A high-resolution dataset for forest disturbance mapping

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 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/Dee4Dist-ResU-net-34) holds the pretrained model weights. Acknowledgement: