Scholarlink

Abstract <p>Early blight caused by <i>Alternaria solani</i> remains a major constraint to sustainable potato production, often managed through repeated prophylactic fungicide applications. This study develops a UAV-based, multi-feature machine-learning framework for early blight detection and precision spray zoning using low-cost RGB imagery under realistic field conditions. Ultra-high-resolution UAV images from two growing seasons (2019 and 2022) were analyzed using a publicly available, expert-annotated dataset. In addition to conventional RGB vegetation indices, including the Excess Green Index (ExG) and the Visible Atmospherically Resistant Index (VARI), a Chlorophyll–Water Content Index (CWCI) was evaluated as a stress-sensitive trait and integrated with RGB statistical features, LAB color-space descriptors, and texture metrics. Five machine-learning classifiers were compared using stratified five-fold cross-validation. Ensemble-based models outperformed linear approaches, with Random Forest achieving the most stable performance (mean ROC-AUC ≈ 0.79; accuracy ≈ 0.72). Feature-importance analysis revealed that texture and color-space features were dominant predictors, while CWCI ranked among the most informative spectral traits, demonstrating its complementary value within a multi-feature framework rather than as a standalone index. Classifier outputs were further translated into risk-based spray zones to evaluate agronomic relevance. Simulated variable-rate spraying scenarios indicated that restricting fungicide application to moderate- and high-risk zones could reduce treated area by approximately 44–50% compared with blanket spraying, while maintaining conservative disease coverage. Overall, the results demonstrate that RGB UAV imagery, when combined with biologically informed feature engineering and ensemble learning, can support spatially explicit disease risk mapping and decision-support for integrated pest management. The study highlights both the potential and limitations of RGB-based UAV systems for early blight monitoring and underscores the importance of multi-feature modeling for robust, interpretable, and practically relevant disease detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

UAV-Based Multi-Feature Machine Learning Framework for Early Blight (Alternaria solani) Detection and Precision Spray Zoning in Potato

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Shefali Vinod Ramteke,
  • Pritish Kumar Varadwaj

摘要

Abstract

Early blight caused by Alternaria solani remains a major constraint to sustainable potato production, often managed through repeated prophylactic fungicide applications. This study develops a UAV-based, multi-feature machine-learning framework for early blight detection and precision spray zoning using low-cost RGB imagery under realistic field conditions. Ultra-high-resolution UAV images from two growing seasons (2019 and 2022) were analyzed using a publicly available, expert-annotated dataset. In addition to conventional RGB vegetation indices, including the Excess Green Index (ExG) and the Visible Atmospherically Resistant Index (VARI), a Chlorophyll–Water Content Index (CWCI) was evaluated as a stress-sensitive trait and integrated with RGB statistical features, LAB color-space descriptors, and texture metrics. Five machine-learning classifiers were compared using stratified five-fold cross-validation. Ensemble-based models outperformed linear approaches, with Random Forest achieving the most stable performance (mean ROC-AUC ≈ 0.79; accuracy ≈ 0.72). Feature-importance analysis revealed that texture and color-space features were dominant predictors, while CWCI ranked among the most informative spectral traits, demonstrating its complementary value within a multi-feature framework rather than as a standalone index. Classifier outputs were further translated into risk-based spray zones to evaluate agronomic relevance. Simulated variable-rate spraying scenarios indicated that restricting fungicide application to moderate- and high-risk zones could reduce treated area by approximately 44–50% compared with blanket spraying, while maintaining conservative disease coverage. Overall, the results demonstrate that RGB UAV imagery, when combined with biologically informed feature engineering and ensemble learning, can support spatially explicit disease risk mapping and decision-support for integrated pest management. The study highlights both the potential and limitations of RGB-based UAV systems for early blight monitoring and underscores the importance of multi-feature modeling for robust, interpretable, and practically relevant disease detection.