Hyperspectral imaging-driven deep learning approach: Asymptomatic stage detection and severity grading of tomato yellow leaf curl disease
Tomato yellow leaf curl disease (TYLCD), a devastating disease in tomato cultivation, causes reduced yields or even total crop failure in infected plants, severely compromising both productivity and quality. The prolonged asymptomatic infection period and non-distinct early symptoms of TYLCD impose...
Saved in:
| Main Authors: | , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S277237552500512X |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Tomato yellow leaf curl disease (TYLCD), a devastating disease in tomato cultivation, causes reduced yields or even total crop failure in infected plants, severely compromising both productivity and quality. The prolonged asymptomatic infection period and non-distinct early symptoms of TYLCD impose limitations on traditional detection methods, whereas hyperspectral imaging technology, capable of capturing rich spectral and spatial features, provides a technical foundation for early lesion identification and dynamic disease progression monitoring. This study proposes a spectral-spatial dual-branch residual network (SSDBRN) designed to fully extract and leverage complex features within hyperspectral images, aiming to achieve detection of TYLCD during its asymptomatic stage and classification of disease severity levels. Comparative experiments with three baseline models (2D CNN, 3D CNN, and HybridSN) demonstrated that SSDBRN achieved an early-infection detection accuracy of 84.2 % (5–7 days post-inoculation) and 96.2 % severity classification accuracy across six stages, significantly outperforming the baseline models. Furthermore, ablation studies validated the contributions of the key components of SSDBRN, namely the spectral channel attention module and the deformable convolution module, while generalization capability studies confirmed the model's robustness to complex backgrounds and lighting variations. Visualization analysis of spectral activation curves and spatial feature activation maps further elucidated the model's classification basis at both spectral and pixel-level dimensions. The findings establish that hyperspectral imaging integrated with SSDBRN enables non-destructive, early asymptomatic-stage detection and precise severity classification of TYLCD, providing an efficient diagnostic solution for the agricultural industry. |
|---|---|
| ISSN: | 2772-3755 |