Accurate detection of full-surface ear rot in maize using hyperspectral imaging and deep learning

Maize ear rot severely restricts maize yield and quality, making the breeding of disease-resistant varieties the core strategy for disease prevention and control. Due to the highly uneven spatial distribution of lesions on maize ears, precise full-surface detection is essential for objectively quantifying disease severity. However, traditional manual disease grading is highly subjective, and conventional RGB-based detection methods struggle to precisely identify lesion regions associated with maize ear rot. These limitations hinder the precise identification and quantitative analysis of maize ear rot infection regions, thereby limiting the reliability of phenotypic data used for resistance evaluation and subsequent genome-wide association studies (GWAS). To address these challenges, this study developed an integrated full-surface hyperspectral imaging system featuring line-scan imaging and synchronous rotation control. Non-redundant full-surface ear images were then generated using the oriented FAST and rotated BRIEF (ORB) algorithm combined with random sample consensus (RANSAC), hereafter referred to as ORB-RANSAC. Furthermore, after Savitzky-Golay (SG) preprocessing and feature selection using a genetic algorithm (GA), three machine learning models and three deep learning models were established, and their classification performance was compared. The results showed that the convolutional neural network-bidirectional long short-term memory network (CNN-Bi-LSTM) model achieved the best average performance, with an average overall accuracy (OA) of 95.61 ± 0.36%. It also achieved higher overall accuracy than traditional machine learning models such as random forest (RF), indicating that CNN-Bi-LSTM can achieve high-precision pixel-level detection of lesion regions showing Fusarium-associated maize ear rot symptoms. Additionally, this model was deployed in locally developed automatic analysis software, enabling an integrated analysis workflow from raw hyperspectral data input to the quantification of disease-related phenotypic parameters. This study not only fills the technical gap in the non-destructive full-surface detection of maize ear rot but also provides an efficient and reliable automated tool for high-throughput phenomics research, which holds great significance for accelerating the discovery of maize resistance genes and ensuring food security.