The tomato leaf disease dataset used in this study was sourced from the “Tomato Leaf Diseases Detect” standardized dataset released by the Roboflow open platform. It contains six typical disease categories (bacterial spot, early blight, late blight, leaf mold, target spot, and black spot) and healthy leaf samples, covering the early, middle, and late stages of disease development. In total, it includes 3,469 high-resolution RGB images. The six tomato leaf diseases of interest in this study are highly prevalent in major tomato-growing regions worldwide, causing yield losses of 20% to 65% (Lu et al., 2018; Panno et al., 2021). Table 1 provides a detailed breakdown of the final image distribution across all categories after augmentation and splitting. The common Bean Dataset was captured at the Guoxin Modern Agricultural Base in Changchun City, Jilin Province, and the public dataset Bean Disease Dataset. The Potato dataset is from the public Potato disease dataset which includes three categories: health, early disease and late disease (https://gitcode.com/open-source-toolkit/829ec). The inclusion of these diverse datasets from different crops is intended to rigorously validate the transferability of the features learned by our model from tomatoes to other commercially important crops.

To enhance the robustness of the model, the dataset was expanded to 7370 images using data augmentation techniques. In order to prevent the original image and enhanced image from appearing simultaneously in the training set and validation set, the original image is initially divided into a training set, validation set, and test set in a ratio of approximately 8:1:1. Subsequently, five techniques including horizontal flipping, vertical flipping, grayscale, contrast adjustment, and brightness adjustment were randomly applied to the original data to enhance the image and labels. The enhanced example image is shown in Figure 1 , and the label distribution is shown in Table 1 .

In order to simulate the complex weather changes in the real tomato cultivation environment, this study adopts the RGB channel synthesis technology based on the atmospheric scattering model to generate enhanced images with controllable weather features on 50% of the typical samples in the training set. The synthesis formula is shown in Equation 1.

In this study, x denotes the pixel coordinate, I ( x ) signifies the synthesized image, J ( x ) represents the original image, the transmittance map t ( x ) is constrained within the interval [0.2, 0.8] and controls the weather intensity gradient, and the atmospheric light value A restricts the amplitude of illumination attenuation. Figure 2 presents an image of medium-intensity synthetic weather.

While general-purpose detectors like YOLOv11 have demonstrated strong performance on common datasets, their inherent architecture is not optimally designed for the unique challenges presented by complex agricultural environments, such as severe scale variation of lesions, leaf occlusions, and pervasive background noise. These limitations often lead to feature misalignment, reduced sensitivity to small objects, and compromised robustness under fluctuating lighting conditions.

To address these specific issues, we propose the Tomato Multi-scenario Adaptive Scale Detector (ToMASD), a novel lightweight architecture specifically engineered for high-precision disease detection in real-world field settings. The overarching design philosophy of ToMASD is to achieve an optimal balance between computational efficiency and detection accuracy by introducing three dedicated core modules that work in concert throughout the feature extraction and fusion pipeline.

As illustrated in Figure 3 , the Two-branch Adaptive Alignment Module (TAAM) is integrated into the backbone network. Its purpose is to dynamically align and calibrate multi-scale features at the earliest stage, effectively mitigating the semantic misalignment between healthy and diseased tissue regions caused by occlusion and scale variance. The Faster-Gated Linear Unit (Faster-GLUDet) is embedded within the neck network. This module acts as an adaptive feature refiner, leveraging a gating mechanism to suppress irrelevant background noise. The Multi-scale Decoupling Head (MDH) is designed as the detection head. It replaces the conventional coupled head with a decoupled structure, allowing for independent optimization for classification and regression tasks at different feature scales. This synergistic design ensures that ToMASD is uniquely capable of handling the complexities of agricultural disease detection.

The present study focuses on the problems of insufficient feature extraction of small-scale targets and redundant shallow computations in the YOLOv11 backbone for leaf lesion detection. It proposes a novel dual-branch adaptive alignment module, the Two-branch Adaptive Alignment Module (TAAM), as shown in Figure 4 . The module under discussion achieves efficient computation through a multi-level feature sharing architecture. Firstly, the Pointwise Spatial Attention Stem (PSAStem) is utilised as the shared initial extraction layer, which pre-calibrates the input features through 1×1 pointwise convolution and an adaptive mechanism, thereby enabling the network to form dynamic spatial focusing capabilities at the input stage. Subsequently, the feature maps are processed through dual paths. The primary pathway integrates two C3k2 modules,each containing three standard 3×3 convolutions with 64 output channels, and a standard 3×3 convolution to preserve intricate features. The secondary pathway employs a 1×1 dimension-reducing convolution (reducing channels by a factor of 2) and subsequently connects to the optimised PSABlock. The initial feature extraction module combines pointwise convolution and spatial attention mechanisms, enabling the network to prioritise key regions in the input image and enhance the dynamic focusing ability on key spatial regions in the input stage while maintaining computational efficiency.

After the dual-branch channels process the features in parallel, they are connected to the Adaptive Alignment Module (AAM), as shown in Figure 5 . The input dual-channel features F₁ and F₂ are first compressed in the channel dimension through 1×1 convolution layers to obtain F₁’ and F₂’, and then adaptive alignment weights - Adaptive Align Weight are generated through the cross-branch feature interaction layer to dynamically balance the contribution of the two paths. They are input into a 3×3 dilated convolution to capture long-range context dependencies and generate dynamic path selection weights α through the Sigmoid activation function. After spatial alignment of the two paths, the key region responses are enhanced through element-wise multiplication, as shown in Equations 2 and 3, where σ represents the Sigmoid activation function and W_α is a learnable weight matrix. Finally, multi-scale feature complementarity is achieved through element-wise addition, and the results are merged and output.

The innovation of this module can be attributed to the following: The following two mechanisms are introduced: 1) A dynamic weight adjustment mechanism is employed to optimise branch weights in real time based on the semantic distribution of the input features. This mechanism alleviates the spatial offset problem of heterogeneous features, such as misalignment interference between leaf lesions and healthy tissues. 2) Through the collaborative design of dilated convolution and gated attention, computational redundancy is reduced while local details and global pathological patterns are jointly modelled.

The neck network of YOLOv11 employs depthwise separable convolution and channel pruning strategies, which have been shown to enhance the recall rate of small targets while concurrently reducing the model’s parameters. Tomato leaf diseases frequently manifest as minute spots, and the receptive field of the P5 layer in the feature pyramid is overly extensive, which may impede the learning of small target features. Secondly, when the brown necrotic spots of tomato late blight are similar in colour to the soil, the feature pyramid network may confuse the target with the background. In order to address these issues, we propose Faster-GLUDet, whose core lies in enhancing the model’s ability to extract disease features in complex backgrounds through a gating mechanism while maintaining model lightweight. The Faster-GLUDet module integrates FasterNetBlock and Convolutional GLU (Chen et al., 2023; Shi, 2023), as illustrated in Figure 6 .

The module’s primary function is the extraction of feature units through the utilisation of FasterNet Block, employing a 3x3 Partial Convolution to extract spatial features from a quarter of the input channels. This approach results in a 25% reduction in computational load when compared with traditional convolution, while retaining edge detail information. Subsequently, two 1x1 convolutions are connected to perform channel dimension reduction and feature fusion, thereby ensuring the effective preservation of multi-scale disease features. In order to enhance the feature interaction between modules, a dynamic gated feature enhancement unit, known as the Convolutional Gated Linear Unit (ConvGLU), is connected after the Faster Block. The model employs a dual-branch convolution to generate feature maps and gating signals, dynamically suppressing background noise and enhancing the response in the lesion area through element-wise multiplication. In the gating branch of the traditional gated linear unit (GLU), ConvGLU decomposes the standard 3x3 convolution into a cascade structure of depthwise convolution (DWConv) and pointwise convolution (PWConv), and combines a linear projection layer and GELU activation function to construct a lightweight feature enhancement path. A lightweight 3x3 depthwise convolution operation is introduced prior to the activation function in order to construct a gating channel attention mechanism based on neighbourhood features. The design converts global channel attention into local context-aware dynamic weight adjustment through the local receptive field characteristics of the convolution kernel. This retains the important information filtering ability in the channel dimension while significantly reducing computational complexity. ConvGLU employs convolution operations to capture local features in the image, thereby enhancing its efficacy in processing local information in comparison to traditional FFN. It is also capable of adaptively enhancing small target features.

The integration of Faster-GLUDet has been demonstrated to enhance the model’s feature extraction and expression capabilities to a considerable extent. The lightweight design of FasterBlock provides low-latency input for CGLU, while the gated weight generation module of CGLU further optimises multi-scale feature interaction, enabling the model to maintain lightweight while enhancing the diversity and hierarchy of features. This, in turn, helps better capture the details and context information of the target object. The system has been engineered to achieve dynamic regulation of feature channels, thereby further enhancing the semantic segmentation and spatial understanding capabilities of features. The combination of these two approaches has been shown to more effectively fuse multi-scale features and alleviate the problem of information loss, significantly improving the model’s performance in challenging plant disease recognition tasks.

In order to address the issue of uneven detection accuracy of traditional detection heads for small and large-scale lesions, the Multi-scale Decoupling Head (MDH) has been proposed, as illustrated in Figure 7 . The core process is as follows: MDH receives three different-scale feature maps - P3, P4 and P5 - from the Feature Pyramid Network (FPN) in parallel, which respectively carry high-resolution details, medium-scale information and large receptive field context, thereby constructing a multi-scale perception foundation. The features of each scale first enter a unified feature enhancement pathway, which is composed of a series of grouped normalized convolutional modules: First, the channel dimension is adjusted and fused through a 1×1 Conv_GN, and then two 3×3 Conv_GN modules are continuously used to enhance the spatial feature expression. At the same time, group normalization is utilized to ensure the stability of the model under small-batch training. After feature enhancement, the network flow is completely decoupled into two independent branches dedicated to their respective functions: The classification branch precisely extracts features through two consecutive 1×1 convolutional layers and ultimately outputs a probability graph with a dimension of nc, accurately determining the category of the target within each anchor box; The regression branch adopts the same structure, but its output dimension is 4 × reg_max, which indicates that it uses an advanced distributed focus loss mechanism. By predicting the discrete distribution of bounding box coordinates, it greatly improves the accuracy of lesion location, where reg_max defines the flexible maximum value of the distribution. Ultimately, the outputs of the two branches are respectively normalized and integrated through a scale layer to generate the final detection results.

In this paper, the metrics Precision, Recall, and mAP are utilised to evaluate the detection performance of the model. TP, FP, and FN represent the number of true positive, false positive, and false negative samples, respectively. C denotes the set of object categories, and |C| is the total number of categories. As shown in Equations 4, 5 and 6

P denotes the proportion of correctly detected disease samples among all positive detections, thereby reflecting the model’s capacity to avoid false positives. R signifies the proportion of correctly detected disease samples among the actual existing disease samples, thus measuring the model’s ability to reduce false negatives. AP quantifies the detection performance of the model for a single disease category by calculating the area under the precision-recall curve. mAP is the average of APs for all categories, and a higher mAP indicates that the model’s detection effect on various disease categories is more balanced and accurate.

To comprehensively evaluate the detection and generalization performance of the proposed ToMASD model, we conducted extensive comparative experiments with thirteen state-of-the-art object detection models on the same tomato leaf disease dataset. As summarized in Table 3 , ToMASD achieved the highest scores in both precision and mAP, significantly outperforming all other contenders. RT-DETR achieves high accuracy but at the cost of high computational complexity, making it unsuitable for edge deployment. In comparison with the unimproved YOLOv11n, ToMASD has increased P, mAP, and Recall by 6.6%, 7.8%, and 5.9%, respectively, demonstrating its superior ability in target localization and classification in complex scenarios. Despite the fact that YOLOv11n exhibits a modestly diminished number of FLOPs in comparison with ToMASD, a notable deterioration in accuracy is evident, suggesting the potential for optimisation shortcomings within the feature ex-traction process. Despite the advantages in parameter quantity and computational cost of YOLOv5n and YOLOv11n, their accuracy still lags significantly behind ToMASD, further confirming the dual improvements in accuracy and recognition performance of the ToMASD model. Whilst sustaining its position of being lightweight, it has considerably surpassed the constraints of prevailing algorithms in the trade-off between ac-curacy and computational resource consumption. As illustrated in Figure 8 , a performance comparison of the six models with the highest mAP is presented. A selection of six algorithms with comparable performance was made for the purpose of a comprehensive comparison, as illustrated in Figure 9 . The performance of the metric is optimised by the distance of each axis of the curve from the intersection point. The area enclosed by the curve is positively correlated with the strength of the algorithm’s comprehensive performance. The comparison results demonstrate that the ToMASD model proposed in this paper exhibits advantages in all metrics, not only improving performance but also achieving lightweight, thus rendering it more suitable for practical scenarios.

To verify the performance of the TAAM module, this study systematically evaluated the effects of four mainstream attention modules as the Stem, as shown in Table 4 . PSA achieved the best balance between accuracy and computational efficiency, reducing the computational complexity by 17.7% compared to the ECA module with the sec-ond-highest accuracy, verifying the performance of PSA.

Figure 10 shows the detection performance of the five models with the highest accuracy under different weather conditions, where gray boxes indicate missed detec-tions and black boxes indicate false positives. Under foggy conditions, except for To-MASD, the other four models misjudged fog points as diseases. When the light intensity was too high, the comparison models also failed to effectively suppress the exposed areas, resulting in missed detections of some real lesions. Figure 11 shows the detection results of the four lightweight models in conventional environments. Comprehensive comparative analysis shows that ToMASD exhibits the most superior performance under various complex weather conditions and has superior feature extraction and denoising capabilities in real agricultural environments. Figure 12 shows ToMASD’s detection of potato and common bean leaf diseases, and the experimental results show that ToMASD maintains high accuracy in the cross-crop task, achieving 92.1% and 93.5% accuracy in the detection of bean and potato leaf diseases, respectively, demonstrating its efficient generalization ability and transferability nature of cross-species training. Table 5 shows the recognition accuracy of ToMASD for different spots, Figure 13 shows the confusion matrix of this experiment, the model still maintains stable recognition performance in the category imbalance dataset, the recognition accuracy of Late Blight, Early Blight is close to 90%, Leaf Mold has a similar chromaticity of the yellow spot and the healthy tissues, which leads to the relatively low detection accuracy. Bacterial Spot and Target Spot have similar water-damaged spot characteristics, but the model still achieved 82.1% and 74.1% mAP values through multi-scale texture analysis, indicating the effectiveness of the feature decoupling mechanism.

The proposed ToMASD model is based on YOLOv11n and has been optimised by introducing TAAM, Faster-GLUDet, and MDH. In order to evaluate the performance of each optimisation module, an experiment was conducted using the variable control method. The training and testing were carried out on the same dataset and training parameters, and the results are shown in Table 6 . It is evident that Model C attained an mAP% of 77.9%, while concurrently sustaining a minimal computational cost. However, Model D, which introduced TAAM and Faster-GLUDet, exhibited an increase in computational cost to 12 FLOPs/G, attributable to parameter redundancy, yielding an accuracy of only 79.2%. The experiments indicated that the joint application of TAAM and MDH caused feature decoupling conflicts. The Model E achieved an 82.1% P% and 80.1% mAP%, thereby demonstrating the viability of multi-module collaborative optimisation through its lightweight design. The ToMASD model proposed in this paper was found to achieve the optimal balance in parameters, computational efficiency, and performance. In comparison with the baseline model A, it enhanced the accuracy by 9.4% whilst escalating the computational cost by a mere 0.4G. This outcome serves to demonstrate the efficacy of the collaborative design of multi-dimensional attention mechanisms and lightweight architectures for object detection tasks.

The present study employed the gradient-weighted class activation mapping technique (Selvaraju et al., 2017) to visualise the small target detection mechanism of the ToMASD model ( Figure 14 ). The experimental findings demonstrate that the tomato leaf disease detection model, based on transfer learning, exhibits adapted feature capabilities in different crop disease recognition tasks. When the model is transferred from the source domain of tomatoes to the target domain of common beans, the heatmap analysis indicates that the lesion areas can still be effectively captured, although the the extent of the activated regions of highlighted areas is lower than that in the source domain. This suggests that the model has initially acquired the ability to locate disease spots across species through transfer learning. When the heatmap is transferred to potatoes with more distinct morphological features, more concentrated highlighted areas are shown, which may be related to the reticulate vein structure of potato leaves, thereby enhancing the distinguishability of texture features. It is noteworthy that the heatmaps of all three crops demonstrate a substantial contrast between the lesion areas and healthy tissues, thereby confirming that the model, while retaining key pathological features, has achieved adaptive adjustments to different crop leaf diseases through weight transfer. This cross-species disease recognition capability provides a feasible solution for intelligent diagnosis of multiple crop pests and diseases under resource-limited conditions.