Technical Summary: Hedgementation = Hedgerow Segmentation: A Remote Sensing Benchmark

Problem Statement

Hedgerows—small patches of natural habitat in agricultural landscapes—are critical for biodiversity, carbon storage, and ecosystem services. However, their extent is declining, and existing inventory data is geographically and temporally limited, often restricted to specific regions like France or England and spanning only a few years. Collecting large-scale data manually is costly. While satellite imagery combined with deep learning offers a potential solution, progress is hindered by sparse labels, poor model transferability to new regions, and a lack of reproducible benchmarks. The specific challenge lies in detecting fine-scale, linear features (hedgerows) that vary significantly across different climatic and agricultural zones.

Methodology

The Hedgementation Benchmark

The authors propose Hedgementation, a new benchmark for evaluating machine learning models on hedgerow mapping at a country scale (France) with a 10m² spatial resolution. The benchmark integrates three primary data sources:

1. Imagery: Harmonized, surface-reflectance Sentinel-2 data (10m resolution, 5-day intervals) from September 2021 to October 2022, accessed via Google Earth Engine.
2. Embeddings: Annualized embeddings from the Alpha Earth Foundations (AEF) model, providing learned representations at 10m resolution.
3. Ground Truth: Labels from the French BD Haie inventory (Version 2), produced by the IGN institute. These labels are derived from aerial photo interpretation, validated by farmers in agricultural zones, and supplemented by automatic segmentation elsewhere.

Data Processing:

• Rasterization: Since BD Haie labels are line geometries without width, the authors rasterize them by applying a 7m buffer (based on average German hedgerow width and French government accounting standards), rasterizing at 0.5m resolution, and downsampling to 10m via averaging.
• Patch Generation: The national data is split into 1.28km x 1.28km square patches (128x128 pixels).
• Stratification: Patches are assigned to climatic zones (Temperate vs. Subtropical) based on Global Agro-Ecological Zones data and categorized by agricultural status using French agricultural parcel boundaries.

Experimental Design

The benchmark evaluates model generalization across two dimensions:

1. Spatial Generalization: France is divided into large tiles separated by a 28km buffer. Patches are assigned to "Near" test sets (potentially adjacent to training data) and "Far" test sets (at least 28km away from any training patch) to measure transferability across distance.
2. Climatic Generalization: Models are tested on temperate and subtropical zones, including cross-zone training scenarios (e.g., training only on temperate data and testing on subtropical data).

Baselines Evaluated:

• Deep Learning Models:
  • PASTIS U-TAE: A U-Net with a Temporal Attention Encoder, trained from scratch on raw satellite imagery.
  • FTW (Fields of The Worlds): A model using fewer spectral bands and two seasonal input images.
• Embedding-Based Models:
  • k-Nearest Neighbors (kNN): Applied to AEF embeddings with 3x3 spatial context and center weighting.
  • Random Forest: Applied to AEF embeddings.
  • Logistic Regression: Applied to AEF embeddings.

Evaluation Metric:
Performance is measured using Intersection over Union (IoU), a standard metric for pixel-wise classification that is invariant to scale and sparsity.

Key Results

Baseline Performance

• Deep Learning: The PASTIS U-TAE model achieved the highest performance, with an average IoU of 41.0% on near test sets and 38.9% on far test sets. The FTW model performed lower, with an average IoU of roughly 31% (near) and 29% (far).
• Embedding Models: Models using AEF embeddings (Random Forest, kNN, Logistic Regression) showed lower absolute accuracy (IoUs ranging from ~15% to ~24%) but demonstrated robust spatial generalization. Notably, the Random Forest model showed almost identical performance on near and far test sets, suggesting the AEF embeddings capture global features effectively.

Generalization Analysis

• Spatial Distance: As expected, deep learning models (PASTIS, FTW) performed slightly better on "Near" patches than "Far" patches (a gap of ~2-5% IoU). However, embedding-based models showed negligible degradation over distance.
• Climatic Zones: Subtropical patches were consistently more difficult to predict than temperate patches. Models trained exclusively on subtropical data performed poorly on temperate data, and vice versa, though training on the full dataset yielded the best overall generalization.
• Agricultural vs. Non-Agricultural: Prediction was significantly easier in agricultural landscapes. The PASTIS U-TAE model achieved an IoU of 45.9% on agricultural pixels compared to 41% on the full dataset. This is attributed to higher label quality in agricultural zones (farmer validation) and the prevalence of hedgerows in these areas.

Significance and Claims

The paper positions Hedgementation as a necessary step toward reproducible and comparable research in remote sensing for ecological monitoring. Its primary contributions are:

1. A New Benchmark: It provides a standardized dataset and evaluation protocol for hedgerow segmentation at a national scale, addressing the lack of accessible, high-quality benchmarks for fine-scale agricultural features.
2. Generalization Focus: Unlike many benchmarks that focus solely on accuracy within a single region, Hedgementation explicitly tests spatial and climatic generalization, highlighting the challenges of transferring models across diverse agro-ecological zones.
3. Reproducibility: The authors provide the code to automatically create the benchmark and replicate baseline results, aiming to overcome the "black box" nature of previous remote sensing studies.
4. Modest Baselines: The authors acknowledge that while the best baseline (PASTIS U-TAE) reaches >40% IoU, this is insufficient for precise agricultural and ecological analysis. The benchmark serves as a foundation to drive further progress in model architecture, pre-training strategies, and data efficiency.

The paper concludes that while significant improvements are needed, the current version of Hedgementation successfully informs model development by quantifying the difficulties of spatial and agricultural generalization in hedgerow mapping.