A STUDY TO DERIVE HIGH RESOLUTION L-VOD INFORMATION BY EXPLOITING SMOS LEVEL 2 AND SATELLITE LIDAR DATA

C. Vittucci¹, P. Richaume², Y. Kerr²

¹Tor Vergata University of Rome, ²Center for the Study of the Biosphere from Space (CESBIO)

In addition to continuous soil moisture monitoring, passive microwave L-band missions have introduced a valuable new tool for vegetation studies: the retrieval of L-band Vegetation Optical Depth (L-VOD). The latter is particularly useful for forest ecosystems monitoring because it’s sensitive to both tree height and biomass. Previous research has demonstrated that L-VOD values increase with forest height and biomass, reaching saturation at significantly higher biomass levels compared to measurements from active microwave sensors [1]. This unique characteristic makes L-VOD a robust variable for monitoring dense forests.

The retrieval of L-VOD in the Soil Moisture and Ocean Salinity (SMOS) is based on an iterative approach, aiming at minimizing a cost function whose main component is the sum of the squared weighted differences between measured and modeled TB data, for a collection of observation angles. In this approach the retrieval accuracy can benefit from a proper initialization of the parameter [2], which is based on maximum Leaf Area Index (LAI) and LAI seasonal changes [3].

A primary research objective is to enhance the understanding of the L-VOD and specific plant characteristics. This improved understanding is critical for more accurately quantifying changes in vegetation water content and for advancing the electromagnetic modeling of this parameter.

A significant challenge, however, is the relatively coarse spatial resolution of L-VOD products, which is a direct consequence of the radiometer’s footprint. The spatial resolutions for the SMAP (Soil Moisture Active Passive) and SMOS missions are approximately 36 km and 45 km, respectively. This low resolution poses a considerable limitation for validating L-VOD data against high-resolution ground-based measurements, which are essential for refining the accuracy and, consequently, improve the modelling, of both L-band and higher-frequency VOD products.

The availability of LiDAR missions like GEDI and ICESat-2 offers a new way to improve L-VOD retrieval. Recent research from Vittucci et al. [4, 5] shows a strong correlation between SMOS L-VOD data and LiDAR-derived plant variables like top-of-canopy height (RH100) and Plant Area Index (PAI). When averaged to match SMOS’s resolution, this correlation is very high for tropical forests (R>0.88) and remains relevant at boreal latitudes (R>0.7 in summer and R≥0.62 during winter).

Such interesting results encouraged the investigation about the study of a novel methodology to derive L-VOD map at 1 km of spatial resolution, by exploiting the small footprint size of satellite lidar instruments.

The primary objective of this study is to leverage RH100 and PAI to estimate L-VOD across different seasons and forest types. Following the methodology established in [4,5], LiDAR variables from GEDI and ICESat-2 were associated with SMOS L-VOD data on the ISEA 4H9 grid at a 12-km spatial resolution. These prior investigations identified linear relationships useful for initializing the L-VOD parameter. The present work applies the coefficients of these relationships to high-resolution LiDAR observations to generate high-resolution L-VOD maps. This procedure required addressing two primary challenges. First, due to the sampling nature of satellite LiDAR sensors, RH100 and PAI products are provided as discrete tracks rather than complete spatial maps. To derive continuous, dynamic L-VOD maps, a gap-filling procedure was required to address these spatial voids. The second challenge was the absence of PAI data for latitudes above 52°N, as this is beyond the GEDI mission’s coverage area, yet it is where most boreal forests are located.

Both issues were resolved by implementing a random forest (RF)-based gap-filling procedure, which can be applied recursively over time and at a continental scale. The RF model’s training data are computed through an iterative process that pairs RH100 and PAI samples with multispectral indices from MODIS products, including NDVI, EVI, and FPAR. Additionally, IGBP land cover classes are included as a supplementary input. The training sets are designed to be representative of the LiDAR parameter dynamics over a 10-day interval for a given geographic region.

For the missing PAI information at high latitudes, the same approach is applied by considering data acquired from the 45°N to 52°N range as training information for the RF. The resulting gap-filled maps of RH100 and PAI have a spatial resolution of 500 m, consistent with native MODIS products. These maps are subsequently downsampled to 1 km, a resolution closer to the ISEA 4H9 grid, before being used as input for the linear relationship equations to compute the final 1-km resolution L-VOD maps. The results of the proposed procedure will be shown during the conference.

Reference

1. Vittucci, C., et al. “SMOS retrieval over forests: Exploitation of optical depth and tests of soil moisture estimates.” Remote sensing of environment 180 (2016): 115-127.

2. Kerr, Yann H., et al. “The SMOS soil moisture retrieval algorithm.” IEEE transactions on geoscience and remote sensing 50.5 (2012): 1384-1403.

3. Rahmoune, Rachid, et al. “SMOS level 2 retrieval algorithm over forests: Description and generation of global maps.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 6.3 (2013): 1430-1439.

4. Vittucci, Cristina, et al. “SMOS L-VOD retrieved by level 2 algorithm and its correlation with GEDI LIDAR products.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14 (2021): 11870-11878.

5. Vittucci, Cristina, Leila Guerriero, and Paolo Ferrazzoli. “Influence of vegetation height, plant area index, and forest intactness on SMOS L-VOD, for different seasons and latitude ranges.” IEEE Transactions on Geoscience and Remote Sensing 61 (2023): 1-11.