Combining LiDAR and Sentinel-2 Data to Improve Leaf Area Index Assessment in Forest and Refine Understory Microclimate Models

Refined modeling of the links between microclimate and canopy properties is needed to identify more resilient forest management practices addressing challenges raised by climate change. These practices should foster the temperature-buffering role of forest canopies, with a positive impact on multiple ecosystem services linked to biodiversity, biogeochemical cycles, and recreational services, among others.

Leaf area index (LAI) is an important input variable for state-of-the-art microclimate models. However, accurately measuring this biophysical variable in the field is challenging due to both technical and logistical difficulties. Relying on Earth observation data and processing techniques currently available appears as a promising solution to extend LAI assessment over space and time.

Standalone methods can be used to assess LAI from data acquired with different types of remote sensing sensors with advantages and limitations identified for each sensor in an operational perspective. Airborne Light Detection And Ranging (LiDAR) technology captures detailed information about forest structures at a very high spatial resolution. However, the high cost of acquisitions limits its use for monitoring LAI dynamics over extended areas. Multispectral satellite imagery provides information on vegetation properties over extended areas, with frequent revisit and fine spatial resolution. However, the signal is known to saturate with high LAI values, which prevents accurate assessment of the LAI in dense canopies. Multi-sensor approaches have the potential to help reduce the uncertainty associated with estimates of the canopy properties and alleviate the limitations of individual sensors.

Intending to improve input data for microclimate models, we developed a framework to better understand the differences between LAI estimated from either Sentinel-2 multispectral imagery or LiDAR data and further introduced a method to improve LAI assessment based on the combination of both data types. The study site is the French National Forest of Mormal, a lowland broadleaved forest located in northern France.

First, forest Plant Area Density (PAD) profiles were derived from data from a single leaf-on airborne LiDAR survey over the forest. From the profiles, Plant Area Indexes (PAI) corresponding to vegetation layers of different depths are assessed. Then we parameterized physical model inversion based on the PROSAIL model to assess LAI from Sentinel-2 canopy reflectance and optimize the correlation with LiDAR-derived PAI. Multiple strategies are currently explored to optimize this parameterization. The several sets of PAI are compared and the potential sources of discrepancies (e.g., height, cover heterogeneity) are analyzed.

In the next step, a deep learning model fusing LiDAR and Sentinel-2 data (including reflectance and higher-level products) will be developed. A domain-specific network architecture will be implemented for each data source, followed by the fusion of each network to assess the LAI. Results will be validated using digital hemispherical photographs (DHPs).

This study is part of the ANR MaCCMic project, which aims to develop new tools to help managers increase the resilience of forests and promote the ecological, recreational, and climatic services they offer.