Convolutional Neural Networks Regression Model with Uncertainty Estimates to predict GEDI Canopy height at 30m resolution using multisource SAR and optical observations

Accurate estimates of forest aboveground biomass (AGB) are essential for assessing forest carbon stocks and their change over time to support policies for climate change mitigation, resource management, and biodiversity conservation. Among the methodologies to estimate AGB, those that include accurate forest canopy height (CH) information present better estimates due to the direct relationship between AGB and CH. Therefore, to cover large areas, Light Detection and Ranging (LiDAR) remote sensing technology is preferred because it can provide highly accurate and precise measurements of the distance from the ground to the top of the canopy. However, developing continuous acquisition campaigns using LiDAR technology at continental scales at high-spatial resolution is too expensive.  

The Global Ecosystem Dynamics Investigation (GEDI) offers a unique opportunity to overcome this challenge. The GEDI mission uses a laser instrument mounted on the International Space Station (ISS) to measure the distance from the ISS to the Earth’s surface with high accuracy and spatial resolution. However, GEDI does not provide a spatially continuous CH map. Instead, it captures 25 m spatial resolution footprint samples over the Earth’s surface following a sparse-grid-based sampling pattern between 51.6° N and 51.6° S. In this acquisition setup, the samples are spaced every 60 m in the along-track direction and 600 m in the across-track direction.  

To estimate CH for areas not covered by the sparse GEDI mission, we propose a non-linear mapping function using Convolutional Neural Networks with Uncertainty estimates (UCNNs) with input data from other satellites and output a continuous estimate of CH with a measure of uncertainty. Specifically, we use coregistered multitemporal data from Sentinel-1, Sentinel-2, and ALOS PALSAR. From Sentinel imagery, we use bimonthly composites each year from April-May, June-July, and August-September to capture the dynamics of the spectral and structural tree information in Canada. From ALOS PALSAR, we use the one-year composite, and from GEDI data, we use strong-beam samples from June to July from the corresponding year, while excluding all low-quality samples. Experiments were conducted for 2020 in the Province of Ontario, Canada, whose climate is considered continental, with temperatures ranging from humid in the south, with cold winters and warm summers, to sub-Arctic in the north. To avoid overfitting, we apply spatial cross-validation splitting the study region into five non-overlapping areas. The cumulative uncertainty histogram shows that 90% of samples present an uncertainty of CH less than 5 meters. These results are the first step towards spatially continuous mapping of canopy height using multitemporal and multisource satellite data, with implications for improving assessment of forest biomass estimation and carbon monitoring from space.