Leveraging Sentinel-2 Multi-Spectral Images and GEDI Observations to Estimate Above Ground Biomass in Costa Rica

Tropical forests represent the world's largest live carbon sink and therefore play a critical role in sequestering carbon that would otherwise contribute to climate change. Estimating and monitoring the amount and dynamics of carbon stored in tropical forests greatly adds to our understanding the behavior and stability of this critical carbon sink. To develop continuous, accurate and high resolution estimates of above ground biomass (AGB) in tropical forests we leverage and fuse remote sensing observations of structural complexity (Global Ecosystem Dynamics Investigation project (GEDI)) and canopy surface reflectance properties (Sentinel-2). GEDI uses space-borne lidar to measure forest structure and predict AGB between 51.6°N and 51.6°S latitudes. However, GEDI is expected to be active on the International Space Station for a relatively short period of time (about three and a half years). We seek to extract the relationship between vegetation structural properties and canopy spectral properties using two rich data sources to enable the seasonally varying estimates of AGB using multi-spectral images from Sentinel-2, which are expected to have continued availability beyond the operational span of GEDI. We focus our study on tropical vegetation in Costa Rica. We collected data for ten high resolution (10m or 20m) bands from Sentinel-2 at monthly time frequency between July 2019 to October 2021. Reflectance data were filtered for noise/clouds and gap filled temporally using polynomial regression and spatially using a convolutional neural network. GEDI Level 4A footprint level AGB measurements and GEDI Level 2A footprint level observations containing canopy height profiles, were also collected within the same time frame. The level 4A data were used to train an ensemble of regression models: random forest, extreme gradient boosting and a convolutional long short-term memory (LSTM) neural network to estimate AGB using Sentinel-2 reflectances. All models were hyperparameter tuned using 4-fold spatial cross-validation with root mean squared error as the scoring metric. The results show that decision-tree based modeling techniques appear ineffective, while LSTM is capable of producing more accurate AGB estimates from multi-spectral images.