Improved YOLOv26-ACearly-ASPPdeep for citrus disease and pest detection
- 1Dongguan Polytechnic, Dongguan, China
Abstract
This study addresses early and accurate detection of citrus diseases and pests under complex orchard conditions, which is essential for timely intervention and yield protection. Existing detectors often struggle with small lesions, subtle early symptoms, and cluttered foliage backgrounds. We propose YOLOv26-ACearly-ASPPdeep, a lightweight one-stage detector that functionally decouples feature enhancement across network depth: AdaptiveConv (ACearly) is embedded in shallow C3k2 blocks to strengthen fine-grained, content-aware filtering before aggressive downsampling, while ASPP-based multi-branch context aggregation (ASPPdeep) is placed in deep backbone/neck stages to enlarge receptive-field diversity for irregular lesion morphology. SCDown modules further reduce information loss during downsampling. Experiments on a Roboflow citrus disease/pest dataset (4,358 images; blackspot, canker, greening, and healthy) show that the proposed model improves max precision from 0.75565 (baseline YOLOv26) to 0.78032 (+3.26%) and raises mean Average Precision at IoU = 0.5 ([email protected]) from 0.791 to 0.863. Under identical training settings, it also outperforms representative detectors including YOLOv5n, YOLOv7-tiny, YOLOv8n, and YOLOv11n in Precision, Recall, F1-score, and per-class AP, while remaining suitable for near real-time orchard monitoring (15.1 ms/image on an RTX 3090). The results indicate that depth-aware AdaptiveConv–ASPP coupling provides a practical accuracy–efficiency trade-off for smart citrus disease and pest surveillance.
Related
- IoT-RiceMobileNet: An improved lightweight MobileNetV2 Model for real-time multi-class rice disease detection using IoT
- Improved detection and spatiotemporal spectral analysis of neural traveling waves
- Deep learning-based detection of construction waste in complex scenarios with an improved lightweight algorithm
- Yolov8-CA-EYHL: A bridge crack detection algorithm based on improved YOLOv8
- Replication-Enhanced Detection of Quantitative Traits Evolving Adaptively (REDQuanTEA): an improved statistical framework to detect locally adaptive traits
BibTeX
@article{li2026improved,
title = {{Improved YOLOv26-ACearly-ASPPdeep for citrus disease and pest detection}},
author = {Huiling Li and Xiaoxia Fang and Yufeng Shu and Yichun Zhan and Jiankai Chen},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0360634},
url = {https://doi.org/10.1371/journal.pone.0360634},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Improved YOLOv26-ACearly-ASPPdeep for citrus disease and pest detection
AU - Huiling Li
AU - Xiaoxia Fang
AU - Yufeng Shu
AU - Yichun Zhan
AU - Jiankai Chen
PY - 2026
DA - 2026/10/15
JO - PLOS One
DO - 10.1371/journal.pone.0360634
UR - https://doi.org/10.1371/journal.pone.0360634
AB - This study addresses early and accurate detection of citrus diseases and pests under complex orchard conditions, which is essential for timely intervention and yield protection. Existing detectors often struggle with small lesions, subtle early symptoms, and cluttered foliage backgrounds. We propose YOLOv26-ACearly-ASPPdeep, a lightweight one-stage detector that functionally decouples feature enhancement across network depth: AdaptiveConv (ACearly) is embedded in shallow C3k2 blocks to strengthen fine-grained, content-aware filtering before aggressive downsampling, while ASPP-based multi-branch context aggregation (ASPPdeep) is placed in deep backbone/neck stages to enlarge receptive-field diversity for irregular lesion morphology. SCDown modules further reduce information loss during downsampling. Experiments on a Roboflow citrus disease/pest dataset (4,358 images; blackspot, canker, greening, and healthy) show that the proposed model improves max precision from 0.75565 (baseline YOLOv26) to 0.78032 (+3.26%) and raises mean Average Precision at IoU = 0.5 ([email protected]) from 0.791 to 0.863. Under identical training settings, it also outperforms representative detectors including YOLOv5n, YOLOv7-tiny, YOLOv8n, and YOLOv11n in Precision, Recall, F1-score, and per-class AP, while remaining suitable for near real-time orchard monitoring (15.1 ms/image on an RTX 3090). The results indicate that depth-aware AdaptiveConv–ASPP coupling provides a practical accuracy–efficiency trade-off for smart citrus disease and pest surveillance.
SN - 1932-6203
ER -