NUST Wheat Rust Disease Dataset: High Resolution Annotated Multileaf Images for Wheat Stripe Rust Semantic Segmentation

Overview

The NUST Wheat Rust Disease Dataset, known as NWRD, is an open source image dataset developed for the semantic segmentation of wheat stripe rust disease. It contains real world images of healthy and diseased wheat crops collected under natural field conditions.

The dataset was designed to support the precise identification and pixel level localisation of wheat stripe rust. Unlike conventional classification datasets that assign a single disease label to an entire image, NWRD provides detailed segmentation masks that show the specific areas affected by rust.

The dataset is intended for computer vision, deep learning, plant disease analysis, precision agriculture, and automated crop monitoring research.

Research background

Wheat stripe rust disease is a serious threat to wheat production and food security. The disease can spread rapidly through a field and can cause substantial crop losses when it is not detected and treated at an early stage.

Traditional field inspection is normally performed by trained personnel. This process is time consuming, labour intensive, and difficult to apply across large agricultural areas. Computer vision and deep learning can support faster and more consistent detection of infected crop areas.

Many existing wheat disease datasets focus on image classification. These datasets identify whether an image contains a healthy or diseased leaf, but they do not provide the exact location or shape of the infected region. The NWRD dataset addresses this limitation by providing manually prepared pixel level annotations for wheat stripe rust segmentation.

Dataset purpose

The main purpose of the dataset is to support the development and evaluation of semantic segmentation models that can identify rust affected areas within complex wheat field images.

The dataset can help researchers develop systems that estimate the location and extent of disease spread. This information may support early intervention, targeted treatment, crop monitoring, and more efficient use of fungicides.

A segmentation model trained on this type of data can potentially be used with field cameras, mobile devices, drones, or other agricultural monitoring systems.

Dataset name

The complete name of the dataset is:

NUST Wheat Rust Disease Dataset

The abbreviation used for the dataset is:

NWRD

NUST refers to the National University of Sciences & Technology. WRD refers to wheat stripe rust disease.

Dataset size

The NWRD dataset contains 100 high resolution images of wheat crops.

The images show multiple wheat leaves within each scene. They include healthy leaves, rust affected leaves, crop structures, soil, shadows, overlapping plants, and natural field backgrounds.

The paper describes NWRD as a segmentation dataset with a slightly wide angle view and multiple leaves in each image. This makes it different from many existing wheat disease datasets that contain only one isolated leaf against a simple background.

Data collection period

The field data were collected during one wheat crop season.

The wheat crop was cultivated in November 2021 and harvested in May 2022. Wheat stripe rust generally began to appear in February and continued to spread during March and April.

Images were collected progressively during the growing season. Data acquisition was carried out during both morning and afternoon periods to capture natural changes in illumination and field appearance.

Data collection location

The images were collected from wheat fields at the National Agricultural Research Centre in Islamabad, Pakistan.

The data collection was conducted in collaboration with NARC. The centre provided access to experimental farms, crop fields, and other agricultural facilities required for the research.

The paper identifies the collection location as approximately six kilometres southeast of Islamabad.

Image acquisition

The dataset contains camera based RGB images of healthy and rust affected wheat crops.

Images were captured under different illumination conditions during the morning and afternoon. The collection includes different viewing angles, field backgrounds, levels of disease visibility, and crop densities.

The reported image resolutions include:

• 6016 × 4000 pixels
• 4608 × 3456 pixels

These high resolution images preserve small rust regions and detailed crop structures. They also create a demanding segmentation task because the disease may occupy only a small part of the complete image.

Annotation process

Each image was manually annotated at a fine level to identify rust affected regions.

The ground truth annotations represent the disease regions as binary segmentation masks. These masks separate rust affected pixels from non-rust pixels.

Manual annotation was necessary because wheat stripe rust does not always appear in a regular or predictable geometric shape. Disease regions can be thin, scattered, curved, partially hidden, or distributed across several overlapping leaves.

The dataset includes densely annotated images containing several diseased leaves within a single image. The annotations are designed to preserve the complex and irregular boundaries of visible rust regions.

Segmentation classes

The dataset supports binary semantic segmentation using two principal classes:

Rust class

Pixels representing visible wheat stripe rust disease.

Non-rust class

Pixels representing healthy leaves, stems, soil, background objects, shadows, and other areas not labelled as rust.

The rust class normally represents a smaller portion of each image. This creates a class imbalance between rust and non-rust pixels.

Image characteristics

The dataset contains challenging real world conditions rather than controlled laboratory images.

Important characteristics include:

• Multiple leaves in a single image
• High resolution images
• Dense manual annotations
• Arbitrarily shaped disease regions
• Natural field backgrounds
• Busy and complex scenes
• Clear and fuzzy disease regions
• Partially occluded leaves
• Different stages of disease spread
• Variations in illumination
• Different viewing distances
• Aerial, ground, and front views
• Strong imbalance between rust and non-rust regions

These properties make the dataset suitable for evaluating whether segmentation models can operate under realistic agricultural conditions.

Image complexity categories

The images were grouped into two broad complexity categories:

Easy images

These images provide a closer view of wheat leaves and make the disease regions comparatively easier to identify.

Difficult images

These images use a slightly wider viewing angle and contain more leaves, background clutter, occlusions, and smaller visible disease regions.

This distinction can support experiments on model robustness and performance under different levels of visual complexity.

Recommended machine learning task

The primary task associated with the dataset is binary semantic segmentation.

A model receives a wheat field image as input and produces a pixel level mask identifying the areas affected by wheat stripe rust.

The dataset may also support related tasks such as:

• Wheat rust detection
• Disease region localisation
• Disease severity estimation
• Infected area measurement
• Patch classification
• Crop health monitoring
• Early disease detection
• Agricultural image analysis
• Domain adaptation
• Transfer learning
• Class imbalance research

Benchmark model

The associated research evaluated the dataset using a UNet semantic segmentation model.

Because the original images are high resolution, the researchers divided the images and masks into smaller patches of 128 × 128 pixels. Three patch generation methods were investigated:

• Grid patching
• Adaptive patching
• Adaptive patching with feedback

Adaptive patching selects image areas containing useful rust information. Adaptive patching with feedback also introduces difficult non-rust regions that the model previously misclassified. This allows the model to learn from rust regions and false positive predictions.

Preprocessing

The paper describes several preprocessing operations used for model development.

The images were downsampled to reduce computational requirements. Lanczos filtering was used to preserve visual details during resizing.

The images and corresponding masks were divided into 128 × 128 pixel patches. Patch level data augmentation included horizontal and vertical flipping.

For the benchmark experiments, the images were randomly divided into:

• 80 percent training
• 10 percent validation
• 10 percent testing

The preprocessing pipeline and model were implemented using Python and PyTorch.

Benchmark results

For the complete set of 100 images, the adaptive patching with feedback approach using an input stride of 32 achieved the following results for the rust class:

• Precision: 0.506
• Recall: 0.624
• F1 score: 0.557

The paper also reports that the adaptive patching with feedback technique reduced the number of unnecessary patches and decreased model training requirements compared with standard grid patching.

These results provide a baseline for future research. They should not be treated as a maximum performance level for the dataset.

Potential applications

The dataset may be used in research and development involving:

• Wheat stripe rust segmentation
• Plant disease detection
• Crop disease severity estimation
• Precision agriculture
• Smart farming
• Agricultural robotics
• Drone based crop inspection
• Mobile crop monitoring
• Automated field surveillance
• Computer vision
• Deep learning
• Image segmentation
• Sustainable crop protection
• Targeted fungicide application
• Food security research
• Agricultural decision support systems

The segmentation masks may also support methods that calculate the proportion of an image affected by disease.

Dataset strengths

The main strengths of the dataset include its real world field setting, high image resolution, manually prepared annotations, multiple leaves per image, and complex natural backgrounds.

The dataset captures disease regions with irregular shapes and varying levels of visibility. It therefore provides a more challenging test environment than datasets containing isolated leaves against uniform backgrounds.

Limitations

The dataset contains 100 images, which is relatively limited for training large deep learning models from the beginning.

The rust class occupies a much smaller area than the non-rust class. This creates class imbalance and may cause a model to favour healthy or background regions.

Some disease regions are small, blurred, partially hidden, or difficult to distinguish from natural changes in leaf colour.

The dataset was collected at one research location in Islamabad during one crop season. Differences in geography, wheat variety, camera equipment, weather, soil, crop management, and disease stage may affect how well a trained model performs in other environments.

The dataset focuses on wheat stripe rust. It should not automatically be treated as a labelled dataset for leaf rust, stem rust, powdery mildew, or other wheat diseases.

Responsible use

The dataset is intended for research, education, agricultural monitoring, and development of plant disease analysis systems.

Predictions produced by models trained on this dataset should be validated before they are used for crop treatment decisions. Disease management and fungicide recommendations should involve qualified agricultural specialists.

The dataset should not be presented as a replacement for professional field inspection. It is more appropriately used as a research resource and as a component of decision support systems.

Data availability

The paper states that the NWRD dataset, pretrained models, and associated implementation were made publicly accessible through the project repository.

Repository:
https://github.com/dll-ncai/NUST-Wheat-Rust-Disease-NWRD

Funding

The research was partially supported by Carl Zeiss Stiftung, Germany, through the Sustainable Embedded AI project, project number P2021-02-009.

The research also received funding from the German Academic Exchange Service, DAAD.

Related publication

Anwar, H., Khan, S. U., Ghaffar, M. M., Fayyaz, M., Khan, M. J., Weis, C., Wehn, N., and Shafait, F. 2023. The NWRD Dataset: An Open-Source Annotated Segmentation Dataset of Diseased Wheat Crop. Sensors, 23, 6942.

DOI: 10.3390/s23156942