AMFITRITE-Sentinel2-HAB-RDNet

This model is a deep learning classifier designed to detect Harmful Algal Blooms (HABs) in inland waters using Sentinel-2 Level-2A satellite imagery. It operates as a binary classifier, determining whether a water surface patch contains a bloom or not.

The model is built on the RDNet Base architecture. It was initialized with weights pretrained on BigEarthNet v2.0 (rdnet_base-s2-v0.2.0) and fine-tuned on the Amfitrite-Inland-Waters-HAB-Sentinel2 dataset.

Dataset & Training Details

The model was fine-tuned using the Amfitrite-Inland-Waters-HAB-Sentinel2 dataset.

• Original Classes: The source dataset contains three indicative classes: High, Moderate and Low.
• Preprocessing for Detection: To create a robust binary detector, these classes were merged:
  • Bloom (Class 1): Merged High and Moderate samples.
  • No Bloom (Class 0): Mapped from Low samples.

This merging strategy allows the model to focus on the critical distinction between "Clear Water" and "Algal Presence".

Spectral Awareness

Utilizes 10 spectral bands from Sentinel-2 in the following order (B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12) like the BigEarthNet v2.0 RDNet Base pretrained only with Sentinel-2 data.

Performance

The model was evaluated on a strictly stratified test set to ensure reliability.

Metric	Test Set	Validation Set
F1 Score	0.926	0.934
Accuracy	0.898	0.909
Balanced Acc	0.892	0.909

Limitations

• The model expects standard L2A Sentinel-2 surface reflectance values (scaled 0-10000).
• Performance may vary on water bodies that were not well-represented in the training set (e.g. with extremely high sediment loads).

Preprocessing Requirements

To successfully run inference with this model, your input data must meet the following criteria:

1. Dimensions: The input image must be resized or cropped to exactly 256 x 256 pixels.
2. Channel Order: The tensor must contain 10 bands in the following specific order: B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12.
3. Normalization: The raw Sentinel-2 L2A pixel values (usually ranging from 0 to ~10000) must be divided by 10000.0 so that the input values generally fall between 0.0 and 1.0.
4. Shape: The final tensor passed to the model must have the shape (Batch_Size, 10, 256, 256).

How to Use

This model requires the timm library.

1. Installation

pip install torch timm rasterio

2. Inference Code

import rasterio
import torch
import timm
import numpy as np
import torch.nn.functional as F

# 1. Load Model
model = timm.create_model('rdnet_base', pretrained=False, num_classes=2, in_chans=10)

# Load weights (assuming the file is downloaded locally as 'model.pth'), use huggingface_hub to download directly
state_dict = torch.load("model.pth", map_location='cpu')
model.load_state_dict(state_dict)
model.eval()

def predict_patch(img_tensor):
    """
    Args:
        img_tensor (torch.Tensor): A tensor of shape (10, Height, Width) containing Sentinel-2 bands.
    """
    # 1. Add batch dimension -> Shape: (1, 10, H, W)
    img_tensor = img_tensor.unsqueeze(0) 
    
    # 2. Force resize to the required 256x256
    if img_tensor.shape[-2:] != (256, 256):
        img_tensor = F.interpolate(img_tensor, size=(256, 256), mode='bilinear', align_corners=False)
    
    # 3. Normalize Sentinel-2 DN values (0-10000) to 0-1 range
    img_tensor = img_tensor / 10000.0
    
    # 4. Predict
    with torch.no_grad():
        output = model(img_tensor)
        prob = torch.softmax(output, dim=1)[0, 1].item() # Probability of Bloom
        
    return prob

# --- Example Usage ---
# Dummy tensor representing a 10-band image that is 365x365 pixels
sample_image = torch.rand(10, 365, 365) * 5000
score = predict_patch(sample_image)
print(f"Score: {score:.1%}")

Citations & References

If you use this model in your research, please cite the following papers for the underlying architecture and pre-training:

1. This Model: Pikounis, K. (2026). Sentinel-2 HAB Detector (RDNet Base). Amfitrite Project.

@misc{pikounis2026hab,
  author       = {Pikounis, Kostas},
  title        = {Sentinel-2 HAB Detector (RDNet Base)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/kostaspic/AMFITRITE-Sentinel2-HAB-RDNet}},
  organization = {Amfitrite Project}
}

2. Amfitrite Dataset:

• kostaspic/amfitrite-inland-waters-hab-sentinel2

3. BigEarthNet v2.0 (Pretraining Source): K. Clasen, L. Hackel, T. Burgert, G. Sumbul, B. Demir, V. Markl, "reBEN: Refined BigEarthNet Dataset for Remote Sensing Image Analysis", IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2025.

@inproceedings{clasen2025refinedbigearthnet,
  title={{reBEN}: Refined BigEarthNet Dataset for Remote Sensing Image Analysis},
  author={Clasen, Kai Norman and Hackel, Leonard and Burgert, Tom and Sumbul, Gencer and Demir, Beg{\"u}m and Markl, Volker},
  year={2025},
  booktitle={IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
}

4. Library & Framework: L. Hackel, K. Clasen, B. Demir, "ConfigILM: A General Purpose Configurable Library for Combining Image and Language Models for Visual Question Answering.", SoftwareX 26 (2024): 101731.

@article{hackel2024configilm,
  title={ConfigILM: A general purpose configurable library for combining image and language models for visual question answering},
  author={Hackel, Leonard and Clasen, Kai Norman and Demir, Beg{\"u}m},
  journal={SoftwareX},
  volume={26},
  pages={101731},
  year={2024},
  publisher={Elsevier}
}