Colour-Mix Augmentation Dataset for Rice Sheath Blight Severity Estimation

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.