Neural Network-Based Early Detection of Wheat Stripe Rust Disease for Enhanced Crop Management

For majority of the world's population, wheat is a basic food source and among the most grown crops in the planet. Still, wheat output is beset with difficulties; illnesses seriously compromise food security and yield. Of these diseases, stripe rust-caused by the fungus Puccinia striiformis f. sp. tritici, is most destructive [1]. If not early on recognised and controlled, stripe rust disease can seriously reduce wheat harvests. Preventing the quick spread of the illness depends on early identification, which also enables quick interventions including the use of resistant types or fungicides. Although efficient, traditional approaches of spotting stripe rust-field inspections and laboratory-based methods-often demand specialised knowledge, time, and effort [2], these approaches might not always enable real-time observation over large agricultural tracts. Early-stage detection is crucial for efficient disease control, so development of precise and automated detection systems employing cutting-edge technology has attracted increasing interest [3]. One such interesting method uses artificial intelligence (AI) and, more especially, neural networks to early on detect stripe rust in wheat harvests [4]. A subset of machine learning and artificial intelligence, neural networks have shown amazing performance in many picture recognition and classification challenges, so they are a good choice for the detection of plant diseases. These networks may learn patterns and features from big datasets, therefore enabling the identification of minor visual symptoms that might not be readily apparent to human observers. Using neural networks, scientists hope to build systems that can automatically examine wheat plant photos and identify early stripe rust infection signals. Early identification could help farmers and agricultural professionals to respond right away, therefore reducing crop damage and guaranteeing maximum harvests [5, 6]. Systems for disease detection in plants have been driven even more by recent developments in computer vision, a branch of artificial intelligence that lets robots read and understand visual data. These systems can record and examine the minute features of plant leaves using high-resolution images, therefore spotting the early start of stripe rust depending on distinctive visual symptoms such yellowish streaks along the veins of the leaves [7-9].

Early identification of stripe rust disease using neural networks presents various benefits over more traditional techniques. First, by rapidly and effectively processing vast amounts of data, neural networks enable constant real-time surveillance of wheat fields. In big-scale agricultural operations where hand inspections are not feasible, this is especially helpful. Second, these artificial intelligence-based systems can be combined with ground-based sensors, satellites, or drones to gather photos across large distances, therefore offering a whole picture of crop condition [10]. Moreover, neural networks can be trained to identify between several diseases that might have similar symptoms, therefore lowering the possibility of misdiagnosis [11-13]. Implementing focused treatments and avoiding pointless treatments that can damage the crop or the surroundings depend on this degree of accuracy. The possibility for scalability of neural networks is another important factor in their application for illness diagnosis. From mobile devices to cloud-based systems, once a neural network model has been trained on a strong dataset of infected and healthy plant photos it can be used on many platforms. Small-scale and large-scale farming operations benefit much from this adaptability since it lets farmers and agronomists access the technology anywhere [14, 15]. Furthermore, major gains in the accuracy and dependability of these detection systems result from ongoing developments in neural network topologies like convolutional neural networks (CNNs). Popular choice for researchers working on agricultural uses, CNNs-which are especially well-suited for image-based tasks-have showed considerable potential in spotting disease trends in plants [16-18].

Although neural networks show promise for early-stage stripe rust detection, some issues still need to be resolved if we are to guarantee general acceptance of this approach. The necessity for high-quality, annotated datasets of wheat plants afflicted with stripe rust at several phases of growth presents one of the main difficulties. Establishing such datasets calls for cooperation between AI researchers and agricultural professionals to properly classify photographs and guarantee that the model picks the right attributes linked with the disease [19, 20]. Furthermore, influencing the accuracy of image-based detection systems are environmental elements including camera resolution, lighting conditions, and the existence of other plants or weeds in the field. Dealing with these difficulties calls for greater investigation and the creation of more complex algorithms able to change depending on the field conditions. a major progress in agricultural technology is made with the early-stage identification of stripe rust in wheat employing neural networks [21-23]. Using artificial intelligence and machine learning will help one create systems that can automatically identify diseases at an early stage, hence allowing quick and efficient treatments. Although there are obstacles to overcome, the possible advantages of applying neural networks for disease detection are significant; they provide a scalable, accurate, efficient answer to one of the most urgent problems in wheat output. The integration of artificial intelligence-based detection systems in agriculture could be very important in guaranteeing world food security and sustainability as research in this field develops changes [24-26].

This work used a dataset of wheat crops gathered from the IARI field during the rabi season of 2019-2020, using a Sony IMX363 RGB camera to get high-quality photos. Images of nitrogen deficient wheat leaves with abiotic stress and leaf rust-common biotic stress signs visible at the booting stage-are included in the dataset and Figure 1 shows the proposed flowchart.

Healthy leaf photos were also included as a baseline for comparison and constant illumination conditions were kept to guarantee accurate analysis. Otsu-based background segmentation, scaling images to 135x270 pixels, contrast stretching, and Contrast Limited Adaptive Histogram Equalisation (CLAHE) to improve stress indicator visibility included preprocessing processes. Variability was raised via rotation and scaling among other data augmentation methods. Class distributions and visual aspects were evaluated by exploratory data analysis (EDA), therefore guaranteeing equal representation among the stress categories. CropStressNet is a deep learning model designed to categorise photos into three: nitrogen-deficient, leaf rust-affected, and healthy. With categorical cross-entropy as the loss function and Adam optimiser for maximum efficiency, the model architecture included cutting-edge ideas such residual and squeeze-excite blocks. Hyperparameters were painstakingly tweaked to maximise classification accuracy.

1.png

Figure 1. Proposed flowchart

3.1 Data collection

The dataset was gathered from wheat crops in the IARI field during the rabi season of 2019-20 using a Sony IMX363 RGB camera. The camera was set to a resolution of 4032×3024 pixels, with ISO set at 100 and a shutter speed of 1/200 seconds to ensure high-quality, sharp images. Images were captured under consistent natural daylight conditions between 10:00 AM and 2:00 PM to minimize variations in lighting. No artificial lighting or flash was used to maintain field authenticity. It centres on wheat leaves showing nitrogen deficit, abiotic stress, and leaf rust-a common biotic stress-both of which are evident in leaves at the booting stage of the crop. These pressures cause symptoms that greatly change leaf colour, texture, and structure, therefore allowing obvious variation. Consistent illumination was used to guarantee dependability and clarity for analytical needs on images. Accurately distinguishing and identifying stressed leaves depends on a baseline, thus this dataset comprises healthy wheat leaf images acquired under controlled conditions. From chlorosis brought on by nitrogen shortage to rust pustules unique to fungal infection, the images in this collection mirror several stress expressions. The dataset has been split into training, validation, and test sets as well as arranged into several folders allowing simple separation between stress and health. This system and inclusion of control images provide a strong basis for teaching deep learning models to precisely identify and classify indicators of biotic and abiotic stress in wheat, hence advancing precision agricultural methods.

3.2 Data preprocessing

Preprocessing in this dataset consisted in several phases to equip photos for deep learning investigation. Background segmentation was first used using Otsu-based masking to isolate leaf structures and eliminate undesired background features, therefore focusing the attention on stress patterns in the leaves. Images were then downsized consistently to 135×270 pixels to preserve interoperability across deep learning models, therefore offering a consistent training input size. Then, contrast stretching was applied to improve fine detail visibility-especially for stress-induced discolouration or texture changes-that are absolutely necessary for nitrogen deficiency and leaf rust detection. CLAHE further enhanced local contrast by varying brightness in smaller areas without increasing noise, hence emphasising stress signs. Custom Python routines were also used to apply brightness and contrast changes, therefore guaranteeing ideal visual clarity for analysis. Random noise was included into the dataset to simulate possible real-world settings and diversity the training data, hence adding variability and strengthening of the model. Emphasising stress-related properties in wheat leaf images under consistent lighting and quality, these preprocessing techniques together guaranteed the dataset was well-suited for accurate and resilient deep learning model performance.

3.3 Exploratory data analysis of wheat yellow rust dataset

Image distribution and features of the dataset were evaluated using EDA. Class distribution graphs provide information on the image count among nitrogen-deficient, leaf rust-affected, and healthy wheat leaves. Analysing these distributions guaranteed that every class was represented fairly, so enabling balanced model training. Maintaining consistent input size for the deep learning models was confirmed by image dimension consistency. Visual analysis of sample photos from each category revealed clear trends including chlorosis in nitrogen-deficient leaves, indicated by yellowing resulting from lack of chlorophyll, and the presence of rust pustules in sick leaves, suggestive of fungal infection. These visual characteristics were meticulously observed since they are essential markers for different forms of stress classification in the next model building stage. Understanding these picture features and class-specific patterns helped us to improve our method of feature selection by concentrating on qualities most important for stress identification. Building an accurate and efficient model for biotic and abiotic stress in wheat leaves depends on a deeper knowledge of the dataset, which this extensive EDA guaranteed.

Figure 2 shows the dataset's class distribution. The x-axis shows several class labels ranging from -0.5 to 3.5; the y-axis shows the count of instances for every class. Class 3 stands out as having the most frequency-more than 350 cases-which suggests a possible dataset imbalance.

2.png

Figure 2. Bar graph of dataset's class distribution

In Figure 3, four classes of leaf samples-Class 0, Class 1, Class 2, Class 3-are shown in the illustration. Every class reflects unique leaf traits that are vital for our work on crop stress classification. These graphic examples help one to grasp the differences related with nitrogen deficit and disease influence.

3.png

Figure 3. Four classes of leaf samples

4.png

Figure 4. Distribution graphs of the image dimensions

In Figure 4, two distribution graphs showing the image dimensions in the dataset help to depict this. Centred on 135 pixels, the left plot displays the width distribution with a homogeneous count of roughly 1,200 photos at this width. With about 1,200 photos, the correct plot shows the height distribution-mostly around 270 pixels. Essential for homogeneity in image processing and analysis, these visualisations show a constant size over the dataset.

3.4 Contrast enhancement of wheat leaf images using CLAHE

By selectively increasing contrast, particularly in areas afflicted by stress that show faint discolourations, CLAHE was used to improve the visibility of important elements in the dataset images. Divining an image into small pieces, or tiles, CLAHE essentially modulates the intensity values and improves local contrast by individually performing histogram equalisation to each tile. In this instance, a tile grid size of (8, 8) was selected to split the photos into reasonable sections to emphasise localised changes in texture and colour over the leaf surfaces. A 2.0 clip limit was also established to stop noise from over-amplifying and therefore masking important information. While preserving the natural gradient transitions of healthy leaf tissues, CLAHE reduces the contrast in each tile, so enhancing stress-related details like the chlorosis in nitrogen-deficient leaves or rust pustules in infected leaves. By sharpening edges and increasing the visibility of discolouration patterns, this preprocessing phase helps the model to distinguish between stress-affected and healthy areas, hence enabling more accurate feature extraction and classification during model training. By improving relevant features without overloading the model with pointless noise, the balanced approach of CLAHE therefore best prepares the images for deeper investigation.

3.5 Model implementation

CropStressNet is the name given to the deep learning model created for nitrogen-deficient, leaf rust-affected, and healthy classification of wheat leaves. This name stresses its use in precision agriculture and indicates the particular attention of the model on stress in crop leaves.

Three categories- nitrogen-deficient, leaf rust-affected, and healthy-were established from wheat leaves using a deep learning-based model. Residual blocks, squeeze-excite blocks, and self-attention mechanisms are among the sophisticated approaches the architecture combines to help it to capture minute details from the images. Resizing the model's input photos to a uniform dimension of 270×135 270×135 pixels helped to enable fit with the deep learning architecture. With a 70:15:15 ratio, the dataset was split into training, validation, and test sets guaranteeing accurate evaluation. Data augmentation using an Image Data Generator applied changes like rotation, scaling, and flips to boost variability and resilience during training. The loss function for the model optimisation was categorical cross-entropy, formally stated as.

$L(y, \hat{y})=-\sum_{i=1}^c y_i \log \left(\hat{y}_i\right)$           (1)

where, $y$ is the true distribution and $\hat{y}$ is the predicted distribution across C classes. The Adam optimiser was chosen because of its adaptive learning features, which enable dynamic change of the learning rate. This improves the crop stress detecting precision of the model. Multiple convolutional layers formed part of the architecture, and channel-wise feature responses were recalibrated using a squeeze-excite method. Under different circumstances, the model's performance was assessed using accuracy as the metric, therefore guaranteeing its ability to clearly differentiate stressed from healthy leaves. We modified hyperparameters including dropout rates, learning rates, and filter count to maximise performance even more and Table 1 shows hyperparameter details apply for improvement of result.

Table 1. Hyperparameter details

The Adam optimizer can be summarized by the following parameter update equation:

$\theta_t=\theta_{t-1}-\frac{\alpha \widehat{m}_t}{\sqrt{\widehat{v}_t+\in}}$         (2)

where,

$\theta_t$: Updated parameters at time (t)

$\theta_{t-1}$: Previous parameters

α: Learning rate

$\widehat{m}_t$: Bias-corrected first moment estimate

$\hat{v}_t$: Bias-corrected second moment estimate

$\in$: A small constant to prevent division by zero

Guiding model optimisation for multi-class classification tasks, categorical cross-entropy quantifies the difference between actual and expected class distributions.