RiceDetect-Net: a lightweight real-time detection framework for rice diseases

Abstract

Rice disease detection is vital for food security, prevention efficiency, pesticide reduction, and sustainable agriculture. Challenges like poor model applicability, low accuracy, and limited datasets make this research essential. Existing models face issues with large parameters, complex computations, and insufficient semantic information capture. This paper introduces a large rice disease dataset and proposes the RiceDetect-Net model to address these challenges. The model integrates a brand-new lightweight detection head LE-Head to reduce the parameter quantity and computational complexity. To boost accuracy, the model integrates the newest FCA attention mechanism into its high-level semantic processing component, strengthening its capacity to interpret complex semantic data. Testing on a custom rice disease dataset comprising 54,240 images, the model attained 94.3% accuracy with a parameter count of 2.32 M. The enhanced model achieves a 0.4% increase in accuracy while reducing parameters by 10% relative to the baseline YOLOv11. The detection model is more lightweight, can adapt to the computing power of field detection equipment, is more suitable for practical scenario applications, and provides technical support for the development of smart agriculture.

Introduction

Rice is a major staple crop in China and holds a critical position in the nation’s agricultural production. China, as the global leader in rice production, cultivates the crop across extensive regions under varied climatic conditions. However, it also faces severe threats from various rice diseases [1]. In recent years, diseases such as rice blast, sheath blight, and false smut have occurred with increasing frequency, exhibiting widespread distribution, rapid propagation, and substantial damage, posing significant challenges to agricultural productivity [2]. Therefore, accurately identifying the spatial distribution and temporal dynamics of rice diseases across different growth stages is of paramount importance. This approach provides a scientific basis for effective disease management, while cutting excessive pesticide use and environmental pollution. It also boosts control efficiency, secures stable high grain yields, and advances the sustainable development of agriculture [3].

Traditional rice disease identification mainly relies on manual judgment by plant protection experts, but there are problems such as low efficiency, high misdiagnosis rate, and limited identification range. Current control methods, such as chemicals, are effective but inhibit seedling growth, and chemical residues threaten human health and the ecological environment [4]. As the market demand for organic agricultural products grows, reducing the use of chemical pesticides has become a trend, highlighting the importance of early and accurate identification. Applying computer technology to disease identification improves prevention and control accuracy and reduces pesticide dependence. This is critical to raising rice yields, ensuring agricultural product safety and advancing agricultural sustainability [5].

Currently, image-based crop disease recognition methods are divided into two types mainly based on the differences in feature extraction techniques: one type utilizes traditional machine learning algorithms, and the other type uses deep learning methods for detection and classification. Regarding the application of machine learning on rice diseases, VK Shrivastava et al. [6] proposed a color feature-based rice disease classification method, which achieved a classification accuracy of 94.65% by extracting 172 features in 14 color spaces and using a Support Vector Machine (SVM) classifier, thus providing an effective solution for the automated detection of rice diseases. B Patel et al. [7] proposed an adaptive learning algorithm based on machine learning, which improves the overall accuracy of crop disease detection by selecting the appropriate classification process, achieving 99.2% accuracy, which is superior to methods such as SVMs. S. Ramesh [8] developed a machine learning-based algorithm that uses image processing and feature extraction for the automatic detection of rice blast disease. This method achieved 99% accuracy for rice blast-infected images and 100% accuracy for healthy images during the training phase, and 90% and 86% accuracy, respectively, during the testing phase. While machine learning has advanced rice disease recognition, it remains constrained by dependencies on manual feature extraction and limited adaptability to varied environments [9]. This makes its performance in complex contexts less than ideal. Deep learning has the ability of automatic feature extraction, which can effectively learn and extract multi-level features from images without human intervention [10, 11]. Deep learning models are highly adaptable to complex environments and can handle more diverse and complex disease images, thus improving the accuracy and robustness of recognition [12, 13]. Beyond rice, deep learning excels in disease detection and classification for other crops. For apple leaf diseases, CNN models were evaluated, with EfficientNetV2_m achieving 100% accuracy and F1 score [14]. For grape leaves, fine-tuned CNN and Vision Transformer models including Swinv2-Base reached 100% accuracy on two datasets [15]. In tomato leaf disease detection, ResNet family models like Res2Next50 attained 99.85% accuracy [16]. For sugarcane leaves, EfficientNet-b6 and DeiT3-Small performed effectively with 93.39% and 93.79% accuracy respectively, with no direct link between model complexity, depth and accuracy [17, 18]. In summary, deep learning is versatile and high-performing, supporting precise multi-crop disease control.

Consequently, deep learning holds greater promise for rice disease identification, delivering more precise diagnostics when handling diverse pathogens and complex field conditions. Pan et al. [19] RiceNet model successfully realized the recognition of four major rice diseases, namely, rice scurf-neck blight, rice curculio, rice leaf blight, and rice straw blight, by combining the two-stage approach of YOLOX and Siamese Network. Four major rice diseases were efficiently detected and accurately recognized, with a mAP of 95.58% for YOLOX in the detection stage and 99.03% for Siamese Network in the recognition stage. CR. Rahman et al. [20] introduced a deep learning approach to detect rice pests and diseases, fine-tuning large-scale CNNs such as VGG16 and InceptionV3 to attain 93.3% accuracy. They also designed a compact CNN for mobile devices, cutting model size by 99% relative to VGG16. Deng et al. [21] developed an ensemble approach incorporating DenseNet-121, SE-ResNet-50, and ResNeSt-50 sub-models for automated diagnosis, enabling efficient detection of five rice diseases—leaf blight, curculio, spike-neck blight, stalk blight, and fine-stripe blight—with 91% overall accuracy. A Haridasan et al. [22] introduced an automated method for detecting and classifying rice diseases using computer vision. This method combines image processing, machine learning, and deep learning techniques to effectively detect five common rice diseases: bacterial leaf blight, rice curculio, brown spot, rice blast, and rice stalk rot. In the validation phase, it attained a peak accuracy of 91.45%.

Recent studies from 2024 to 2025 have further advanced deep learning-based rice disease detection technologies. In a 2025 study, Li et al. [23] developed the RDRM-YOLO model to detect rice diseases in complex field conditions, attaining a 93.5% mAP while maintaining a compact model size of 7.9 M parameters. Huang et al. (2025) [24] introduced the GDS-YOLO model for rice disease identification, which attained a mAP of 85.3% while utilizing only 2.32 M parameters. Zhou and Wei (2025) [25] introduced the YOLO-DP model for detecting 15 common rice diseases and pests, achieving a mAP50 of 77.8%. Zhang et al. (2024) [26] presented the YOLO-CRD model, designed for rice disease detection in natural environments, achieving a mAP of 90.2%. Regarding DETR-based methods, Song et al. (2025) [27] designed the ADAM-DETR model to detect rice diseases, attaining 94.76% mAP on the RiDDET-5 dataset while containing 14.3 M parameters. Yang et al. (2023) [28] proposed the DHLC-DETR model for rice disease detection, which achieved a mAP of 93.5%, but with a high parameter count of 46 M. Notably, among the YOLO-based rice disease detection models published in 2025, most exhibited moderate mAP performance, with some even falling below 90%, and there is still room for optimization in terms of parameter size. On the other hand, DETR-based models typically contain 14.3 M to 46 M parameters, resulting in larger model sizes that hinder their application on devices with limited resources, despite achieving competitive mAP scores. Existing YOLO-based agricultural variants and attention-enhanced crop disease models suffer from an imbalance between lightweight performance and fine-grained detection, as well as a lack of targeted attention mechanisms, leading to inadequate practical field performance and an urgent need for a lightweight model with targeted fine-grained detection capabilities. To tackle key challenges in rice disease identification like scarce samples, high model complexity, and significant inter‑class similarity, this study presents an efficient lightweight detection algorithm built on an improved YOLOv11 architecture. It is specifically developed to accurately recognize nine typical rice diseases, such as rice blast, sheath blight, bacterial leaf blight, bacterial streak, brown spot, and rice false smut.

This paper makes the following key contributions:

1. A Comprehensive Rice Disease Dataset: Aiming at the scarcity and incompleteness of existing rice disease data, this study constructs a dataset covering 9 typical rice diseases and 1 healthy category. Unlike many datasets that rely solely on online-sourced images, our dataset combines field-collected images from Huzhou, Zhejiang, with curated online resources, providing a robust and diverse foundation for training. Table 2 presents the total sample size for each disease category, ensuring transparency and reproducibility for future research.

2. A Lightweight Detection Head (LE-Head) for Edge Deployment: To address the computational complexity of existing YOLOv8/YOLOv10 agricultural variants and RTMDet-based lightweight detectors, we propose the Lightweight Enhanced Detection Head (LE-Head). By redesigning the classification and regression branches, it significantly reduces computational load while maintaining detection accuracy, making it uniquely suited for resource-constrained devices like plant protection UAVs, a limitation that plagues current state-of-the-art models.

3. An Adaptive Fine-Grained Channel Attention (FCA) Module for Higher Accuracy: Concerning the low accuracy of existing attention-enhanced crop disease models, we introduce the FCA module. Unlike computationally heavy attention mechanisms (e.g., CBAM, SE-Block), FCA efficiently transforms fine-grained details into discriminative semantic features, enabling more precise disease type classification. This provides robust technical support for early detection and accurate pesticide use, and outperforms standard YOLO variants in suppressing complex field background interference. The entire research process is illustrated in Fig. 1.

Fig. 1.

Real-time detection scheme for rice diseases based on YOLO and computer vision systems

Materials and methods

Dataset

The image data were primarily collected using multiple mobile phones of different models in rice fields located in Huzhou, Zhejiang Province, as well as through online image searches. Comprising 54,240 manually curated original images in total, the dataset consists of 40% field-collected samples and 60% online-sourced images from public repositories and professional agricultural databases, which are strictly used for non-commercial academic research in full compliance with fair use principles. Based on ongoing seasonal monitoring, the research team gathered authentic images that document each stage of rice’s complete growth cycle. These images were captured under natural light, featuring varied backgrounds, complex field conditions, and different illumination intensities. Such environmental variety provides a comprehensive dataset that strengthens the detection model’s robust and generalizable nature.

To diversify the image dataset, this study gathered data through both offline field collections and online sources. Usable images were manually selected by research team members and then uniformly resized to 640 × 640 pixels. As a result, a rice disease dataset was constructed, covering nine distinct categories: Rice blast, Bacterial leaf blight of rice, Rice leaf spot, Dead heart disease, Downy mildew, False smut of rice, Sheath blight, Stripe disease and Tungro disease, along with an additional ordinary category. Fig. 2 illustrates on-site data collection by the researchers, while Fig. 3 presents sample images from the rice disease dataset.

Fig. 2.

On-site data collection by researchers

Fig. 3.

Rice disease image dataset. a Rice blast (b) Bacterial leaf bligh (c) Brown spot (d) Bacterial foot rot (e) Downy mildew (f) False smut (g) Sheath blight (h) Stripe disease (i) Tungro disease

The overall framework of rice disease detection model RiceDetect-Net

The intelligent and precise identification of crop diseases is pivotal to advancing precision agriculture and green plant protection. Given rice’s status as a staple food crop, early and accurate identification of its diseases is critical for formulating effective prevention strategies and minimizing yield losses. Modern computer vision technology, especially object detection, lays the technical groundwork for automated disease diagnosis using field images. After years of development, object detection has formed two core paradigms: Two-stage and Single-stage. Classic two-stage detectors (e.g., R-CNN series) prioritize accuracy by generating candidate regions before classification and localization, but suffer from high computational cost and low efficiency [29]. To meet the real-time demands of agricultural field diagnosis and monitoring, this study adopts the YOLO framework [30], a single-stage detector that enables end-to-end target localization and identification, offering superior speed ideal for mobile or edge device deployment. We use YOLOv11 [31], which is optimized based on YOLOv3 [32] and YOLOv5, as our baseline model. Its key architectural advancements are threefold: first, replacing the C3 module with the optimized C3K2 module (CSP architecture), which reduces computational load while preserving feature expression capability; second, integrating the C2PSA attention mechanism to focus on key image regions dynamically; third, adopting multi-scale feature fusion for accurate detection of targets of varying sizes. These improvements enhance YOLOv11’s efficiency and accuracy, laying a solid foundation for agricultural scenario-specific optimization.

Aiming at key challenges in rice disease intelligent detection, including small lesion targets, leaf occlusion and high inter-disease visual similarity, this study selects YOLOv11n as the backbone network. It conducts architectural transformation while balancing the model’s real-time performance and edge device deployment efficiency. This choice aims to make it adaptable to edge computing platforms with limited computing power. To effectively break through the limitations of the basic model in detection accuracy and computational efficiency for small lesions, this study proposes and implements two core structural innovations: first, in terms of model head design (Detection Head), the traditional detection head structure with independent parameters is abandoned, and a lightweight detection module based on a parameter sharing mechanism is innovatively constructed and introduced. This design substantially cuts model parameters and lowers computational resource use during inference, enhancing the model’s overall inference speed. Second, in terms of strengthening feature extraction capabilities, focusing on the relatively weak high-level semantic information modeling capability of deep networks, a new Frequency Domain Channel Attention mechanism is introduced into the deep layer of the model. By performing frequency-domain analysis on feature maps and dynamically tuning channel weighting, FCA selectively amplifies channels containing discriminative disease information while suppressing non-essential background noise. This markedly strengthens the model’s capacity to detect subtle lesions under complex field conditions. Figure 4 illustrates the complete architecture.

Fig. 4.

Overall architecture of the model. This architecture integrates a lightweight detection head based on parameter sharing and a Frequency Domain Channel Attention mechanism to improve the accuracy and efficiency of small lesion detection and meet the application requirements of resource-constrained edge computing platforms

Lightweight enhanced detection head

Although existing disease detection technologies have achieved remarkable progress in theory, they still face many challenges in practical deployment, especially in agricultural application scenarios where the actual effectiveness is less than ideal. There are few deployment cases of existing rice disease detection models, mainly due to the following reasons: First, conventional models typically depend on significant computational resources and high-performance hardware, requirements that are often incompatible with the limited capabilities of low-resource devices like edge computing systems in real-world agricultural settings. Second, factors such as lighting changes, complex backgrounds, and the diversity of different disease characteristics in agricultural environments make many traditional models perform unstably in complex field conditions. In conclusion, although many models perform well in controlled laboratory environments, they often face difficulties in balancing detection accuracy with computational efficiency when deployed in real-world scenarios. This often results in prohibitively high hardware costs and limits their broader adoption in farmland monitoring scenarios.

This study introduces the LE-Head, a module designed specifically for the RiceDetect-Net model to tackle these challenges. This detection head deeply integrates the design concepts of cutting-edge models such as RTMDet [33], takes structural simplification and feature sharing enhancement as the core optimization directions, and innovatively introduces Detail-Enhanced Convolution (DEConv) technology to inject explicit gradient prior knowledge into the model. Inspired by differential convolution, DEConv parallelly deploys Horizontal Differential Convolution (HDC), Vertical Differential Convolution (VDC), and ordinary convolution, integrating local gradient descriptors into the convolutional layer to significantly enhance the model’s ability to capture fine-grained disease features, such as lesion edges and texture variations. Its core lies in the explicit encoding of gradient prior information: HDC and VDC respectively learn the horizontal and vertical gradient features of images, and after fusing with the features of ordinary convolution, form a multi-dimensional detail-enhanced representation to effectively address low-contrast disease features under complex field backgrounds. To further balance model complexity and performance, LE-Head adopts re-parameterization technology to simplify DEConv’s parallel convolution into a single standard convolution in the inference stage. This maintains identical parameter count and computational cost while reducing redundancy and optimizing efficiency. In specific implementation, LE-Head first uses a 1 × 1 convolutional layer to perform channel dimension reduction on the input feature map, reducing data transmission and computation overhead based on channel redundancy analysis. Its central element, the Shared Detail-Enhanced Convolution Layer (DEConv_GN), addresses issues of multi-scale feature stability and parameter redundancy. By employing a cross-layer detail feature sharing mechanism integrated with Group Normalization, it optimizes the consistency of feature distributions. In terms of functional module design, the regression layer and classification layer of LE-Head have clear divisions of labor and efficient collaboration. Employing the reg_max strategy, the regression layer achieves precise bounding box modeling through maximizing the probability distribution of the predicted values. The classification layer, based on the representation learning ability of deep neural networks, outputs reliable target category confidences. To address variable target scales in rice disease images, LE-Head introduces an adaptive scale adjustment mechanism. This dynamically adjusts feature map resolution and receptive field size to enhance the model’s detection robustness to targets of different scales. At the training strategy level, LE-Head draws on the advanced experience of RTMDet and uses a dynamic soft label assignment scheme. By introducing a cosine similarity-based label matching algorithm, it optimizes the positive and negative sample assignment strategy during training, effectively alleviating the training instability problem caused by traditional hard label assignment. The structure diagram of LE-Head is shown in Fig. 5.

Fig. 5.

LE-Head structure. This dedicated detection head for RiceDetect-Net integrates Detail-Enhanced Convolution technology, reparameterization technology, an adaptive scale adjustment mechanism and a dynamic soft label assignment scheme to enhance fine-grained disease feature capture capability and multi-scale detection robustness

Adaptive fine-grained channel attention

The successful extraction and fusion of high-level semantic information are crucial in object detection, as they serve as the key factor in distinguishing between object categories and directly affect classification performance [34]. Through successive convolution and pooling layers, deep learning models progressively refine low-level features like edges and textures into discriminative semantic representations, allowing for accurate object category recognition. Despite advancements in rice disease detection, current models continue to encounter challenges, especially when it comes to capturing fine-grained details. Rice diseases typically manifest as fine-grained lesions and texture variations, which are often difficult to accurately identify in complex agricultural environments, especially under low contrast and high background noise. As a result, traditional rice disease detection models have limitations in recognizing detail features such as edges and textures, leading to unstable detection results, low classification accuracy, and even misclassification in some scenarios.

In the evolution from YOLOv5 to YOLOv11, multiple innovative modules have been introduced for object detection tasks. The Spatial Pyramid Pooling - Fast (SPPF) module, as an efficient variant of spatial pyramid pooling, successfully addresses the network’s adaptability to images of different scales through a multi-scale pooling strategy, providing a stable feature foundation for multi-scale object detection [35]. However, as convolution and pooling operations are applied repeatedly, some original semantic information is inevitably lost during feature extraction. This issue becomes especially problematic when processing fine-grained rice disease characteristics, where it may impair the model’s target recognition performance in complex settings. To address these challenges, this paper incorporates an innovative FCA mechanism following the SPPF module, enhancing high-level semantic information fusion and boosting detection accuracy for diverse targets. Drawing on the innovative design of Sun et al. [36] in UBRFC‑Network, the FCA mechanism adaptively optimizes channel weights through global‑local information fusion, thereby strengthening the model’s sensitivity to fine‑grained features. Specifically, the FCA mechanism extracts global semantic information per channel via global average pooling, then employs 1D convolution to explore inter-channel local dependencies. It then deeply fuses global and local information through cross-correlation operations to construct an association matrix for capturing information interaction patterns at different granularities. Finally, through an adaptive fusion strategy, learnable parameters dynamically adjust the weights for combining global and local information. These refined weights are then applied to the original feature maps, enabling fine-grained feature enhancement. By employing a multi-scale and multi-granularity fusion method, the FCA mechanism successfully counteracts semantic loss inherent in conventional feature extraction. This offers reliable technical support for rice disease detection models to achieve accurate identification in complex agricultural environments. In contrast to conventional approaches, the FCA mechanism excels at capturing finer details like disease boundaries and textures, thereby significantly enhancing the model’s accuracy and robustness in classification. Its structural diagram is shown in Fig. 6, and the specific mathematical derivation process is detailed in Eqs. 1–6.

Fig. 6.

FCA structure

1

2

3

4

5

6

The FCA mechanism begins by compressing the input feature map (C、H、W respectively represent the number of channels height and width) into a channel descriptor through global average pooling , detailed in Eq. [1]. It then derives localized inter-channel interactions using a band matrix with weight coefficients , where denotes the number of adjacent channels for local inter-channel interaction, to generate the local channel information vector in Eq. [2], while capturing holistic dependencies across all channels via a diagonal matrix with weight coefficients , where is the total number of channels of the feature map, to obtain the global channel information vector as shown in Eq. [3]. The correlation between these global and local representations is computed via cross-correlation, producing the global-local channel information correlation matrix in Eq. [4]. The global channel weight vector and local channel weight vector are then adaptively fused using a learnable gating parameter , which regulates the fusion ratio of global and local weights and is regulated by the sigmoid function , to generate the fused channel weight matrix which serves for feature weighted enhancement, as formulated in Eq. [5]. Equation [6] generates the weighted output feature map by applying these computed weight matrix to the original input feature map via the element-wise multiplication operator , thereby accomplishing precise feature extraction and enhancement.

Results and analysis

Experimental platform and parameter settings

The complete experimental setup is specified in Table 1. It details the distinct hardware and software configurations used for model training on a Linux server and for subsequent validation on a local Windows system, along with the key hyperparameters that were kept consistent across both phases.

Table 1.

Experimental configuration

Parameter	server	local
Operating System	Linux	Windows 11
Deep Learning Framework	PyTorch	PyTorch
CPU	Intel(R) Xeon(R) CPU E5-2650 v3 @ 2.30 GHz	Intel Core i9 14900HX
System Memory (RAM)	64 GB	16 GB
GPU	NVIDIA V100 (32 GB Memory)	NVIDIA GeForce RTX 4060 (8 GB Memory)
CUDA Version	12.1	12.1
CUDNN Version	8.2	8.2
Input Image Size	640 × 640 pixels	640 × 640 pixels
Batch Size	32	32
Training Epochs	200	200
Optimizer	Stochastic Gradient Descent (SGD)	Stochastic Gradient Descent (SGD)
Learning Rate	0.01	0.01
Momentum	0.937	0.937
Weight Decay	0.005	0.005
Dataset Split Ratio	7:2:1	7:2:1

Data set segmentation

In this study, a dataset comprising 54,240 rice images spanning 10 categories was utilized, including 9 classes of distinct rice diseases and 1 class representing healthy rice plants. The extensive dataset offers robust support for both model training and performance evaluation. To guarantee effective training, stability, and generalization, the dataset was divided into three subsets: 70% for training, 20% for validation, and 10% for testing. A randomized sampling strategy was adopted to minimize manual bias and ensure fairness and randomness in the experimental results. During partitioning, particular attention was paid to maintaining class balance by carefully selecting and distributing samples from each category across the subsets. This approach ensured a representative and uniform class distribution, thereby enabling fair model evaluation and efficient training. The rice disease experimental dataset is summarized in Table 2.

Table 2.

Data on different categories of rice diseases

Classification	Training set	Val set	Test set	Total
Rice blast	4291	1226	613	6130
Bacterial leaf blight of rice	3791	1083	543	5417
Rice leaf spot	3270	934	468	4672
Dead heart disease	3612	1032	516	5160
Downy mildew	1870	534	268	2672
False smut of rice	6556	1873	938	9367
Sheath blight	3504	1001	501	5006
Stripe disease	4354	1244	622	6220
Tungro disease	5682	1623	813	8118
ordinary	1034	296	148	1478
Total	37,972	10,846	5430	54,240

Analysis of model convergence

The rice disease detection model developed in this work demonstrates effective convergence throughout the training phase. During training and validation, essential loss functions including bounding box loss, classification loss, and Distribution Focal Loss all exhibit a steady pattern of rapid decrease followed by stabilization. This reflects marked improvements in the localization and classification of rice disease targets, alongside a consistently stable and dependable training cycle. Meanwhile, precision and recall show a marked rise in the initial phase, approach saturation later, and eventually stabilize at elevated levels. These trends reflect a steady enhancement in the model’s capacity to identify disease-related targets. The detection metrics mAP50(B) and mAP50-95(B) consistently yield favorable outcomes across various thresholds, stabilizing near 0.9 and 0.65 respectively. Such results confirm the model’s effective generalization under different IoU settings. Furthermore, the close alignment between training and validation curves, absent notable overfitting, verifies the model’s reliable convergence and practical value for rice disease detection applications. The model convergence curve is shown in Fig. 7.

Fig. 7.

Model training curve

Ablation studies

Ablation studies are conducted in this paper to verify the efficacy of the novel lightweight detection head LE-Head and the FCA mechanism for rice disease detection, and to demonstrate the overall performance advantage of the proposed model. These experiments encompass both ablation studies on individual modules and comparisons among various attention mechanisms. By gradually removing or replacing each module in the model, the experiments analyze the impact of LE-Head and the FCA mechanism on model performance. For example, after removing LE-Head, the model’s detection accuracy and inference speed significantly decreased, demonstrating that LE-Head plays a key role in accuracy optimization; while after removing the FCA mechanism, although the calculation speed increased, the accuracy significantly declined, indicating that the FCA mechanism is crucial for the model’s detail capture ability. Findings in Tables 3 and 4 substantiate the significant impact of every module and mechanism on the model’s performance.

Table 3.

Ablation experiments with different modules

YOLOv11	FCA	LE-Head	Parameter	Precision/%	Recall/%	mAP@0.5/%	GFLOPs/G
✔			2.58 M	91.2	89.6	93.9	6.3
✔	✔		2.64 M	91.2	90.5	94	6.3
✔	✔	✔	2.32 M	91	91.1	94.3	4.9

Table 4.

Ablation experiments with different attention

YOLOv11	SimAM	TA	EMA	FCA	mAP@0.5/%	Recall/%
✔					93.9	89.6
✔	✔				93.6	90
✔		✔			94	90.2
✔			✔		93.4	89.6
✔				✔	94	90.5

Analysis of Table 3 and Fig. 8 indicates that incorporating the advanced attention mechanism for global-local information fusion alongside the LE-Head detection module reduces the model’s computational load to 77% and its parameter count by 10%, compared to the baseline. Contrary to expectations, the model’s mAP increased by 0.4% points despite its lightweight design, demonstrating performance improvement alongside reduced complexity.

Fig. 8.

Comparison of ablation experiments with different modules

To assess the advantages of the FCA attention mechanism within the proposed rice disease detection method, comparative experiments were carried out with several widely recognized attention mechanisms in recent years, namely SimAM [37], TA [38], and EMA [39], as presented in Fig. 9. The experimental findings presented in Table 4 demonstrate that the FCA attention mechanism reached an accuracy of 94% in detecting rice diseases, surpassing all other methods tested. These findings clearly affirm the superior efficacy of the FCA attention mechanism for rice disease detection, confirming its current status as the preferred approach.

Fig. 9.

Comparison of ablation experiments with different attention modules

Comparative experiments

A series of methodical tests were carried out to measure the effectiveness and benefits of the suggested rice disease detection model, employing a tailored dataset of rice diseases. In particular, the research contrasted various cutting-edge detection models, including the traditional convolution-based YOLOv5, YOLOv8, YOLOv10 and RT18-DETR frameworks. The experimental outcomes presented in Table 5 plainly illustrate the exceptional performance of the proposed model in detecting rice diseases, thereby confirming its notable strengths and suitability for this field.

Table 5.

Comparison of different models

Model	Parameter	Precision/%	Recall/%	mAP@0.5/%
YOLOv5	2.1 M	88.7	87.2	92.3
YOLOv8	2.6 M	89.9	89.4	93.7
YOLOv10	2.7 M	87	85.4	91.3
YOLOv11	2.58 M	91.2	89.6	93.9
RT18-DETR	20.1 M	91.3	92	95.1
RiceDetect-Net	2.32 M	91	91.1	94.3

As evidenced by the analysis in Table 5 and Fig. 10, the proposed RiceDetect-Net achieves optimal performance in both parameter count and average precision for rice disease detection tasks. RiceDetect-Net achieves a 0.4% point higher average precision than the state-of-the-art YOLOv11 model. It reaches 94.3% mAP@0.5 which is highly competitive with RT18-DETR’s 95.1% mAP@0.5. It only uses 2.32 M parameters representing a significant reduction compared to RT18-DETR’s 20.1 M parameters. Moreover, the proposed framework markedly decreases both model parameters and computational demands, confirming its ability to deliver accurate results while efficiently identifying rice diseases with lower resource consumption. This dual optimization of precision and efficiency establishes RiceDetect-Net as a practical solution for agricultural computing scenarios requiring resource-aware deployment.

Fig. 10.

Performance analysis of different models

Based on data from Table 6 and Fig. 11, RiceDetect-Net achieves higher average accuracy across various rice diseases, demonstrating enhanced performance in the recognition task. Specifically, accuracy improved by 0.9% for rice leaf blight, enhancing the capture of disease features; by 0.5% for downy mildew, showing more stable recognition; and by 1.5% for sheath blight, further optimizing detection performance for this disease. Overall, the RiceDetect-Net model significantly enhances the accuracy of rice disease identification and offers stronger technical support for intelligent agricultural diagnostics.

Table 6.

Average precision analysis for various rice diseases

Classification	YOLOv11 mAP@0.5/%		RiceDetect-Net mAP@0.5/%
Rice blast	90.8	90.6
Bacterial leaf blight of rice	95.8	96.7
Rice leaf spot	99	99.1
Dead heart disease	99.5	99.5
Downy mildew	98.7	99.2
False smut of rice	98.2	98.1
Sheath blight	68.1	69.6
Stripe disease	91.6	92.3
Tungro disease	97.7	98.9
ordinary	99.5	99.5

Fig. 11.

Comparison between YOLOv11 and RiceDetect-Net for different diseases, where (a) is YOLOv11 and (b) is RiceDetect-Net

Analysis of the detection results

In this study, a novel RiceDetect-Net model is proposed based on the YOLOv11 framework for intelligent detection of rice diseases. Thorough ablation experiments and broad comparative analysis confirm that the model achieves 94.3% detection accuracy with only 2.32 million parameters. This design successfully balances precision with a lightweight architecture, substantially improving computational efficiency and real-world applicability. Furthermore, specialized evaluations were conducted across nine categories of rice diseases. The outcomes show steady performance gains for each disease category, verifying the model’s enhanced capability and general applicability in multi‑class disease identification. Figure 9 presents a comparative visualization of the detection results between RiceDetect-Net and YOLOv11 across the nine rice disease categories and healthy samples.

Figure 12 reveals that the RiceDetect-Net model introduced in this study demonstrates clear superiority over the conventional YOLOv11 model for rice disease detection, particularly in detection accuracy and region completeness. Specifically, in the detection of rice blast, RiceDetect-Net successfully identifies 9 disease regions, while YOLOv11 can only detect 6 regions, indicating that RiceDetect-Net performs better in terms of the comprehensiveness of disease region coverage and the sensitivity of detection. Meanwhile, in other disease detection tasks, RiceDetect-Net further improves detection accuracy and area boundary integrity compared to YOLOv11. This improved performance is attributed to the improved design of RiceDetect-Net, including the optimized feature extraction mechanism and the efficient fusion strategy of global and local information, which enables it to capture disease features more accurately. RiceDetect-Net showcases superior detection performance across diverse rice diseases, delivering an effective and practical solution for their precise identification.

Fig. 12.

Comparison of rice disease detection results (a) rice blast (b) Rice leaf spot (c) Dead heart disease (d) False smut of rice (e) Bacterial leaf blight of rice (f) Downy mildew (g) Tungro disease (h) Sheath blight Stripe disease (i) Stripe disease (j) Normal rice

Discussion

RiceDetect-Net lowers parameter count while preserving high detection accuracy via its lightweight LE-Head and FCA module. It validates the efficacy of its lightweight architecture and feature fusion strategy for rice disease detection and offers a feasible solution for resource-limited agricultural equipment. The design of the LE-Head and FCA modules is biologically motivated by the specific morphological characteristics of rice diseases. Rice pathogens often present unique visual features that demand high spatial fidelity: for example, Magnaporthe oryzae (Rice Blast) typically manifests as small, dark, spindle-shaped lesions, while Xanthomonas oryzae (Bacterial Leaf Blight) appears as elongated water-soaked streaks. Standard detection heads often rely on aggressive downsampling to reduce parameters, which risks smoothing out these fine-grained features and leads to missed detections. The Detail-Enhanced Convolution (DEConv) in our LE-Head addresses this by preserving high-frequency spatial gradients, ensuring that even minute discolorations or micro-lesions are retained in the feature map. Meanwhile, the FCA module enhances the model’s focus on these lesion features, suppressing interference from complex field backgrounds. This advantage creates the core difference between this study and published works by addressing key limitations of existing rice disease detection models. Unlike most YOLO agricultural variants that struggle to balance lightweight performance and fine-grained detection, DETR models are overly large and incompatible with field equipment. The fused LE-Head and FCA maintains model compactness while being optimized specifically for the unique morphological characteristics of rice lesions, targeting their scattered distribution and blurred edges. It also enhances anti-interference in complex field backgrounds, filling the gap of lightweight models tailored for rice disease diagnosis. However, the adaptability of the model in complex field environments still needs to be improved, and the subtle differences of similar disease spots and leaf shading may lead to recognition errors, and the detection accuracy of a few types of diseases is also affected by data imbalance [40].

This issue is typical in Sheath blight detection: morphologically, its lesions appear as irregular, water-soaked patches spreading along leaf sheaths, often merging into the complex texture of leaf collars and overlapping with healthy tissues. The lack of distinct visual boundaries makes it difficult for the model to distinguish lesions from background clutter, unlike diseases with well-defined spots such as Rice Blast. Meanwhile, the dataset has an imbalance in Sheath blight samples with insufficient proportion, and the limited data hinders the model from learning robust and generalizable features for this disease.

In terms of dataset construction, although the dataset includes nine common rice diseases, the data acquisition conditions remain relatively limited, lacking sufficient diversity in lighting conditions, weather variations, rice varieties, and growth stages. This limitation reduces the model’s capacity to generalize in practical deployment. Additionally, the manual annotation process is subject to subjectivity and inconsistent labeling standards, which may introduce labeling errors and subsequently affect the quality of model training and overall performance.

In practical application scenarios, deploying models on mobile devices faces multiple challenges related to energy consumption, stability, and real-time performance. Factors such as high temperature, humidity, and dust in agricultural environments can interfere with device operation. Long-term continuous detection imposes higher requirements on the battery life of hardware, while performance differences among different edge computing devices may also lead to unstable model inference efficiency.

Future research can be optimized from multiple dimensions. The model architecture integrates an adaptive attention mechanism [41] to improve the detection of subtle features, and combines this with knowledge distillation [42] to enhance overall detection performance; in terms of the dataset, the scope of the sample collection is enlarged by using the technology of unmanned aerial photography and multi-spectral imaging, and the category imbalance is alleviated by the active learning algorithm [43]; in terms of the application, the deep integration of the model with the IOT device is promoted to construct an integrated system of “detection-warning-decision”. In terms of application expansion, we promote the deep integration of the model with IoT devices to build an integrated system of “detection-warning-decision-making”, and at the same time, we develop multimodal prediction models by combining the environmental data to realize the early warning of rice diseases. In addition, through the model compression and quantization technology to reduce the consumption of hardware resources, to promote the low cost and portability of detection equipment, to help the widespread application of agricultural intelligent technology.

Conclusions

This research tackles the constraints of limited computational power and storage capacity in devices used for rice disease detection. RiceDetect-Net is presented as a computationally efficient yet highly accurate single-stage detection model, specifically engineered for reliable rice disease identification within agricultural settings where hardware resources are often limited. By incorporating the dedicated LE-Head module, the architecture achieves a significant reduction in both parameter count and computational demands, thereby enhancing its suitability for real-time operation on portable field devices. RiceDetect-Net incorporates the FCA attention module to strengthen its capacity for extracting advanced semantic features, leading to more precise classification. When tested on a collection of 54,240 rice leaf images covering nine distinct diseases, the model attained a detection accuracy of 94.3%. RiceDetect-Net surpasses the original YOLOv11 by increasing mAP@0.5 by 0.4% and decreasing parameters by 10%. These outcomes highlight its enhanced precision, speed, and robustness, all while maintaining a lightweight profile ideal for real‑world agricultural deployment. Future research efforts will focus on the following directions: first, enhancing the adaptability and versatility of the model in diverse field environments; second, expanding the dataset scale by incorporating samples of more rice varieties and both abiotic and biotic stress scenarios; third, exploring the scheme for deep integration of the model with IoT-driven agricultural monitoring systems; fourth, optimizing the computational performance of the model on low-power agricultural edge devices to improve real-time detection efficiency; fifth, conducting long-term field validation trials to comprehensively verify the stability and reliability of the model in practical large-scale applications.

Authors’ contributions

Conceptualization, Y.W.; methodology, X.Y. and T.W.; validation, Q.X. and G.Z. and L.G.; formal analysis G.Z.; investigation, G.Z. and Y.W.; resources, L.J. and Q.X.; writing-original draft preparation, X.Y.; writing-review and editing, X.Y. and T.W.; visualization, Y.W. and L.J; supervision, Q.X. and T.W.; project administration, T.W. and L.G.; funding acquisition, Y.W. and L.J. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Zhejiang Provincial Natural Science Foundation Public Project (LTGN23C130002, LGN22C190029); the Jiaxing Agricultural Science Research Institute and Jiashan Agricultural Science Research Institute (H20220562); the Zhejiang Province “Three Rural Areas and Nine Parties” Science and Technology Cooperation Project (2024SNJF007).

Data availability

The data presented in this study are available upon request from the corresponding author.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.