High-throughput spike detection and refined segmentation for wheat Fusarium Head Blight in complex field environments

Abstract

Fusarium Head Blight (FHB) is a devastating disease of wheat worldwide. It is an explosive epidemic disease that can severely reduce or even fail wheat production. Estimating the disease ear rate and disease severity is crucial for effective plant protection. Manual assessment is labor-intensive and time-consuming. Accurately and quickly segmenting wheat ears and areas affected by Fusarium head blight (FHB) in complex field environments is essential for quantitative assessment of wheat trait phenotypes and FHB in wheat plants. This paper presents DeepFHB, an automated method for efficiently detecting, locating, and segmenting dense wheat spikes and diseased areas in digital images captured under natural field conditions. The experiment consists of three steps:Firstly, the process begins by generating initial coarse-grained mask predictions at lower resolutions to provide a rough segmentation. Secondly, a quadtree-based method is employed to identify and refine multi-scale inconsistent regions. Finally, a transformer-based refinement network is introduced to predict highly accurate instance segmentation masks. The results demonstrate that the DeepFHB algorithm outperforms traditional methods in detecting and segmenting diseased areas. Our DeepFHB model achieves state-of-the-art single-model results of 64.408 box AP and 64.966 mask AP on the FHB-SA dataset. This study is capable of rapidly and accurately segmenting wheat spikes and wheat scab lesions in agricultural scenarios with high field density, high crop occlusion, and high background interference. This provides a foundation for subsequent targeted research to assist agricultural workers in assessing the severity of wheat diseases.