Deep Learning For Forest Disturbance mapping (Deep4Dist)

The Deep4Dist dataset is a comprehensive, high-resolution remote sensing data product specifically designed for forest disturbance mapping. It comprises approximately 17,500 georeferenced image patches extracted from high-resolution digital orthophotos acquired in Rhineland-Palatinate, Germany. Each image patch measures 500 × 500 pixels at a spatial resolution of 20 cm and contains five spectral channels: red, green, blue, near-infrared (NIR), and a normalized digital surface model (nDSM). Together, these channels capture both spectral and structural information essential for distinguishing various forest disturbance types, including bark beetle damage, clear-cuts, and windthrow events. Key Features: High Resolution: 20 cm spatial resolution enables fine-grained mapping of forest disturbances. Multiple Disturbance Classes: Bark beetle damage, clear-cut and windthrow. Multispectral & Structural Data: Five channels (RGB, NIR, and nDSM) provide detailed spectral and structural insights. Large-Scale Coverage: ~17,500 georeferenced image patches support robust statistical analysis and deep learning applications. Rigorous Curation: Data were generated from high-resolution digital orthophotos and ground disturbance records. Extensive quality control, including expert-based external validation. Deep Learning Ready: The dataset is organized and annotated for direct use in semantic segmentation tasks. Train (~70%), validation (~25%) and test (~5%) splits are provided. Applications: Deep4Dist is ideally suited for: Developing and validating deep learning models for forest disturbance mapping. Investigating the spatial dynamics of forest disturbances. Supporting adaptive forest management and conservation strategies. Integrating with medium-resolution satellite data for multi-modal forest disturbance mapping. Class Description: The classes in the segmentation masks are encoded as integers ranging from 0 to 3, corresponding to: 0: Background 1: Bark beetle damage 2: Clear-cuts 3: Windthrow Metadata Description: The metadata.csv file contains the following fields and information: tile_name: corresponding to the image/mask name split: the assigned data partition set (train, validation, test) x_center: the x coordinate of the tile centroid (EPSG:25832) y_center: the y coordinate of the tile centroid (EPSG:25832) acquisition_date: aerial image acquisition/flight date disturbance_agent: disturbance agent/agents affecting the tile. Spatial Data Description: The tile_geometries.gpkg is a vector file containing the geometries (polygons) for each image sample (EPSG:25832). Coordinate Reference System: Both, images and masks are georeferenced using the EPSG: 25832 (DE_ETRS89_UTM32) crs. Folder Structure: Each folder-set (train, validation and test), contains the subfolders "image" and "mask", where the 5-channels aerial images and dense pixel labels are stored. Additional Resources: Code : The GitHub repository (https://github.com/enmanuelrodpau/deep4dist) contains code for model training and dataset validation. Model weights: The HuggingFace repository (https://huggingface.co/enmanuelrp/Deep4Dist-ResU-net-34) holds the pretrained model weights. Acknowledgement: We gratefully acknowledge the Hunsrück-Hochwald National Park and Landesforsten Rheinland-Pfalz for providing reference data, as well as the Landesamt für Vermessung und Geobasisinformation Rheinland-Pfalz for granting free access to digital orthophotos. We thank the Allianz für Hochleistungsrechnen Rheinland-Pfalz for providing access to high-performance computing resources. We also express our appreciation to Asli Ozdarici Ok, Levent Yorulmaz, Patrick Christen, Sharad Kumar Gupta, Sonila Papathimiu, and Tomasz Wojciechowski for their kind participation in the external evaluation of the Deep4Dist dataset. Version history: 1.0.0 - Initial version. 1.0.1 - Metadata file updated. Added a new column with per tile disturbance agents.