Colour-Mix Augmentation Dataset for Rice Sheath Blight Severity Estimation
Description
This dataset is based on the hypothesis that accurate plant disease severity estimation under real field conditions requires high-quality, field-collected images that capture natural variability in lesion color, texture, illumination, and background context, as severity discrimination in rice sheath blight is driven more by subtle chromatic and spatial cues than by structural deformation. The dataset comprises original images of rice plants affected by sheath blight, collected directly from agricultural fields under natural lighting and environmental conditions, without any synthetic augmentation or post-processing. Images were acquired across multiple observation periods and include inherent variations in soil background, water presence, plant orientation, and growth stage. Each image is annotated into one of six disease severity levels, representing progressive stages of infection, based on visual assessment of lesion extent, lesion height progression along the sheath, and symptom color intensity following standard plant pathology practices. The data reveals fine-grained transitions between severity classes, where differences are primarily reflected in lesion density, chromatic intensity, and spatial spread rather than major anatomical changes. As such, the dataset provides a realistic benchmark for developing and evaluating disease severity estimation models that must operate under uncontrolled field conditions. It can be readily reused for training and benchmarking machine learning and deep learning models, studying color- and texture-driven disease progression, and supporting field-deployable decision-support systems for crop health monitoring.
Files
Steps to reproduce
Images of rice sheath blight were collected from paddy fields located in Kottekkad, Kuttur, Thrissur, Kerala, India (latitude: 10.56581, longitude: 76.19060. The images were captured from the rice variety Jyotsna during the Puncha season (January to March), which corresponds to the Summer/Third Crop in the region. Data acquisition was performed using an iPhone 6s 12-megapixel camera. The collected images are in JPG format and exhibit natural variations in backgrounds, illuminations, orientations, and dimensions, reflecting real-world field conditions. A total of 1,609 images were collected after removing blurred and duplicate images. These were then reviewed by a domain expert from the Department of Plant Pathology, College of Agriculture, Padannakkad, Kasargod for disease severity classification. Visual scoring of incidence and severity of disease was done using the standard evaluation scale. For reproducibility of the augmentation experiments, all augmentation methods were implemented using a standardized Python-based workflow. Conventional augmentations, Mixup augmentations and GAN based augmentations were implemented following their original formulations. The proposed Mask-Guided Colour-Mix augmentation was developed by first converting RGB images to HSV color space to decouple illumination from chromatic symptom features, followed by hierarchical logical masking operations to isolate leaf and lesion regions using threshold-based and morphological operations. Refined foreground masks were then used in an inverted masking process to perform bidirectional exchange of foreground plant regions and background environmental contexts between image pairs. All experiments were executed on a Linux-based system using Python
Institutions
- APJ Abdul Kalam Technological University