[Paper Review] High-resolution canopy height map in the Landes forest (France) based on GEDI, Sentinel-1, and Sentinel-2 data with a deep learning approach

Martin A. Schwartz, Philippe Ciais+12 more|arXiv (Cornell University)|Dec 20, 2022

Remote Sensing and LiDAR ApplicationsEnvironmental Science35 references30 citations

TL;DR

The paper develops a deep learning U-Net model using Sentinel-1/2 data to produce a 10 m canopy height map for the Landes de Gascogne forest (France) with GEDI-derived height as reference, achieving MAE of 2.02 m.

ABSTRACT

In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10 - 20 meters) is arguably needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-stream remote sensing measurements to create a high-resolution canopy height map over the "Landes de Gascogne" forest in France, a large maritime pine plantation of 13,000 km$^2$ with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-net models based on a combination of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each instrument in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole "Landes de Gascogne" forest area for 2020 with a mean absolute error of 2.02 m on the Test dataset. The best predictions were obtained using all available satellite layers from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region.

Motivation & Objective

• Capture fine-scale canopy height heterogeneity in European managed forests where stands are small and heterogeneous within stands.
• Create a high-resolution (10 m) canopy height map for Landes de Gascogne using multi-sensor Sentinel data.
• leverage GEDI-derived height as reference to train/deploy a deep learning model across a large area.
• Evaluate multiple U-Net configurations to assess the contribution of Sentinel-1 and Sentinel-2 inputs.

Proposed method

• Train seven U-Net models on multi-band Sentinel-1/2 inputs with composite time-averaged features.
• Use GEDI-derived tree height as the target for supervised learning.
• Evaluate band configurations to determine the importance of each instrument.
• Generate a 10 m resolution canopy height map for the 2020 time step for the Landes de Gascogne area.
• Validate predictions with external forest inventory plots and a Skysat-based stereo 3D reconstruction model at specific locations.
• Compare performance to existing canopy height models in the same region.

Experimental results

Research questions

• RQ1 Can a deep learning model using multi-band Sentinel-1/2 data accurately predict GEDI-derived tree height across a large managed forest.
• RQ2 How do different Sentinel-1/2 band combinations influence canopy height prediction accuracy?
• RQ3 Is a 10 m canopy height map for Landes de Gascogne achievable with acceptable error (MAE) across validation datasets?
• RQ4 How does the proposed model compare with previous canopy height models in the region?

Key findings

• Achieved a 10 m resolution canopy height map for 2020 over Landes de Gascogne with a mean absolute error of 2.02 m on the test dataset.
• Best predictions were obtained using all available Sentinel-1 and Sentinel-2 layers, though using a single satellite source also yielded good results.
• Outperformed previous canopy height models for coniferous forests in the region on the validation datasets.
• Utilized external validation data from forest inventory plots and Skysat-based 3D reconstructions for evaluation.