Open-Canopy: A Country-Scale Benchmark for Canopy Height Estimation at Very High Resolution

🌈 Abstract

The article introduces Open-Canopy, an open-access and country-scale benchmark for very high resolution (1.5 m) canopy height estimation. It covers more than 87,000 km2 across France, combining SPOT satellite imagery with high resolution aerial LiDAR data. The authors also propose Open-Canopy-∆, the first benchmark for canopy height change detection between two images taken at different years. The article evaluates a comprehensive list of state-of-the-art computer vision models for these tasks.

🙋 Q&A

[01] Open-Canopy Dataset

1. What are the key characteristics of the Open-Canopy dataset?

• The Open-Canopy dataset covers more than 87,000 km2 across France, combining SPOT 6-7 satellite imagery at 1.5 m resolution with high resolution aerial LiDAR data.
• It is divided into training, validation, and test sets, with a 1 km buffer between the test set and other sets to avoid data contamination.
• The dataset includes SPOT images, ALS-derived canopy height maps, and ALS-derived classification rasters.
• It provides a total of 42,455,312,381 annotations, with one height value per 1.5 m pixel.

2. Why was France chosen as the geographic focus for this benchmark?

• France offers valuable open-source data through recent national initiatives, including DINAMIS for SPOT 6-7 VHR satellite imagery and LiDAR-HD for airborne 3D point clouds.
• France exhibits a wide range of climates and forest types, making it a representative test bed for canopy height estimation.

3. How was the vegetation mask constructed for the dataset?

• The vegetation mask was created by taking the union of an ALS-derived mask for vegetation over 1.5 m in height and the official forest plot outlines provided by IGN.
• This covers a wide range of vegetation types, including trees, shrubs, and urban trees, beyond just forest areas.

[02] Canopy Height Estimation Models

1. What types of computer vision models were evaluated for canopy height estimation?

• The authors evaluated a range of state-of-the-art models, including convolutional networks like UNet and DeepLabv3, Vision Transformers (ViT) and their hierarchical variants (HVIT, SWIN, PCPVT, PVTv2).
• They also explored the impact of pretraining the models on different datasets, including ImageNet, DINO, CLIP, and satellite imagery.

2. What were the key findings from the model evaluations?

• Contrary to trends in natural image analysis, convolutional approaches like UNet and HVIT outperformed standalone ViT models.
• However, hierarchical ViT architectures like SWIN, PCPVT, and PVTv2 achieved the highest precision, highlighting the importance of multi-scale structure for this task.
• Pretraining on ImageNet performed better than pretraining on large-scale vision-language models like CLIP, suggesting the importance of task-specific features.

3. How did the performance of the Open-Canopy models compare to existing canopy height products?

• The models trained on the Open-Canopy dataset significantly outperformed existing canopy height maps, which were typically derived from lower resolution imagery like Landsat or Sentinel.
• The Open-Canopy UNet and PVTv2 models achieved a Mean Absolute Error (MAE) of 2.67 m and 2.52 m, respectively, compared to MAEs ranging from 4.83 m to 9.22 m for other products.

[03] Canopy Height Change Detection

1. What is the goal of the Open-Canopy-∆ benchmark?

• Open-Canopy-∆ is the first open-access benchmark for detecting significant reductions in canopy height between two consecutive VHR satellite images (2022 and 2023).
• The ground truth for this task was derived from two ALS acquisitions covering a 166 km2 area in the Chantilly forest.

2. How was the ground truth for canopy height change detection created?

• The ground truth change mask was generated by subtracting the 2022 ALS-derived height map from the 2023 map, and then applying thresholds to identify areas with significant (>10 m) and contiguous (>100 m2) decreases in canopy height.

3. How did the performance of the PVTv2 model compare to other approaches for change detection?

• The PVTv2 model trained on the Open-Canopy dataset achieved an Intersection over Union (IoU) of 22.5% for the binary change mask, outperforming the Sentinel-based approach from Schwartz et al. (8.0% IoU).
• However, canopy height change detection remains a challenging task, with further improvements needed to precisely identify all areas of significant disturbance.