Three-channel Recognition Model Based on Visible Light Images for Crop Disease Incubation Stage
- VernacularTitle:作物潜育期病状可见光图像三通道识别模型
- Author:
Hao PANG
1
;
Yan ZHANG
1
Author Information
- Publication Type:Journal Article
- Keywords: deep learning; color features; visible light images; incubation period; early blight
- From: Progress in Biochemistry and Biophysics 2025;52(10):2650-2662
- CountryChina
- Language:Chinese
- Abstract: ObjectiveEarly detection of crop diseases is crucial for effective agricultural management and yield protection. While visible light imaging has been widely applied for disease detection due to its accessibility and non-destructive nature, most existing methods primarily focus on identifying diseases during the symptomatic phase, when visual symptoms are already prominent. However, detecting plant diseases during the incubation period—when symptoms are still subtle or invisible—remains a major challenge due to the lack of distinctive visual cues and limited research methodologies. This study aims to address this gap by proposing a novel three-channel recognition model to accurately identify early blight symptoms during the incubation stage in Solanaceae crops, particularly in chili and tomato, using only visible light images. MethodsWe established a controlled experimental setup in which healthy leaves and leaves inoculated with early blight pathogens were photographed continuously over time. A total of 1 258 visible light images were collected, capturing various stages of disease progression. From these images, lesion regions were manually annotated. To quantitatively characterize early and subtle color changes within the lesion areas, we extracted color moments—first-order (mean), second-order (standard deviation), and third-order (skewness)—from multiple color spaces, including Lab and HSV. By analyzing the temporal variation of these color moments across disease progression stages, we identified the first-order moment of the saturation (S) channel in the HSV color space as the most sensitive indicator of lesion development on inoculated leaves. Using this insight, we defined four disease categories: healthy, incubation stage, early stage, and late stage. Subsequently, a three-channel classification model was constructed by integrating features from three color channels that provided complementary information. Three-channel models were constructed based on R-G-B, L-a-b, and H-S-V color spaces, respectively, to evaluate performance across different crops and to determine which color representation provides the most discriminative power for identifying disease symptoms during the incubation period. ResultsThe proposed models demonstrated strong classification performance. The three-channel model built using the Lab color space achieved a 94.44% accuracy in recognizing the incubation stage of early blight in pepper, effectively distinguishing subtle pre-symptomatic features from healthy tissue. The model based on the HSV color space achieved 100% accuracy in detecting incubation-stage symptoms in tomato, underscoring the discriminative power of S-channel variations in this context. These results confirm the model’s capability to identify early blight before visible lesions become pronounced, which is essential for timely disease intervention. ConclusionThis study presents a new technical pathway for early-stage disease detection using visible light images by focusing on subtle color feature changes during the incubation period. The proposed three-channel recognition model effectively identifies early blight in both chili and tomato, offering a non-destructive, low-cost, and easily deployable solution for early warning and precision agriculture. Furthermore, this framework can be generalized to other crops and diseases where early detection plays a critical role in minimizing yield losses and ensuring sustainable production. The method lays a solid foundation for future research in pre-symptomatic plant disease recognition and provides valuable tools for intelligent crop monitoring and precision management systems.
