Rice (Oryza sativa) is the main source of food to more than 3.5 billion people and is the economic backbone of the agricultural communities in South and Southeast Asia. Pest control in India is a national food security issue as over 43 million hectares of rice are grown. Examples of insects that result in yield loss of up to 20-80 percent, depending on the extent of infestation, the crop developmental stage, and access to prompt treatment, include brown planthopper, yellow stem borer and leaf folder.

The traditional method of pest monitoring (in decades) was manual field scanning by trained agronomists, which is labour-intensive, subjective and cannot be replicated across the extensive paddy fields of rural Asia. Farmers in most instances end up on blanket application of pesticides without proper identification of the pest species concerned, thus wasting money and degrading the environment. A camera-based, automated, detection system, capable of discerning 26 species of rice pests in real-time, based on field images, would radically redefine this image.

The fast progress of deep learning has already given rise to a number of detector families that might be used to meet this requirement. Single-stage YOLO-based models are capable of providing real-time inference with high accuracy; transformer-based detectors such as RT-DETR are capable of strong attention-based feature modelling; two-stage models such as Faster R-CNN are capable of high localisation accuracy at the expense of speed; and large vision-language models such as Florence-2 are capable of zero-shot generalisation without task-specific training. Nevertheless, none of the studies has directly compared the four families on a multi-source, 26-class rice pest, large-scale benchmark under the same experimental conditions.

This gap is covered in this paper. We make the following contributions:

(1) We compile a single 26-class rice pest benchmark of 11, 319 images across four Roboflow sources.

(2) You can refer to our proposed YOLOv8n+CBAM+BiFPN that incorporates the channel-spatial attention as well as bi-directional multi- scale fusion into the YOLOv8n backbone.

(3) We perform a strict six-model comparison with the same training conditions.

(4) We document the surprising result that BiFPN achieves less inference latency than the YOLOv8n baseline, because it has depthwise separable convolutions.

(5) We verify that, with no domain-specific adaptation, the fine-grained rice pest taxonomy can not be inferred at all using the zero-shot version of Florence- 2.

Our dataset was built by merging four rice pest datasets sourced from the Roboflow platform: rice-pest- bb (ds1), Rice Pest (ds2), rice pest disease detection (ds3), and rice pest detection 4 (ds4). All images were converted to YOLO bounding-box annotation format during the merge process. After de-duplication and format normalisation, the merged dataset contains 8,546 training images and 2,773 validation images—11,319 images in total across 26 pest and disease classes. No separate test split was designated; model performance is therefore reported on the validation split.

The 26 classes are: brown-planthopper, green- leafhopper, leaf-folder, rice-bug, stem-borer, whorl- maggot, paddy stem maggot, rice gall midge, rice hispa, rice leaf hopper, rice leaf roller, rice plant hopper, rice stem borer, rice thrips, rice water weevil, Bacterial Leaf Blight, Brown Spot, Dirty Panicle, Narrow Brown Spot, Rice Blast Disease, Rice Leafhopper, asiatic rice borer, brown plant hopper, rice leaf caterpillar, small brown plant hopper, and yellow rice borer.

It is notable that some names of classes are shared between source datasets, such as brown-planthopper and brown plant hopper are considered different classes to maintain fidelity of the annotation to its original sources. This overlap is recognized as a limitation of the dataset and is probably one of the factors that led to the relatively small mAP50 scores of all models. Fig. 1 shows the estimated split of classes in the training split.

All models were evaluated on the validation split (2,773 images) using five metrics: Precision, Recall, mAP@50 (mean Average Precision at IoU ≥ 0.50), mAP@50–95 (averaged over IoU thresholds 0.50–0.95 in 0.05 steps), and inference latency in milliseconds per image. Inference latencies were measured on a 640×640 dummy input tensor after 10 warmup passes, averaged over 100 timed runs on CUDA. Table I presents the complete results.

Proposed model. Faster R-CNN early-stopped at epoch 70. All speeds: RTX 3060, 640×640 input, 100 timed runs on CUDA.

The proposed model scores the highest precision among all YOLO-based models at 0.5888 and runs faster than every other model tested at 2.40 ms per image. Compared to the plain YOLOv8n baseline, it is 2.5 times faster while its mAP@50 drops by only 0.02 (from 0.4894 to 0.4694). In short, we got a faster and more precise model than the one we started from — which is not the usual outcome when adding modules to a detector.

The addition of CBAM to YOLOv8n without BiFPN did not change the precision much, but reduced recall (0.5187 to 0.4887) and mAP50 (0.4894 to 0.4494). This

informs us that attention by itself, without the appropriate multi-scale fusion, can lead to the model overlooking detections which it would have otherwise detected. Adding BiFPN on top of CBAM fixed this: recall went back up to 0.4957 and mAP@50 recovered to 0.4694. The two elements are true necessities to work together.

RT-DETR has the highest precision of any model at 0.6563, but its mAP@50 of 0.3744 is the lowest among all the neural network models. High precision with low recall means the model is being overly cautious — it only predicts a box when it is very sure, so it misses a lot. With around 329 training images per class on average, there simply is not enough data for a transformer to learn 26 distinct pest classes reliably in 100 epochs.

Faster R-CNN stopped training at epoch 70 with a mAP@50 of just 0.0963. The likely culprit is the batch size of 4 — with so few images per update, the gradient estimates are too noisy for the model to converge properly on a 26-class problem. This is a useful finding in itself: two-stage detectors need much larger batches to work well on fine-grained agricultural datasets, which in practice means they need significantly more GPU memory than most field-deployment scenarios can provide.

Florence-2 got zero precision and zero recall across all 200 test images, taking 180.72 ms per image — 75 times slower than our proposed model. Asking a general- purpose vision-language model to identify a paddy stem maggot or rice hispa from a text prompt alone does not work. These are specialist terms that barely appear in general web data. Without showing the model examples of what these pests look like, it has no chance of identifying them correctly. This result is an important warning for anyone planning to use large vision-language models for agricultural detection without task-specific fine-tuning.

Two main findings stand out from this study. First, combining CBAM and BiFPN with YOLOv8n gives you a faster and more precise model than the baseline — without sacrificing meaningful accuracy. The BiFPN neck deserves special attention here: most people would expect adding components to slow a model down, but BiFPN actually speeds things up because its depthwise separable convolutions are lighter than the standard ones in PANet. This means replacing PANet with BiFPN could be a useful approach for speeding up any YOLOv8-based model, not just in pest detection.

Second, neither the transformer model (RT-DETR) nor the zero-shot foundation model (Florence-2) could match a basic fine-tuned YOLO model on this task. RT- DETR needs more training data per class than is available here. Florence-2 needs domain-specific examples it has never seen. In practical terms, if you are building a rice pest detection system today, a fine-tuned YOLO-family model is still your best option — not a larger, slower general-purpose model.

The 0.02 mAP@50 gap between the proposed model and the baseline is worth explaining honestly. The merged dataset has class name overlaps — for example, brown-planthopper from one source and brown plant hopper from another are treated as different classes even though they refer to the same insect. A more discriminative model like ours is more affected by this kind of label noise than a simpler baseline, which may explain why the mAP gap exists even though our model performs better on clean distinctions.

This paper presented a systematic and comprehensive comparison of six object detection architectures for automated rice pest detection, evaluated under identical conditions on a unified 26-class benchmark of 11,319 images assembled from four Roboflow sources. The proposed YOLOv8n+CBAM+BiFPN architecture demonstrated that careful integration of attention mechanisms and efficient multi-scale fusion can simultaneously improve detection precision and inference speed relative to the baseline—an outcome that is not commonly achieved when adding complexity to a neural network.

At the core of our approach, CBAM provides the backbone with the ability to dynamically focus on pest- discriminative features while suppressing the dominant green foliage background that characterises rice field imagery. BiFPN then ensures that these attention- enhanced features are propagated and fused effectively across all three detection scales (P3, P4, P5), using bidirectional connections and learned fusion weights. Crucially, BiFPN’s use of depthwise separable convolutions makes the entire neck more computationally efficient than the standard PANet neck it replaces—explaining the counterintuitive result that the proposed model is faster than the YOLOv8n baseline it is built upon.

In addition to the proposed model per se, this work contributes to the agricultural detection community more generally by offering the most direct multi-family comparison of a large scale rice pest taxonomy to date. Our findings support the fact that single-stage YOLO- family models are still the best option in this category of problem: they stabilize on the given data, extrapolate to the validation dataset, and provide inference rates that can be used in practice by deploying a drone or smartphone in real-time. Models that are based on transformers have potential but need significantly more data per class to achieve it. The current state of zero-shot foundation models, such as Florence-2, are not suitable to the fine-grained taxonomic separation required by rice pest monitoring, an observation that the community must not ignore before such models can be used in production agricultural systems.

Regarding the practical deployment implication, the 2.40 ms inference time of the proposed model implies that it will be capable of operating at a rate of more than 416 frames/s on an RTX 3060 GPU. Real-time inference at the drone video inference rates (25 to 30 fps) is completely feasible even on more modest edge hardware. This makes the model a true contender to be integrated into precision agriculture systems where it is needed to have pests detected promptly and spatially in order to implement targeted pesticide application and minimize both costs and environmental footprint.

There are a number of directions which would be natural extensions of this work. To begin with, the merging process of the databases added the noise of labeling classes at the level of classes by having near-duplicate names of classes in the sources (e.g., brown-planthopper and brown plant hopper). The close re-annotation exercise to fix these overlaps into a clean 2022 class taxonomy would probably enhance mAP 50 in all models and give a cleaner point of reference to compare to in future. Second, substituting the standard CIoU bounding box regression loss with Wise-IoU or Shape-IoU should enhance localisation accuracy of small-bodied pests including rice thrips and whorl maggots. The loss functions are tailored to address imbalanced regression challenge between simple and challenging examples, which is especially pertinent when the dataset has a large variation in the sizes of pest bodies.

Third, the fact that Florence-2 failed completely in zero- shot mode does not rule out its usefulness as a few-shot or fine-tuned model. Specific study of domain-adapted prompting techniques, or a few-shot visual fine-tuning of Florence-2 on representative rice pest images, would clarify whether big vision-language models can be usefulized to this domain with limited labelled data. This direction is becoming increasingly similar in relation to the multimodal foundation models that have been developed at a rapid pace. Fourth, we tested all models with a constant 640×640 input resolution. A resolution ablation experiment at 416x416, 640x640, and 800x800 inputs would help understand the tradeoff between small- object recall and inference speed in this particular pest taxonom, and can help inform configuration decisions in various deployment scenarios (e.g., fixed camera traps versus UAV video streams). Lastly, the practical deployment case presented in this paper and the practical implementation of the proposed model on real edge hardware, like a Jetson Nano or a Raspberry Pi 5 with an AI accelerator, and actual field conditions would confirm that the proposed model is practically deployable and offer the community directly actionable benchmarks against which agricultural edge computing can be assessed.