martinkorelic/dpr-zoo-models

🛰️ DPR Zoo Segmentation Models for Sentinel-2

This repository hosts a collection of semantic segmentation models fine-tuned specifically for Sentinel-2 satellite imagery, developed as part of the DPR Zoo Segmentation Hub framework.

These models are designed to perform pixel-level classification on unseen Sentinel-2 data and can be used for:

• ✅ Single-scene segmentation of satellite imagery
• 🔄 Temporal analysis using time series of Sentinel-2 observations
• 📈 Extraction of key environmental statistics (e.g., NDVI, NDMI, NDWI, etc.) based on segmented pixel classes
• ☁️ Cloud-aware filtering to ensure robust statistics across time

Each model is trained on domain-specific datasets and is integrated into a modular serving backend, enabling easy evaluation, comparison, and deployment of new models. The system also supports integration with S3 for data persistence and dataset generation.

Model details

Model Name	Architecure	Dataset Link	Train/Test Dataset Size	Epochs	Number of Labels	Sentinel-2 Bands	Mean IoU	F1 Score (Weighted)	Accuracy
Satlas_Amazon_epoch30	Sentinel2_SwinB_SI_RGB	Amazon/Atlantic forest	499/20	30	2	B02, B03, B04	0.916	0.956	0.956
Satlas_Atlantic_epoch30	Sentinel2_SwinB_SI_RGB	Amazon/Atlantic forest	485/20	30	2	B02, B03, B04	0.648	0.818	0.836
Aitlas_Amazon	DeepLabV3	Amazon Rainforest	30/15	40	2	B02, B03, B04	0.846	0.917	0.914
Satlas_RGB1_epoch70	Sentinel2_SwinB_SI_RGB	satlas-pretrain	5656/3702	40	12	B02, B03, B04	0.271	0.679	0.953
Satlas_RGB2_epoch100	Sentinel2_SwinB_SI_RGB	satlas-pretrain	5656/3702	150	12	B02, B03, B04	0.265	0.681	0.95
Satlas_MS_tci-b08_epoch150	Sentinel2_SwinB_SI_MS	satlas-pretrain	5656/3702	70	12	B02, B03, B04, B08	0.335	0.74	0.959
Satlas_MS_tci-b08-b11-b12_epoch40	Sentinel2_SwinB_SI_MS	satlas-pretrain	5656/3702	100	12	B02, B03, B04, B08, B11, B12	0.32	0.741	0.96

Explanation of Columns:

• Model Name: The name of the model trained and evaluated.
• Dataset Link: A link to the dataset used for training and testing the model.
• Train/Test Dataset Size: The number of samples used for training and testing, respectively.
• Epochs: The number of training epochs.
• Number of Labels: The total number of classes or labels in the segmentation task.
• Sentinel-2 Bands: The Sentinel-2 spectral bands used for training (e.g., B02 = Blue, B03 = Green, B04 = Red, etc.).
• Mean IoU: The mean Intersection over Union score, a common metric for evaluating segmentation models.
• F1 Score (Weighted): The weighted F1 score that considers the imbalance of class sizes.
• Accuracy: The overall classification accuracy of the model.

More information on the model details can be found in model_config.json.

DPR Team, 2025

Made as part of Arnes Hackathon 2025.