Hyperspectral Data

Hyperspectral imaging provides the highest spectral resolution, capturing 408 spectral bands across the electromagnetic spectrum. This data type is particularly valuable for detailed spectral analysis of biological organisms.

Data Format

  • File Types: .bil (Band Interleaved by Line) with .hdr header files
  • Channels: 408 spectral bands
  • Wavelength Range: 348-809 nm (equally distributed)
  • Bit Depth: 16-bit per band
  • Source: University of St Andrews Lepidoptera Hyperspectral Database

Data Structure

Each hyperspectral image consists of two files:

.bil File

Contains the raw spectral data in band-interleaved format: 

specimen_001.bil  # Raw hyperspectral data

.hdr File

Contains metadata and wavelength information: 

specimen_001.bil.hdr  # Header with wavelength calibration

Wavelength Information

The 408 bands are equally distributed between 348-809 nm: - Band spacing: ~1.13 nm per band - Spectral range: UV to near-infrared - Exact wavelengths: Specified in each .hdr file

Directory Structure

data/hyperspectral/lepidoptera/
├── species_1/
│   ├── specimen_001.bil
│   ├── specimen_001.bil.hdr
│   ├── specimen_002.bil
│   ├── specimen_002.bil.hdr
│   └── ...
├── species_2/
│   ├── specimen_001.bil
│   ├── specimen_001.bil.hdr
│   └── ...
└── metadata.csv

Configuration

For hyperspectral data training:

data: 
  data_dir: "data/hyperspectral/lepidoptera"

output:
  out_dir: "path/to/output/directory"

model:
  name: 'simclr'  
  backbone: 'vit_l_16'
  weights: 'ViT_L_16_Weights' 

augmentations:
  input_size: 224

Band Sampling

Full Spectrum

Use all 408 bands: 

python train/simclr_birdcolour_kornia_hyperspectral.py

RGB Approximation

Use bands closest to RGB wavelengths (450nm, 550nm, 650nm): 

python train/simclr_birdcolour_kornia_hyperspectral.py --rgb-only

Custom Band Selection

Sample specific bands: 

python train/simclr_birdcolour_kornia_hyperspectral.py \
  --sample-bands 0,50,100,150,200,250,300,350,400

Data Reading

The framework uses a Python reimplementation of the MATLAB read_iml_hyp.m function:

from scripts.read_iml_hyp import read_hyperspectral_image

# Read hyperspectral data
data, wavelengths = read_hyperspectral_image('specimen_001.bil')
print(f"Data shape: {data.shape}")  # (height, width, 408)
print(f"Wavelengths: {wavelengths[:5]}...")  # First 5 wavelengths

RGB Conversion

For visualization, convert to RGB using specific wavelengths:

# Wavelengths closest to RGB
rgb_bands = {
    'red': 650,    # nm
    'green': 550,  # nm  
    'blue': 450    # nm
}

# Find closest band indices
rgb_indices = find_closest_wavelengths(wavelengths, rgb_bands)
rgb_image = data[:, :, rgb_indices]

Training Options

Loss Functions

Choose between different loss strategies:

# Global loss only (standard SimCLR)
python train/simclr_birdcolour_kornia_hyperspectral.py --loss-type global

# Local patch-based loss
python train/simclr_birdcolour_kornia_hyperspectral.py --loss-type local

# Combined global and local loss
python train/simclr_birdcolour_kornia_hyperspectral.py \
  --loss-type combined \
  --global-weight 0.7 \
  --local-weight 0.3

Memory Optimization

For large hyperspectral data:

batch_size: 8  # Reduce for 408-channel data
num_workers: 4
pin_memory: true

Example Usage

# Train on full hyperspectral data
python train/simclr_birdcolour_kornia_hyperspectral.py \
  --config configs/config_kornia_hyperspectral.yaml

# Train on sampled bands
python train/simclr_birdcolour_kornia_hyperspectral.py \
  --sample-bands 0,25,50,75,100,125,150,175,200 \
  --loss-type combined

# Generate embeddings only (skip training)
python train/simclr_birdcolour_kornia_hyperspectral.py \
  --test-embeddings-only \
  --model-checkpoint outputs/hyperspectral_model.ckpt