AnnalisaAppice/SWIFTT-bark_beetle_detection_semantic_segmentation

Bark Beetle Detection - Semantic Segmentation Dataset (SWIFTT Project)

📋 Overview

This collection of datasets is devoleped in fullfilment of the research objectives of SWIFTT project (Satellites for Wilderness Inspection and Forest Threat Tracking), funded by the European Union under Grant Agreement 101082732.

It contains labeled satellite imagery acquired with Sentinel-2 and Sentinel-1, that can be used for developing and evaluating predictive models to detect and map forest damages caused by bark beetles in some spruce European forests.

Research Purpose

⚠️ These datasets have been developed and can be used for research purposes only. They are intended to support scientists and reasearchers in conducting their scientific studies, remote sensing algorithm development, and validation of AI models for automated bark beetle damage detection and monitoring.

Objective

The datasets have been created and made available as annotated and well-structured satellite imagery datasets to train and validate semantic segmentation models capable of automatically identifying forest tree dieback caused by bark beetle infestations.

Geographic and Temporal Coverage

• —🇨🇿 Czech Republic in September 2020 - Bark Beetle Detection & Forest Type Classification
• —🇫🇷 North East France in October 2018 - Bark Beetle Detection
• —🇷🇴 Romania in September 2020 - Bark Beetle Detection

🛰️ Data Creation Methodologies

The satellite datasets have been created by using one of the following methodologies:

1. Best Available Image (Best)

• —Description: It selects the highest-quality entire Sentinel-2 image from a monthly time window based on a cloud index metric. The cloud index is computed using the Scene Classification Level (SCL) algorithm and represents the percentage of cloudy, noisy, and defective pixels. The image with the lowest cloud index per month is selected; if multiple images achieve the same minimum cloud index value, the most recent one is chosen.
• —Methodology: Cloud index-based Best selection as described in (Andresini et al. (2024), DIAMANTE: A data-centric semantic segmentation approach to map tree dieback induced by bark beetle infestations via satellite images. J. Intell. Inf. Syst. 62(6): 1531-1558 (2024), https://doi.org/10.1007/s10844-024-00877-6)
• —Spectral Bands: 13 bands - Sentinel-2 spectral bands resampled at 10m resolution (B1-B9, B11, B12; excluding B10 SWIR Cirrus reserved for atmospheric corrections), SCL, and (optionally) colocated Sentinel-1 bands (VV and VH)
• —Advantages: Superior radiometric quality, minimized cloud cover and defective pixels, no compositing artifacts
• —Available for: An area in Czech Republic (September 2020, Sentinel-1 and Sentinel-2), an area in North East France (October 2018, Sentinel-1 and Sentinel-2)

2. Temporal Median (Median)

• —Description: It computes the median value for each band over a monthly time window, masking cloud cover for the median computation
• —Methodology: Median-based selection as described in (Recchia et al. (2026), ULISSE: Parameter-efficient adaptation of earth vision models to monitor forest disturbance in sentinel-2 time series. Ecol. Informatics 94: 103668 (2026), https://doi.org/10.1016/j.ecoinf.2026.103668)
• —Spectral Bands: 12 bands Sentinel-2 spectral bands resampled at 10m resolution (excluding SCL - Scene Classification Layer and B10)
• —Period: Monthly median aggregation
• —Cloud Masking: Active during processing
• —Advantages: Temporal stability, reduction of artifacts, effective cloud masking avoiding chhoosing a single image with high cloud cover
• —Available for: An area in Czech Republic (September 2019, April-September 2020, September 2019), Romania (April-September 2020)

📊 Dataset Structure

Each dataset is organized as follows:

[COUNTRY]/
├── Bark Beetle/
│   ├── [PROCESSING_MODE]/
│   │   ├── [TEMPORAL_PERIOD]/
│   │   │   ├── sentinel_2/          # Satellite imagery
│   │   │   ├── sentinel_1/          # (optional) Radar data
│   │   │   └── stats/               # (optional) Statistics
│   │   └── masks/
│   │       ├── train/               # Training masks
│   │       └── test/                # Test masks
│   └── Planetary/                   # (optional) Ultra high-resolution imagery
│
└── Forest Type/                     # (optional) Forest type classification
    ├── GEE Images Median/
    │   └── [TEMPORAL_PERIOD]/
    │       └── sentinel_2/
    └── Masks/                       # Forest type masks

📦 Data Formats

• —Satellite Imagery: GeoTIFF (georeferenced)
• —Masks: GeoTIFF (single-band) with values 0 (negative) and 1 (positive)
• —Metadata: Included in TIFF headers (CRS, geotransformation)

🔧 Technical Processing Details

Data Processing and Projection

• —Data Source: Satellite imagery collections are acquired from either Microsoft Planetary Computer or Google Earth Engine (GEE)
• —Uniform Resolution: Satellite imagery collections are resampled at 10m spatial resolution using the nearest-neighbor interpolation
• —Coordinate Reference System: Satellite imagery collections are downloaded and processed in EPSG:3857 (Web Mercator)
• —Sentinel-2 Bands: For Sentinel-2 images, 12 bands (B1-B9, B11, B12) are included in the created the imagery datasets
• —Band B10 Exclusion: The B10 (SWIR Cirrus) band is part of the Sentinel-2 spectral range, but it is excluded from all datasets, as it is reserved for atmospheric corrections and not used in the analysis
• —SCL Band Exclusion: The SCL (Scene Classification Layer) band is excluded if the Median-based methodology is used
• —Radiometric Correction: All Sentinel-2 images use Level 2A surface reflectance products (atmospherically corrected)
• —Radiometric Terrain Correction (RTC): All Sentinel-1 images use Level-1 RTC products

📚 Regional Dataset Details

For in-depth information on each regional dataset, see the region-specific README files:

• —**README - Czech Republic**
• —**README - North East France**
• —**README - Romania**

**List of scientific publications developed in fulfillment of the research objectives of the SWIFTT Project using these datasets**

📄 License

This collection of datasets is released under the CC BY-NC-SA 4.0 license (Creative Commons Attribution Non-Commercial Share-Alike 4.0).

Usage:

✅ Non-commercial use ✅ Scientific research ✅ Education ❌ Commercial use (without authorization)

🤝 Contributions and Feedback

To report issues, suggestions, or contributions to the datasets, please contact the SWIFTT project team at the University of Bari Aldo Moro.

The development of this repository is supported from the SWIFTT Project - Satellites for Wilderness Inspection and Forest Threat Tracking, funded by the European Union under Grant Agreement 101082732