Form related works, and numerous deep learning models (Liu et al., 2024; Dhanka and Maini, 2025; Han et al., 2025; Kamonsukyunyong et al., 2025; Logeswari et al., 2025; Mu et al., 2025; Patel, 2025; Qiao et al., 2025; Zhang C. et al., 2025; Zhang Z. et al., 2025) that have demonstrated promising results in the field of leaf ailment finding have been developed under monoculture conditions and exhibit limitations ( Table 1 ) when applied to diversified cropping systems (Qin et al., 2024; Xu et al., 2024; Liang et al., 2025; Mohanty et al., 2025; Xu et al., 2025). TinyResViT, MobileH-Transformer, and GC-MobileNet have shown high accuracy on benchmark datasets; however, these methods rely heavily on ideal conditions and not generalize to real-world intercropping systems (Upadhyay et al., 2025). A critical observation from the literature review indicates that most existing works do not address leaf disease prediction in intercropping systems. To date, few studies have attempted to explore disease detection (Liu et al., 2024) within such systems, despite their increasing relevance in sustainable agriculture (Xu et al., 2023; Gong et al., 2024; Tu et al., 2024; Wang et al., 2024; Lu et al., 2025). Intercropping systems introduce unique challenges such as overlapping foliage, interspecies spectral interference, and a higher risk of co-infections—factors that are typically overlooked in conventional monoculture-based models (Liu et al., 2024; Qin et al., 2024; Xu et al., 2024; Cai et al., 2025; Chi et al., 2025; Christy and Jeyaraj, 2025; Dhanka and Maini, 2025; Han et al., 2025; Kamonsukyunyong et al., 2025; Liang et al., 2025; Li J. et al., 2025; Lingayya et al., 2025; Li Q. et al., 2025; Logeswari et al., 2025; Maranga et al., 2025; Mohanty et al., 2025; Mu et al., 2025; Patel, 2025; Qiao et al., 2025; Ratmele et al., 2025; Wang and Ruan, 2025; Wu M. et al., 2025; Xu et al., 2025; Zhang C. et al., 2025; Zhang Z. et al., 2025). To fill the research gaps, an intelligent framework for integrated early detection of foliar diseases is proposed, combining hyperspectral imaging with hybrid deep learning applied to precision agriculture. Hyperspectral imaging provides rich spectral information, capable of capturing subtle biochemical and structural changes in crops such as corn-soybean and pea-cucumber ( Figure 1 ).

The dataset used in this study is the publicly available hyperspectral dataset curated by Liu et al (Liu et al., 2024; Han et al., 2025; Zhang Z. et al., 2025). This dataset was collected from maize–soybean and pea–cucumber intercropping fields in Fei Cheng City, Tai’an, Shandong Province, China. It includes hyperspectral images of maize, soybean, pea, and cucumber leaves under various disease conditions. Specifically, the maize–soybean dataset contained healthy maize leaves, maize leaves with leaf spot, rust-infected maize leaves, and samples with combined infections, along with healthy and rust-infected soybean leaves. The pea–cucumber dataset ( Figure 3 ) comprised healthy pea leaves as well as leaves infected with Ascochyta blight, powdery mildew, downy mildew, and Fusarium wilt. Cucumber samples included healthy leaves and those affected by angular leaf spot, powdery mildew, downy mildew, and anthracnose.

The target disease region segmentation plays a vital role by isolating only the infected portions of crop leaves from hyperspectral images while discarding irrelevant background and healthy regions. This step confirms that the subsequent stages of deep feature extraction and classification focus solely on the areas exhibiting disease symptoms, thereby improving accuracy and reducing false detections. To achieve this, a hybrid Synergistic Swarm Optimization (SSO) based on the Crowd Synchronization Algorithm (CSA) is employed (Logeswari et al., 2025). It is inspired by the collective behaviors of birds, fish, and ants, which rely on synchronization and cooperation to achieve optimal solutions. In this context, the input consists of hyperspectral images of maize, soybean, pea, and cucumber leaves, where each pixel carries intensity values across multiple spectral bands in the 400–1000 nm range. Unlike conventional segmentation methods such as K-means or Otsu’s thresholding, which fail under complex intercropping conditions, or deep segmentation models like U-Net that demand extensive annotated datasets and heavy computation, the proposed SSO-CSA hybrid dynamically adapts segmentation boundaries based on spectral similarities, swarm synchronization, and information sharing among candidate solutions (Zhang C. et al., 2025). This prevents premature convergence and ensures robust detection even in the presence of overlapping infections and multi-crop variations. The output of segmentation (Equation 1) process is a binary disease mask that highlights infected regions while suppressing background and healthy tissues, along with segmented hyperspectral sub-images that are passed to the PANet module for deep feature extraction. By focusing only on disease-affected regions, this segmentation approach reduces computational overhead and enhances the downstream classification accuracy of the system.

A matrix P represents fitness solution form the objective function (Equation 2).

where, un and ln stand for the vectors above and below each dimension of the issue space, respectively, and B implicitly represents the dimensions or variables Equation 3 of the given problem. Candidate solutions are updated with (P) (Dhanka and Maini, 2025).

where P N e w ( h ,   g ) denotes the original optimal place g of the h-th applicant explanation, P ( h , g ) denotes g the current position of the h-th contender key, V ( h , g ) and g denotes the place of the h-th contender resolution value. The particles are based on the local and global attraction of states (Equation 4) (Kamonsukyunyong et al., 2025).

The V n e w ( h , g ) value is the disinterest weight value (IWV), describes as trails (Equation 5):

When the inertia weight parameter (z), which dynamically regulates the ratio of exploration to exploitation (Equation 6), has an adaptive mechanism denoted by z. The following formula is used to get the personal best coefficient (PBC) (Patel, 2025):

where, P h solution number h is returned, Rand(xbest) is a unplanned resolution from the available entrant solutions, R1 is a random value, and Eps returns a tiny assessment. The global best coefficient (gbc) Equation 7 is calculated as follows (Qiao et al., 2025):

where, R2 is an accidental value j b e s t s representing the best comprehensive explanation and P h gives the solution number. The diversity preservation constant (dmc) is designed as trails (Equation 8) (Mu et al., 2025):

where, d2 is an extra hastening factor for the variety period and R5 is a random value. D i v e r s i t y h stands for the location in the cluster where the particle’s neighborhood diversity is maximized. The research gaps discussed in Table 1 . Algorithm 1 describes the working process of target region segmentation using SSO.

The feature extraction from the segmented target region of the disease transforms it into a set of meaningful descriptors that describe essential characteristics such as shape, texture, color, and structural patterns of the damaged area. In order for classification algorithms to accurately identify the type and severity of the disease, these extracted features are essential inputs. The fuzzy attention network (PANet) is used to efficiently extract features from the segmented disease areas in leaf images. PANet permits the network to focus more precisely on the most relevant input areas, such as diseased parts, while reducing the effect of healthy or non-relevant areas. Fuzzy attention components (Wang and Ruan, 2025) allow the model to selectively focus on exact feature maps that capture various visual cues such as texture, edges, and gradients at different levels of abstraction.

In the context of identifying plant diseases, feature extraction often leads to a database that contains a variety of in-depth features. However, not all of these features are relevant or necessary for the classification of diseases. Feature optimization aims to reduce the dimensions of the dataset by identifying and retaining only the most important features that contribute efficiently to the classification task. By removing the less useful features, the model becomes more efficient, faster, and less prone to overfitting, thereby improving its performance and generalization capability. To optimize the features in this work, the two-stage Kepler optimization algorithm (DSKO) is used. DSKO is a metaheuristic algorithm inspired by nature and based on Kepler’s laws of planetary motion, which determine the movement of celestial bodies in space (Cai et al., 2025). The populace size is the numeral of planets B x that reflect the optimization problem’s decision parameters, is dispersed randomly over fuzzy sizes in the manner described in Equation 9:

where h-th represents P h , g candidate solution, B x is the numeral of applicant solution in the exploration space R₁; P g ,   L O W and P g , U P denotes the lesser and higher bounds of the g-th optimal parameter, separately. Where ai is the elliptical orbit semi-major axis at time s of object h, which is determined by Kepler’s third law as follows: the minimum value to prevent a divide-by-zero error. An absolute value randomly generated using a regular circulation to signify the orbital period of the objective. P t and P h ; denotes Euclidean distance normalization (Equation 10); defined as follows.

To inform the distance position of every objective from the Sun according to Equation 11 the previous steps:

where P → h ( s + 1 ) denotes an object H’s new position at time s + 1, P h ( s ) denotes the object h’s current location at time s, VHS shows the velocity required for object h to move to the innovative situation, P h ( s ) displays the optimal sun place, which is linked to the greatest explanation with the lowest fitness score, and F is shown as a flag to change the route of the exploration.The normalized values of A b t and a b h respectively (Equation 12),

where F i t K ( a b h ) is the value of the fitness function with respect to each position of the object K at the current time s; Worst(t) denotes the solution candidate with the greatest fitness score (Equation 13). The term ( r b h ), which denotes the normalized value of ( r h ), may be used to determine the Euclidian distance between P h and P t .

To accomplish exploration correctness, µ(s) is a purpose that exponentially debilities with time (s). The fitness optimization is expressed in Equation 14:

The procedure of feature optimization with DSKO is explicated in Algorithm 2 .

Leaf disease prediction is process in the early detection and organization of plant diseases, based on the visual symptoms observed on the leaves of plants. In such an intercropping environment, the prediction of foliar diseases becomes more difficult due to the presence of various plants with different disease susceptibility profiles. Therefore, accurate disease prediction is crucial to prevent significant yield losses and ensure prompt intervention through appropriate disease management strategies. To achieve accurate and efficient predictions of leaf diseases in intercropping systems, a randomly used deep convolutional neural network (R-DCNN) is employed. R-DCNN is an advanced variant of the traditional deep convolutional neural network (DCNN), designed to enhance model performance by introducing random elements into the network structure and learning pathways. In R-DCNN, the convolution layer is used to learn parameters such as the weight matrix (n) and the dependence terms (E) of the convolution kernel (k). The complete convolution calculation ω x is discussed in Equation 15 as follows (Li J. et al., 2025).

The following is the sample size of 1 × 2 if the pooling layer (d) employs uniform sampling ω x y (Equation 16).

For the pooling layer, s₁ and s₂ stands for the random input and output values, respectively. A feature network of R-DCNN is mapping of features from an input image to convolutional kernel and a processing function in convolutional layer (Equation 17).

To determine the neural output by using the nonlinear activation function is d K h : The current layer additive bias k-th feature map is signified by E is the stimulation function d K h (Equation 18) which usually starts at 0 (Chi et al., 2025).

The analysis’s findings demonstrate that the ReLU feature enhances the network’s capacity for recognition and learning. A fully connection layer φ K h is used to finalizes the random of nonlinear mapping and network size optimization (Equation 19).

The results of the ReLU feature improve the network’s recognition and learning capabilities. The objective function j L 2 represents the combined weight and offset factors of layer 1 (Equation 20) input layer 2 units (Equation 21).

Forward propagation relies heavily on finding the appropriate intermediate stimulus value (Equations 22, 23) for each layer.

j ( h ) is the numeral of coatings in the neural net, h for the input layer, and j ( h ) for the output layer. The system’s main issue is the willpower of the excitation value at each buried layer in the neural system forward spread. Algorithm 3 describes how to use DSKO for leaf disease detection and categorization.

The results of the SSO algorithm was linked with present segmentation algorithms—K-means clustering (K-MC), fuzzy C-means (FCM), and particle swarm optimization (PSO)—using dice similarity coefficient and Jaccard index as evaluation metrics. Table 5 presents the comparative results of segmentation algorithms for leaf disease detection in intercropping systems. In the Corn-Soybean field, the SSO algorithm consistently outperformed K-MC, FCM, and PSO across all classes. Dice coefficient improvements ranged from 16.05% to 28.68%, while the Jaccard index increased by 29.74% to 49.93%. Similarly, in the Pea-Cucumber field, SSO achieved notable gains, with Dice increases up to 31.28% and Jaccard improvements reaching 53.75% over existing methods. These results highlight SSO’s superior segmentation accuracy for multi-crop disease detection. In the Pea-Cucumber field, SSO consistently outperformed K-MC, FCM, and PSO across all disease classes. Dice coefficient improvements ranged from 20.76% to 33.79% for pea diseases and 22.79% to 32.00% for cucumber diseases. Corresponding Jaccard index gains varied between 38.52% to 63.34% in pea and 41.30% to 55.44% in cucumber. These results highlight SSO’s superior segmentation performance, achieving 15% to over 60% improvements across metrics, ensuring highly accurate disease detection in intercropping.

The exercise and challenging loss curves of the R-DCNN model for leaf disease prediction in Maize-soybean and pea-cucumber intercropping fields, as presented in Figure 5 , show a consistent decline across increasing epochs, indicating effective model convergence and enhanced learning. In the Maize-soybean field, the exercise loss reduced from 0.402 at epoch 20 to 0.009 at epoch 1000, representing a 97.76% reduction. Similarly, the difficult loss concentrated from 0.417 to 0.032, yielding a 92.33% decrease, which reflects the model’s improved generalization ability. For the pea-cucumber intercropping field, the training loss dropped from 0.392 to 0.014, marking a 96.43% reduction, while the difficult loss decreased from 0.405 to 0.017, which corresponds to a 95.80% reduction. The exercise and difficult accuracy presentation of the R-DCNN model for leaf disease prediction, as illustrated in Figure 6 , shows a substantial improvement across training epochs for both Maize-soybean and pea-cucumber intercropping fields.

The accuracy results of leaf disease detection in Maize-soybean intercropping fields, as shown in Figure 7 , the performance comparison of models, PANet+R-DCNN and PANet+DSKO+R-DCNN, across various learning rates during the 10-fold validation process. The smallest improvement of 0.21% is observed at a learning rate of 0.002, where PANet+DSKO+R-DCNN achieved 0.967, compared to 0.959 for PANet+R-DCNN. PANet+DSKO+R-DCNN show an advantage, with an improvement in accuracy of approximately 0.56% across the ten-fold validation.

The correctness results of leaf illness discovery for pea-cucumber intercropping fields, presented in Figure 8 , show the performance comparison between PANet+R-DCNN and PANet+DSKO+R-DCNN models across various learning rates. The results indicate that PANet+DSKO+R-DCNN consistently outperform PANet+R-DCNN, with improvement in accuracy. The ROC curves illustrated in Figure 9 shows the strong classification performance of the models across different crop disease classes. For both Maize-Soybean and Pea-Cucumber intercropping fields, the curves consistently stay well above the random guess line, indicating high true optimistic charges and low false confident charges.

Table 6 describes the performance evaluation of the proposed model for leaf disease detection in maize–soybean and pea–cucumber intercropping systems demonstrate consistently high accuracy across all disease classes. In the Maize–soybean field, normal maize leaves achieved an accuracy of 99.857%, with precision, sensitivity, specificity, and F-measure values all above 99.8%, indicating almost perfect classification. Leaf spot detection showed slight reductions in accuracy and specificity, while precision, sensitivity, and F-measure remained above 99.1%, highlighting minimal performance degradation. Rust-infected maize leaves achieved 99.364% accuracy, with all other metrics exceeding 99.2%, reflecting robust detection capabilities. Hybrid maize leaves demonstrated 99.396% accuracy, with minor variations in other metrics, showing that the model can accurately handle mixed infections. Soybean leaves, both normal and rust-infected, reached near-perfect scores, with normal leaves attaining 100% across all metrics and rust-infected leaves achieving 99.854% accuracy, with a slight decrease of 0.146% compared to normal leaves. In the Pea–cucumber field, normal pea leaves showed 99.413% accuracy, with corresponding high values for precision, sensitivity, specificity, and F-measure, reflecting consistent performance. Disease classes such as Ascochyta blight, powdery mildew, downy mildew, and Fusarium wilt exhibited accuracies of 99.132%, 99.053%, 99.024%, and 99.293%, respectively, representing minor reductions of 0.281%, 0.36%, 0.389%, and 0.12% compared to normal pea leaves, while maintaining high detection reliability. Similarly, cucumber leaves demonstrated strong performance, with normal leaves at 99.413% accuracy. Angular leaf spot, powdery mildew, downy mildew, and anthracnose achieved accuracies of 99.373%, 99.192%, 99.283%, and 99.182%, respectively, showing marginal decreases ranging from 0.041% to 0.221% compared to normal leaves. The results indicate that the model consistently delivers robust and reliable classification, with improvements ranging from 0.041% to 0.281% compared to the next best-performing disease classes, highlighting its effectiveness in real-world hyperspectral leaf disease detection scenarios.

Table 7 describes the recall of proposed PANet+R-DCNN and PANet+DSKO+R-DCNN models is compared with the existing SOTA models such as SVM+CARS, SVM+SPA, SVM+PCA, BiLSTM+CARS, BiLSTM+SPA, BiLSTM+PCA, DBO-BiLSTM+CARS, DBO-BiLSTM+SPA and DBO-BiLSTM+PCA. These results represent major improvements of up to 5.16% in Corn rust and 3.03% in Corn hybrid compared to even the top-performing DBO-BiLSTM models. PANet+R-DCNN yielded 98.568% recall in Corn rust and 98.148% in Corn hybrid, achieving over 3.86% and 1.54% gains respectively when compared to earlier baselines. Figure 10 describes the accuracy results comparison of proposed and SOTA models on Maize-soybean intercropping system. For the training phase, the proposed model PANet+DSKO+R-DCNN achieved the highest accuracy of 99.858%. Connected to SVM+CARS which recorded 93.6%, there is improvement of 6.66%. Over SVM+SPA and SVM+PCA with accuracy of 89.6% and 84.4% respectively, the gains are 10.26% and 15.26%. PANet+R-DCNN model recorded 98.748% the improvement is 1.05%, proving the effectiveness of the DSKO enhancement.

Table 8 describes the seasonal variations in disease detection accuracy for Maize–soybean and pea–cucumber intercropping systems revealed significant differences, primarily influenced by temperature, lighting, and other environmental factors. During spring (Mar–May), the model achieved the highest detection accuracy, with values ranging from 94.034% to 97.52% for Maize–soybean and 94.014% to 97.411% for pea–cucumber. High-intensity illumination and crop stress caused performance to deteriorate in the summer, although autumn and winter showed comparatively poorer accuracy, especially in the winter when there was no natural light. Thus, variations in temperature and illumination have a significant impact on how accurately diseases are detected in these intercropping systems, with severe circumstances causing performance to noticeably deteriorate.

The descriptive statistics of recall performance for various models in the Maize-Soybean intercropping system, as shown in Table 9 , reveal that the proposed PANet+DSKO+R-DCNN model achieved the highest mean recall of 99.916%, with minimal variability (Std. Dev. 0.148%) and a maximum recall of 100%, indicating highly consistent performance across all disease classes. The PANet+R-DCNN model also performed well with a mean recall of 99.453%, followed by DBO-BiLSTM+CARS with 98.55%. In contrast, classical models such as SVM+PCA and SVM+SPA showed lower mean recall of 81.467% and 87.883%, respectively, with much higher standard deviations, reflecting greater inconsistency and reduced reliability in detecting leaf diseases. The comparative analysis in Table 10 highlights the superiority of PANet+DSKO+R-DCNN over the state-of-the-art models. When compared to SVM+CARS, the proposed model showed the highest improvement in Corn hybrid by 25.54% and Corn rust by 23.16%, while Corn leaf spot and Soybean rust improved by 7.1% and 6.5%, respectively. There was no change in Corn normal and Soybean normal recall. Compared to SVM+SPA, the improvements were substantial, with Corn leaf spot increasing by 29.92%, Corn hybrid by 23.44%, and Soybean rust by 11.5%, while Corn rust improved by 6.55% and Corn normal remained unchanged. Against SVM+PCA, the proposed model achieved remarkable gains of 45.14% for Corn hybrid, 27.44% for Corn rust, 26.24% for Corn leaf spot, and 17.04% for Soybean rust, with minor improvement of 4.2% for Corn normal. Compared with DBO-BiLSTM+CARS, PANet+DSKO+R-DCNN showed moderate improvements, including 5.16% for Corn rust, 3.04% for Corn hybrid, and negligible decrease of 0.14% in Corn leaf spot. Similarly, against DBO-BiLSTM+SPA, the improvements were 9.64% in Corn hybrid, 7.85% in Corn rust, and 4.5% in Corn leaf spot, while the other classes remained unchanged. Compared to DBO-BiLSTM+PCA, the proposed model achieved 27.24% improvement in Corn hybrid, 5% in Corn leaf spot, 4.45% in Corn rust, and no change in other classes. The PANet+DSKO+R-DCNN model consistently outperformed both classical SVM-based models and deep learning-based SOTA models, demonstrating its robustness and effectiveness in accurately detecting leaf diseases across all classes in Maize-Soybean intercropping systems, with improvements ranging from minor gains of 4.2% to major increases of 45.14%.

Despite the promising performance of the proposed hyperspectral intercropping disease detection system, several challenges remain for practical deployment. The model’s accuracy is highly dependent on data quality, and field conditions such as motion artifacts, uneven illumination, and sensor noise can affect predictions. Environmental variability, including changes in lighting, temperature, and humidity, may further influence spectral measurements. Scalability is another concern, as real-time or large-scale applications require substantial computational resources and specialized hardware. Additionally, while the model performs well on maize–soybean and pea–cucumber systems, its generalization to other crops or regions needs further validation. Ensuring consistent hyperspectral imaging precision under field conditions also poses practical challenges. Addressing these limitations will be essential for robust and scalable deployment in precision agriculture.