Answers are grounded in this paper and cite their sources.
Accurate segmentation of pear leaf diseases under complex field conditions is severely hindered by blurred lesion boundaries, substantial scale variations, and background interference from leaf venation and illumination changes, posing critical obstacles to reliable automated disease monitoring in precision agriculture. To address these challenges, we propose an improved DeepLabV3+ architecture incorporating two additional modules. First, a Channel–Spatial–Frequency (CSF) attention module splits the feature channels into high- and low-frequency groups, a structural prior rather than an explicit transform, separating channels that tend to capture boundary detail from those that tend to capture regional structure so as to suppress orchard background noise while enhancing the discriminability of subtle disease patterns. Second, the Crop Lesion Adaptive Routing Kolmogorov-Arnold Network (CLARK) module integrates a soft lesion routing mechanism with a Chebyshev polynomial basis, enabling spatially adaptive, high-order non-linear feature transformations tailored to the heterogeneous visual appearances of plant diseases. Together, these modules produce feature representations suited to timely field diagnosis, supporting precision spraying and reducing unnecessary pesticide application. Experiments on a custom Pear Leaf dataset and the public CropProtection dataset evaluate the proposed method. Our model achieves a Mean Intersection over Union (mIoU) of 89.11 percent on the Pear Leaf dataset and 62.11 percent on the CropProtection dataset, outperforming the baseline and networks such as Mask2Former and UNet++. Crucially, our ablation shows that the two modules are complementary rather than independently beneficial: on the heterogeneous multi-species data, CSF and CLARK each fall below the baseline when used in isolation, and a consistent gain appears only when they are combined, so the proposed design should be understood as a coupled CSF–CLARK pair rather than two stand-alone, plug-and-play blocks. Under this setting, frequency-aware attention combined with spatially adaptive routing improves segmentation in field conditions where boundary ambiguity and illumination variation are the dominant sources of error, and supports downstream applications such as automated severity quantification and variable-rate spraying. The source code is publicly available at https://github.com/vitant-lang/CSF-CLARK-SEG .
No takes yet. Share an insight, caveat, or question.
GAO et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: