Smarter AI Spots Citrus Diseases Early in Cluttered Orchards

Citrus growers may soon have a sharper set of digital eyes in the orchard. A research team reporting in PLOS One has unveiled a lightweight artificial intelligence detector that can spot diseases and pests on citrus trees with markedly improved accuracy, even when early symptoms are small, faint, and buried in dense, cluttered foliage. The system, built on a modern object-detection architecture and refined with two purpose-designed feature-enhancement modules, raises detection precision and overall accuracy well above its baseline while still running fast enough for near real-time field monitoring.

The work addresses a stubborn problem in agricultural computer vision. Detecting a citrus disease in its early stages is one of the most valuable things a monitoring system can do, because timely intervention protects yield and limits the spread of pathogens and pests. Yet the very conditions that make early detection valuable also make it hard. Lesions on leaves and fruit are often tiny relative to the whole image. Early symptoms can be subtle in color and texture. And the natural backdrop of an orchard, with overlapping leaves, branches, variable lighting, and shadows, produces backgrounds that confuse standard detectors. Many existing models either miss small lesions or raise false alarms when foliage patterns resemble disease signs.

The new detector, described by Huiling Li, Xiaoxia Fang, Yufeng Shu, Yichun Zhan, and Jiankai Chen, is called YOLOv26-ACearly-ASPPdeep. The name encodes its two central innovations. It belongs to the YOLO family of one-stage detectors, which process an image in a single pass and are favored for speed. The researchers embedded a module they call AdaptiveConv, abbreviated ACearly, into the shallow, early blocks of the network, and they placed an ASPP-based multi-branch context module, called ASPPdeep, into the deeper backbone and neck stages of the architecture. The design philosophy is what the authors describe as a functional decoupling of feature enhancement across network depth: each module does the job it is best suited for at the scale where it matters most.

AdaptiveConv operates at the shallow end of the network, before the image has been aggressively downsampled. At that stage, the feature maps still retain fine spatial detail, which is exactly what small lesions require. The module strengthens content-aware filtering, allowing the network to emphasize the fine-grained textural and color cues that distinguish a developing lesion from healthy tissue or from look-alike blemishes. By inserting this capability early, the model preserves and sharpens the weak signals of incipient disease before the network compresses spatial information in later stages.

ASPPdeep works at the opposite end of the depth spectrum. Atrous spatial pyramid pooling, the technique behind the module, applies parallel convolutional branches with different dilation rates, letting the network sample context at multiple scales simultaneously. Placed in the deep backbone and neck stages, this multi-branch aggregation enlarges the diversity of receptive fields available to the detector. That matters for citrus pathology because lesions rarely come in uniform sizes or shapes; irregular, sprawling symptom patterns benefit from context aggregation that can integrate information over larger areas without losing resolution. The coupling of early fine-grained enhancement with deep multi-scale context is the core of the architecture’s accuracy gains.

A third refinement targets a common source of information loss. Standard downsampling in convolutional networks can discard detail that later stages need, particularly for small objects. The researchers incorporated SCDown modules, which reduce this loss as the network compresses spatial resolution. Together, the three components, AdaptiveConv in the shallow blocks, ASPP-based aggregation in the deep stages, and gentler downsampling, form a depth-aware pipeline tuned to the specific visual challenges of citrus surveillance.

The team evaluated the detector on a citrus disease and pest dataset drawn from Roboflow, containing 4,358 images spanning four classes: blackspot, canker, greening, and healthy trees. These are among the most consequential citrus threats. Citrus blackspot and canker cause direct fruit and tree damage, while greening, also known as Huanglongbing, is a devastating bacterial disease spread by insects that has reshaped citrus production in affected regions. A detector that can reliably separate diseased from healthy trees across these categories has obvious practical value for growers and for surveillance programs trying to contain outbreaks.

The quantitative results show consistent improvements over the baseline YOLOv26 model under identical training settings. Maximum precision rose from 0.75565 to 0.78032, an increase of 3.26 percent. Mean Average Precision at an Intersection over Union threshold of 0.5, a standard accuracy metric in object detection, climbed from 0.791 to 0.863. The model also outperformed a set of representative lightweight detectors, including YOLOv5n, YOLOv7-tiny, YOLOv8n, and YOLOv11n, across Precision, Recall, F1-score, and per-class Average Precision. These comparisons matter because the competing models are widely used in agricultural detection applications, so beating them on the same data and training regime is a meaningful benchmark rather than an abstract claim.

Speed remains a central consideration for orchard deployment. The proposed model processes an image in 15.1 milliseconds on an RTX 3090 graphics card, a rate that supports near real-time monitoring. That performance envelope suggests the architecture could run on field equipment, whether mounted on ground vehicles, drones, or fixed camera stations, without requiring heavyweight computing infrastructure. The authors characterize the result as a practical accuracy-efficiency trade-off, meaning the model gains accuracy without sacrificing the speed that smart citrus surveillance demands.

The broader significance lies in the design pattern the study demonstrates. Rather than bolting generic enhancements onto a detector, the researchers matched each module to a specific depth in the network and a specific visual challenge: fine-grained filtering early, multi-scale context late, and information-preserving downsampling throughout. This depth-aware coupling of AdaptiveConv and ASPP provides a template that could transfer to other crop-detection problems where small lesions, subtle symptoms, and cluttered backgrounds are the norm. As orchard monitoring systems evolve from experimental prototypes into working tools for yield protection and disease containment, architectures of this kind, accurate, lightweight, and fast, are likely to define the state of the practice.

Subject of Research: Improved YOLO-based deep learning model for early citrus disease and pest detection

Article Title: Improved YOLOv26-ACearly-ASPPdeep for citrus disease and pest detection

Article References: Li, H., Fang, X., Shu, Y., Zhan, Y., & Chen, J. (2026). Improved YOLOv26-ACearly-ASPPdeep for citrus disease and pest detection. PLOS One, 21(10), e0360634. https://doi.org/10.1371/journal.pone.0360634

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360634

Keywords: citrus disease detection, YOLO, object detection, deep learning, AdaptiveConv, ASPP, precision agriculture, plant pathology, computer vision, smart orchard monitoring, Huanglongbing, lightweight neural network