Abstract
Precise detection of crop leaf damage is essential for real-time plant health monitoring and yield estimation. However, conventional deep learning models often exhibit poor generalization when deployed across varying species and complex, unstructured field environments. To address these limitations, we propose a new modeling paradigm that shifts from traditional task-specific training to foundation model adaptation. Specifically, we introduce a novel architecture integrating the DinoV3 foundation model with a Unet framework to achieve robust leaf lesion segmentation. By incorporating a Spatial Prior Module (SPM) and a Projection Module, our approach effectively bridges the gap between general-purpose pre-training and domain-specific requirements. Experimental results on coffee and black gram datasets demonstrate that this paradigm consistently outperforms standard networks, including Unet, Unet++, and SwinUnet. On the coffee leaf dataset, the proposed model achieves an Intersection over Union (IoU) of 78.31% and a Pixel Accuracy of 88.00%, surpassing the baseline Unet by over 10.5% in IoU. Remarkably, the architecture reduces inference time by approximately 93.6% (from 63.41s to 4.07s), proving that high-parameter foundation models can be adapted for extreme computational efficiency in agricultural scenarios. To further validate scalability, we conduct additional experiments on a larger dataset, AMG HS . The proposed paradigm achieves the best overall detection performance while maintaining superior computational efficiency, confirming its robustness under increased data scale. Interpretability analysis reveals that the foundation model backbone effectively captures high-level semantic features of lesions, providing a clear explanation for its superior performance and cross-domain reliability. This research establishes a scalable, high-performance paradigm for intelligent crop protection, demonstrating that coupling customized encoders with foundation models is a superior strategy for cross-domain agricultural tasks.
Abstract from the paper’s OpenAlex record; open copy licensed cc-by. Read at the source ↗
The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.
Claims 8
Claim · model-written summary
The proposed model outperformed Unet, Unet++, and SwinUnet on coffee and black gram datasets.
- Taxon
- coffee and black gram
Vigna mungo
Coffeae0 - Outcome
- outperforms standard networkse0
- Direction
- outperformse0
- Comparator
- Unet, Unet++, and SwinUnete0
Organisms pictured · Representative pictures of the matched species, not figures from the paper.
Verbatim
e0“Experimental results on coffee and black gram datasets demonstrate that this paradigm consistently outperforms standard networks, including Unet, Unet++, and SwinUnet.”
t0“leaf damage”
Claim · model-written summary
On coffee leaves the model achieved 78.31% IoU.
Organisms pictured · Representative pictures of the matched species, not figures from the paper.
Verbatim
e0“On the coffee leaf dataset, the proposed model achieves an Intersection over Union (IoU) of 78.31% and a Pixel Accuracy of 88.00%, surpassing the baseline Unet by over 10.5% in IoU.”
Claim · model-written summary
On coffee leaves the model achieved 88.00% pixel accuracy.
Organisms pictured · Representative pictures of the matched species, not figures from the paper.
Verbatim
e0“On the coffee leaf dataset, the proposed model achieves an Intersection over Union (IoU) of 78.31% and a Pixel Accuracy of 88.00%”
Claim · model-written summary
The model exceeded baseline Unet by over 10.5% in IoU.
Not reportedTaxon
Verbatim
e0“surpassing the baseline Unet by over 10.5% in IoU.”
Claim · model-written summary
The architecture reduced inference time by approximately 93.6%, from 63.41s to 4.07s.
Verbatim
e0“the architecture reduces inference time by approximately 93.6% (from 63.41s to 4.07s)”
Claim · model-written summary
On the larger AMG HS dataset the model had the best overall detection performance.
Not reportedComparator
Verbatim
e0“The proposed paradigm achieves the best overall detection performance while maintaining superior computational efficiency”
Claim · model-written summary
On the larger dataset the model maintained superior computational efficiency.
Not reportedComparator
Verbatim
e0“The proposed paradigm achieves the best overall detection performance while maintaining superior computational efficiency”
Claim · model-written summary
Interpretability analysis indicated that the backbone captures high-level lesion features.
- Outcome
- high-level semantic features of lesionse0
Not reportedComparatorValue
Verbatim
e0“Interpretability analysis reveals that the foundation model backbone effectively captures high-level semantic features of lesions”