A High-Resolution Proximal Multispectral Dataset for Early Detection of Brown Spot (Bipolaris oryzae) in Tropical Rice Crops

Description

This dataset provides high-resolution proximal multispectral imagery specifically designed for the detection and classification of Brown Spot disease (Bipolaris oryzae) in rice crops (Oryza sativa L.). The data were acquired in the tropical savannas of Casanare, Colombia (Yopal and Aguazul municipalities), representing one of the most productive rice regions in the country. 1. Data Acquisition and Sensor Specifications: The imagery was captured using a DJI P4 Multispectral (P4M) platform. The system records six discrete spectral channels: one RGB (visible) and five monochromatic bands centered at Blue (450 nm), Green (560 nm), Red (650 nm), Red-Edge (730 nm), and Near-Infrared (840 nm). To resolve early-stage necrotic lesions (often smaller than 1.5 mm), a proximal sensing strategy was implemented, maintaining a sensor height of 30–45 cm from the canopy with a 45° oblique perspective. 2. Dataset Composition: The dataset consists of 2,772 multispectral sets (totaling over 13,000 individual TIFF files) categorized into two primary classes: Healthy: Asymptomatic rice leaves confirmed under optimal nutrient conditions. Diseased (Brown Spot): Leaves exhibiting characteristic oval lesions with necrotic centers and chlorotic halos, ranging from early to advanced infection stages. 3. Ground Truth Validation: Class labels were assigned through a rigorous \textit{in-situ} inspection protocol. Ground truth was validated by expert agronomists from the TICTROPICO research group (Unitropico) and local technical assistants, following the phytosanitary monitoring standards for Colombian tropical rice. 4. Potential Applications: This dataset is optimized for training and benchmarking Deep Learning architectures (e.g., CNNs, ConvNeXt, Vision Transformers) and for the development of high-dimensional Bio-Spectral Tensors. It serves as a benchmark for precision phytopathology and autonomous crop health monitoring in tropical agricultural ecosystems.

Steps to reproduce

1. Proximal Data Acquisition:Use a DJI P4 Multispectral (P4M) platform with its integrated 6-sensor array (RGB, Blue, Green, Red, Red-Edge, and NIR).Conduct flights or handheld captures at a proximal height of 30–45 cm from the rice canopy, maintaining a 45° sensor inclination to resolve lesions smaller than 1.5 mm.Ensure captures are performed between 09:00 and 13:00 hours under optimal radiometric conditions, utilizing a Spectralon reflectance panel for initial calibration.Follow a systematic W-pattern sampling across the field to ensure representative class distribution (Healthy vs. Diseased).2. Image Pre-processing and Alignment:Perform sub-pixel band registration using an affine transformation or homography model to correct parallax errors between the five monochromatic sensors.Use the Green band (560 nm) as the reference "anchor" for spatial alignment.Apply a binary vegetation mask using Otsu’s adaptive thresholding on the NIR band (840 nm) to isolate foliar tissue from the background (soil and water).3. Feature Engineering and Index Audit:Evaluate a set of 17 candidate multispectral indices to determine the optimal feature space.Quantify the discriminative power of each index using the Fisher Discriminant Criterion (F). Based on the ranking, construct a 3-channel Bio-Spectral Tensor using the Chlorophyll Vegetation Index (CVI), Modified Anthocyanin Content Index (MACI), and Green Normalized Difference Vegetation Index (GNDVI).4. Model Training and Validation:Input the normalized Bio-Spectral Tensors ($224 \times 224 \times 3$) into a ConvNeXt-Tiny architecture.Apply transfer learning from ImageNet-1K and utilize a Cosine Annealing Learning Rate scheduler for optimization.Validate the diagnostic logic through Grad-CAM heatmaps to ensure the network targets the physiological markers of Bipolaris oryzae

Institutions

• Universidad de Antioquia

  Antioquia, Medellín
• Fundación Universitaria Internacional del Tropico Américano

  Casanare Department, Yopal

Categories

Licence