Plant diseases have long been affecting global crop production and food security. For agricultural production, simply determining whether a plant is diseased is far from sufficient; accurately understanding the extent to which the disease has developed is equally important. Disease severity information can reflect the extent of crop damage and provide important references for developing disease control strategies, promoting the rational use of pesticides, and assessing crop yield losses. With the rapid development of artificial intelligence, image-based automated analysis has provided a new technical approach to improving the efficiency and objectivity of disease assessment.
However, accurately determining disease severity based solely on field images is not easy. Leaves affected by plant diseases are often surrounded by complex backgrounds involving soil, branches, other leaves, shadows, and variations in illumination. Overlapping leaves may make it difficult for models to identify the target leaf that actually needs to be analyzed, while some tiny lesions may occupy only a very small portion of the entire leaf. More importantly, diseased and healthy tissues sometimes have highly similar colors and textures, making it difficult for conventional vision models to accurately distinguish subtle lesion boundaries.
Recently, a team led by Professor Qi Wang from Guizhou University proposed a text-guided multimodal analysis framework called AgriMamba. It combines plant disease images with pathological descriptions of lesion color, shape, edges, and spatial distribution, and performs plant disease severity assessment through a coarse-to-fine process of “target leaf localization–fine-grained lesion segmentation–quantitative severity grading.”
Their study was made available online on August 5, 2026 and was published in The Crop Journal .
Explaining the motivation behind this research, Wang says, “Plant disease assessment in field environments is far more complicated than determining whether a leaf is diseased against a clean background. Background interference, leaf occlusion, and tiny lesions can all mislead the model. Therefore, we wanted to develop an analytical method that can not only ‘look at’ images, but also use pathological descriptions to understand which disease symptoms it should focus on.”
To achieve this goal, the research team designed the AgriMamba framework consisting of three stages and constructed a multimodal plant disease grading dataset containing 10,350 samples from 16 crop species. In the first stage, the localization-aware Mamba leaf segmenter first identifies the target leaf from complex environments and excludes irrelevant backgrounds as much as possible. In the second stage, the text-guided Mamba lesion segmenter further combines image features with pathological descriptions to more accurately locate diseased regions on the leaf. Finally, the system calculates the proportion of the target leaf area occupied by lesions and determines the corresponding disease severity grade based on this quantitative result.
“AgriMamba demonstrated competitive overall performance compared with multiple representative deep learning methods in plant disease severity grading,” Wang adds. “More importantly, the model maintained stable recognition performance when lesions were small, scattered, partially occluded, or surrounded by visually confusing background regions.”
In AgriMamba, textual information can be understood as a set of additional “diagnostic clues” provided to the model. When agricultural experts examine diseased leaves, they do not make judgments based only on the overall appearance of the leaf. “We also considered whether the lesions are yellow or brown, circular or irregular, sharply or diffusely bordered, and concentrated or scattered across the leaf surface,” Wang shares. “AgriMamba introduces similar pathological descriptions into the visual analysis process, enabling the model to use more explicit disease characteristics to distinguish visually similar healthy and diseased tissues.”
The researchers further examined which regions the model focuses on during the analysis process. Visualization experiments showed that, in the first stage of AgriMamba, the model’s main responses remain concentrated on the complete target leaf even when branches, soil, other leaves, and natural backgrounds are present in the image. Once the target leaf is identified, the regions of attention in the second stage gradually shift toward abnormal colors, texture changes, and specific lesion locations on the leaf surface. “This change intuitively demonstrates how the model gradually narrows its analysis range from the entire leaf to the regions where disease actually occurs,” Wang says. “We believe that combining visual information with standardized pathological descriptions provides a valuable approach to more reliable plant disease assessment.”
“Rather than asking the model to directly examine a complex field image and provide a severity grade, AgriMamba breaks the entire problem down into several easy-to-understand steps—first finding the leaf that needs to be analyzed, then identifying the lesions, and finally measuring what proportion of the leaf is affected by disease,” adds lead author Qinglong Wu.
The research team believes that this analysis strategy provides a new technical approach for automated plant disease assessment in complex environments. “With further improvements in model computational efficiency and practical field deployment, similar multimodal methods may be applied to automated crop scouting and smart agricultural management, helping agricultural producers respond to plant diseases in a more timely, efficient, and sustainable manner,” Wu adds.
Author contact:
Qi Wang (qiwang@gzu.edu.cn)
The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 300 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).
The Crop Journal
Imaging analysis
Not applicable
AgriMamba-guided multimodal framework with pathology-aware alignment for plant disease severity grading
The authors declare no competing financial or personal interests that could have influenced the work reported in this paper.