Depth-Aware Detection Sharpens Citrus Disease Recognition in Orchard Images
Early identification of citrus disease and pest injury depends on more than raw image classification. In orchards, symptoms often begin as faint lesions, scattered discoloration, or partially obscured damage hidden among dense foliage, and conventional detectors can miss those signals long before visible decline becomes obvious. The model tested here, YOLOv26-ACearly-ASPPdeep, was built around that problem: preserve fine-grained detail at the front end of the network, then expand contextual understanding deeper in the architecture where lesion shape and background clutter become harder to separate.
The design is straightforward in concept but deliberate in execution. AdaptiveConv was inserted into shallow C3k2 blocks as ACearly, where early feature extraction has the greatest influence on whether a small lesion is retained or blurred away during downsampling. Deeper in the backbone and neck, ASPPdeep applied multi-branch context aggregation so the detector could capture irregular lesion morphology over a wider receptive field. SCDown modules were also used to limit information loss during downsampling, an important detail when the target occupies only a small fraction of the frame.
Better Precision and mAP Without Sacrificing Speed
The dataset comprised 4,358 images drawn from Roboflow and included blackspot, canker, greening, and healthy citrus classes. Under identical training settings, the proposed model increased max precision from 0.75565 with baseline YOLOv26 to 0.78032, a gain of 3.26%. Mean Average Precision at IoU 0.5 rose from 0.791 to 0.863, a sizable jump for a detector meant to operate in visually noisy orchard scenes. The architecture also outperformed YOLOv5n, YOLOv7-tiny, YOLOv8n, and YOLOv11n on Precision, Recall, F1-score, and per-class AP.
The speed result matters because orchard surveillance is only useful when it can keep pace with field conditions. The system ran at 15.1 ms per image on an RTX 3090, which keeps it in near real-time territory for monitoring workflows. In practical terms, that means the detector is not merely more accurate; it also remains responsive enough to fit into inspection pipelines where early action can determine whether a localized problem becomes a larger yield threat.
A small lesion missed at the shallow layers is a missed intervention opportunity.
Why Early Feature Preservation Matters in Citrus Scanning
The network logic reflects a familiar challenge in agricultural vision: small, heterogeneous lesions are often the first sign of trouble, yet they are also the easiest to lose when a model aggressively compresses image information too early. Adaptive convolution in the shallow layers helps the model respond to local texture changes and content-specific edges, which is especially useful when symptoms are subtle or partially masked by leaf overlap, shadows, or motion blur. Once those details survive the early stages, deeper ASPP-based processing can integrate larger contextual patterns without collapsing the lesion into background noise.
That layered separation is especially relevant for citrus disease detection because the visual signatures are not uniform. Blackspot, canker, and greening do not present as identical objects, and healthy tissue can sit adjacent to diseased tissue in the same image. A detector that preserves fine detail while still understanding broader context has a better chance of distinguishing a true lesion from a visually similar artifact, such as leaf speckling, reflected light, or shadowing from the canopy.
The architecture also fits the operational reality of orchard surveillance. Field images are rarely pristine. Branches, fruit, leaves, and variable illumination create the kind of clutter that punishes detectors built only for clean laboratory imagery. By combining depth-aware feature enhancement with multi-scale context capture, the model attempts to bridge that gap between algorithmic performance and field usability.
Class-Specific Gains and the Cost of Missing Early Symptoms
Per-class improvements were reported across the citrus categories, with the model outperforming the comparison detectors in AP as well as overall precision and recall. Those class-level gains are important because disease surveillance is not only about whether a model can find “something wrong.” It must identify the specific condition accurately enough to guide management decisions, whether that means targeted treatment, further inspection, or removal of infected material.
In crop protection, delayed recognition tends to compound loss. Early-stage lesions can evolve into widespread canopy involvement, and pests often create secondary damage patterns that make later diagnosis more difficult. A detector that is better tuned to early-stage morphology can therefore do more than improve an abstract benchmark score; it can change the timing of intervention. In a real orchard, that timing affects spraying decisions, labor allocation, and the probability of limiting spread before symptoms become visually obvious across a larger portion of the planting.
The reported performance suggests that the model’s gains came from architecture rather than brute-force scale. That is a useful distinction in deployment settings where faster inference and lower computational burden are valued alongside accuracy. Lightweight detectors often trade away sensitivity to small or irregular objects; this model was designed to resist that compromise by placing the right type of computation in the right depth of the network.
What the Dataset Can and Cannot Prove
The study was performed on a Roboflow citrus disease and pest dataset, and the results establish a clear comparison against several widely used YOLO variants under the same training conditions. The evidence supports the internal claim that depth-aware AdaptiveConv and ASPP coupling improves detection quality while preserving speed. It does not, however, establish performance across every orchard environment, camera setup, lighting condition, or geographic disease distribution.
That boundary matters because agricultural vision models often behave differently once they leave curated datasets and enter messy field workflows. Variability in fruit maturity, cultivar, canopy density, weather, and sensor quality can alter lesion appearance in ways that are difficult to anticipate from a single benchmark. Real-world deployment will need validation across broader orchard settings before the detector can be treated as a general surveillance tool rather than a strong proof of concept.
Even so, the core result is concrete: early-stage feature preservation and deeper context aggregation improved the detector’s ability to recognize citrus disease and pest patterns without slowing inference beyond practical use. For image-based orchard monitoring, that combination is exactly where the operational value lies—finding small problems early enough that they can still be managed.