¹Nanjing Agricultural University, Nanjing, China
• ²Jiangsu Academy of Agricultural Sciences (JAAS), Nanjing, Jiangsu Province, China

A leading concern for global wheat production, Fusarium head blight (FHB) can cause yield losses of up to 50% during severe epidemics. The cultivation of FHB-resistant wheat varieties is widely acknowledged as a highly effective and economical approach to disease management. The disease resistance breeding task depends on accurately evaluating the severity level of FHB. However, existing approaches may fail to distinguish among healthy and slightly infected wheats due to insufficient fine-grained feature learning, resulting in unreliable predictions. To tackle these challenges, this paper proposed the FHBNet model for evaluating the severity level of FHB under an end-to-end manner by simply using image-level annotated RGB images. In total, 6035 RGB aerial images taken from the wheat field were used to construct the dataset and each image was labelled by the light, moderate, or severe category. In FHBNet, we first utilized the multi-scale criss-cross attention (MSCCA) block to capture the global contextual relationships from each pixel, thereby modelling the spatial context of wheat ears. Furthermore, in order to accurately locate small lesions in wheat ears, we applied the bi-level routing attention (BRA) module, which suppressed the most irrelevant key-value pairs and only retained a small portion of interested regions. The experimental results demonstrated that FHBNet achieved an accuracy of 79.49% on the test set, surpassing the mainstream neural networks like MobileViT, MobileNet, EfficientNet, RepLkNet, ViT, and ConvNext. Moreover, visualization heatmaps revealed that FHBNet can accurately locate the FHB lesions under complex conditions, e.g., varying severity levels and illuminations. This study validated the feasibility of rapid and nondestructive FHB severity level evaluation with only image-level annotated aerial RGB images as an input, and the research result of this study can potentially accelerate the disease resistance breeding task by providing high-throughput and accurate phenotype analysis.