README.md

metadata

license: apache-2.0
tags:
  - pytorch
  - image-segmentation
  - remote-sensing
  - earth-observation
  - mangroves
  - carbon-credits
  - climate
datasets:
  - sentinel-2
  - global-mangrove-watch
language:
  - en
metrics:
  - iou
pipeline_tag: image-segmentation

NatureCode-Mangroves

Ultimate mangrove monitoring model for carbon credits, biocredits, and nature credits.

Model Description

NatureCode-Mangroves is a state-of-the-art deep learning model for mangrove ecosystem monitoring using satellite imagery. It combines:

• Swin Transformer backbone for global context understanding
• Mamba (State-Space) blocks for efficient long-range dependencies
• UNet decoder with attention gates
• Multi-task learning with specialized output heads

Performance

Metric	Value
Mangrove IoU	94.91%
Parameters	40.2M
Training Data	96 real Sentinel-2 tiles

Intended Use

This model is designed for:

• Mangrove extent mapping (segmentation)
• Carbon stock estimation
• Change detection over time
• Biodiversity assessment
• Blue carbon project monitoring

Carbon Credit Standards

The model outputs are compatible with:

• Verra VCS VM0033 (Tidal Wetland Methodology)
• Gold Standard Mangrove Methodology
• IPCC Tier 2/3 guidelines
• Blue Carbon Initiative

Usage

import torch
from naturecode_mangroves import create_model

# Load model
model = create_model({'in_channels': 19})
checkpoint = torch.load('pytorch_model.bin', map_location='cpu')
model.load_state_dict(checkpoint['state_dict'])
model.eval()

# Prepare input (19 channels: 10 optical + 4 indices + 3 SAR + 2 extra)
# See prepare_input() function for details
input_tensor = torch.randn(1, 19, 256, 256)

# Run inference
with torch.no_grad():
    outputs = model(input_tensor)
    segmentation = outputs['segmentation']  # (1, 2, 256, 256)
    mangrove_prob = torch.softmax(segmentation, dim=1)[:, 1]  # Mangrove probability

Input Format

The model expects 19-channel input:

1. Bands 0-9: Sentinel-2 optical bands (B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12)
2. Bands 10-13: Spectral indices (NDVI, NDWI, MNDWI, CMRI)
3. Bands 14-16: SAR channels (VV, VH, VV/VH ratio) - can be simulated from optical
4. Bands 17-18: Extra features (NDVI duplicate, NIR-SWIR average)

Output Heads

The model produces multiple outputs:

• segmentation: Binary mangrove/non-mangrove segmentation
• agb: Above-Ground Biomass estimation (Mg/ha)
• soc: Soil Organic Carbon (Mg C/ha)
• health: Ecosystem health score (0-1)
• Additional heads for species, change detection, etc.

Training Data

Trained on real Sentinel-2 imagery from:

• Sundarbans (India/Bangladesh)
• Mekong Delta (Vietnam)
• Amazon Delta (Brazil)
• Niger Delta (Nigeria)
• Queensland (Australia)
• Florida (USA)
• And 50+ other global mangrove regions

Labels derived from:

• Global Mangrove Watch (GMW) v4.0
• Spectral index-based pseudo-labels
• Expert validation

Limitations

• Optimized for 10m resolution Sentinel-2 imagery
• Best performance in tropical/subtropical mangrove regions
• SAR channels are simulated when real SAR not available
• Carbon estimates require calibration for specific regions

Citation

@software{naturecode_mangroves_2026,
  author = {NatureCode Team},
  title = {NatureCode-Mangroves: Deep Learning for Mangrove Ecosystem Monitoring},
  year = {2026},
  url = {https://huggingface.co/naturecode/mangroves}
}

License

Apache 2.0