Towards Generalizable Mapping of Hedges and Linear Woody Features from Earth Observation Data: a national Product for Germany

Technical Summary: Towards Generalizable Mapping of Hedges and Linear Woody Features from Earth Observation Data

Problem Statement
Linear woody features (LWFs), such as hedgerows, tree lines, and riparian vegetation, are critical for biodiversity, climate adaptation, and soil protection. However, systematic, large-scale mapping of these features using Earth Observation (EO) data remains a methodological challenge. Existing approaches are often highly specialized, relying on narrow combinations of specific sensors, spatial resolutions, and acquisition conditions (e.g., leaf-on vs. leaf-off). Consequently, these workflows lack transferability; applying a model trained on one dataset to a different region or sensor type typically results in significant performance degradation or requires extensive retraining. Furthermore, the diversity of available EO data in Germany—from high-resolution orthophotos to canopy height maps—lacks a unified processing pipeline that can handle this heterogeneity while maintaining a single, reusable segmentation model.

Methodology
The authors propose a modular workflow designed to decouple the generation of a binary woody vegetation mask from the morphological separation of linear features. The workflow consists of two independently optimizable components:

1. Heterogeneous Input Data Interface: This module consolidates diverse EO data sources into a unified intermediate product: a binary woody vegetation mask. The study utilizes three distinct input sources for Germany:

   • BKG Data: A combination of 20 cm digital orthophotos (DOP20), 1 m Digital Surface Models (DSM), 1 m Digital Terrain Models (DTM), and building footprints. Due to the mosaic nature of the DOP20 (varying acquisition dates, leaf-on/off conditions, and contrast enhancements), a conditional processing logic is employed. This includes generating a normalized DSM (nDSM), creating a building mask, and applying tile-specific NDVI thresholding based on acquisition metadata to derive the vegetation mask.
   • Canopy Height Map (CHMv2): A global product derived from Maxar imagery (approx. 0.73 m resolution in Germany), processed via a simple height threshold.
   • Liu et al. (2023) Canopy Height Map: A 3 m resolution product derived from PlanetScope imagery, also processed via height thresholding.

2. Deep Neural Network for Morphological Separation: A deep learning model, based on a modified U-Net architecture, is trained to separate linear features from non-linear (patchy) features within the binary vegetation masks.

   • Training Strategy: Crucially, the model is trained exclusively on procedurally generated synthetic data. This avoids reliance on site-specific real-world labels, enhancing spatial transferability. The synthetic data generator creates 1,024 × 1,024 pixel scenes with randomized combinations of background, linear features (angular and organic shapes), and patches (large, medium, tiny).
   • Architecture & Input: The model accepts a three-channel input: the binary vegetation mask, its skeleton, and a distance transform. It utilizes two output heads: one for three-class segmentation (background, linear, non-linear) and a secondary head for skeleton prediction restricted to linear features.
   • Loss Function: The training objective combines weighted cross-entropy, a Dice loss focused on the underrepresented linear class, and a skeleton-based loss to enforce topological continuity.
   • Inference: The model processes the national-scale masks in overlapping chips. Post-processing includes skeleton-based refinement (removing pixels far from the predicted skeleton) and the exclusion of known non-vegetated linear structures (e.g., bridges, solar parks) using vector data.

Key Contributions

• A Conceptual Workflow: A unified pipeline that integrates heterogeneous input data (orthophotos, DSM/DTM, canopy height maps) into a single intermediate binary mask, enabling the use of a single trained model across different data configurations.
• Synthetic Training Methodology: A training approach based entirely on synthetic data generation, allowing for a spatially independent model that does not require retraining when applied to new regions or sensor types.
• National-Scale Products: The derivation of three national-scale linear woody feature maps for Germany at spatial resolutions of 0.73 m, 1 m, and 3 m, demonstrating the workflow's flexibility.
• Evaluation Framework: A comprehensive evaluation using independent reference data from four federal states (Bavaria, Brandenburg, Baden-Württemberg, North Rhine-Westphalia) and a novel skeleton-based metric (AUC of F1 scores over varying distance thresholds) to assess topological agreement, mitigating the sensitivity of pixel-wise metrics to resolution differences.

Results
The study evaluated the three derived maps against reference data and two existing datasets (Huber-García et al., 2025; Muro et al., 2025).

• Performance: The workflow produced competitive results across all evaluation sites. The models derived from the BKG data (1 m) and CHMv2 (0.73 m) generally outperformed the 3 m resolution models (Liu et al. and Muro et al.) in terms of topological correctness.
• Resolution Impact: While higher-resolution inputs yielded more detailed outputs, the skeleton-based evaluation showed that the performance gap between resolutions was less pronounced when focusing on topological alignment rather than strict pixel overlap.
• Generalizability: A single model trained on synthetic data successfully processed three distinct input sources without retraining, validating the workflow's transferability.
• Limitations: The 1 m resolution of the BKG-derived mask (constrained by the DSM) led to the exclusion of low-growing linear features, resulting in a performance gap compared to the 20 cm resolution study by Huber-García et al. Additionally, tiles acquired under leaf-off conditions required a simplified processing branch, occasionally introducing non-vegetated structures (e.g., bridges) that required post-processing removal.

Significance
The paper argues that the proposed workflow addresses the critical bottleneck of spatial transferability in linear woody feature mapping. By abstracting the input data into a binary mask and treating the extraction task as a generalized morphological separation problem solved by a synthetic-data-trained model, the approach eliminates the need for region-specific retraining. This modularity allows domain experts to optimize the vegetation mask generation based on available local data, while deep learning specialists can independently improve the segmentation model. The authors posit that this design provides a foundation for scalable mapping beyond Germany, potentially extending to European or global scales using existing global canopy height products, provided that standardized benchmark datasets are established for future validation.