In a joint cooperative between ESA´s Global Development Assistance (GDA) Climate Resilience project and a WorldBank (WB) funded proof-of-concept for a modern MRV2.0 (Monitoring, Reporting and Validation), the partners aim to improve the accuracies and reduce the efforts required for countries carrying out their GHG (Greenhouse Gas) National Inventory Reports. Multiple countries were selected as test cases, one such case is Mozambique, with an area of interest in the Zambezia province.
With Above-Ground Biomass (AGB) being a integral input for the calculations required as part of the MRV process, this initiative aims at assessing and improving wall-to-wall AGB mapping. This assessing component is to identify the accuracies and uncertainties using a current ‘state-of-the-art‘ methodology.
As with any biophysical parameter we want to estimate from space, this basically entails using ground-truth data, in this case provided by the National Forest Inventory, and using the relationship (if any) with EO data to make predictions over a larger area. For AGB, this is inherently difficult due to the saturation effect of optical sensors at relatively low values. This is due to the parameter being volumetric in nature and mostly below the surface of the canopy. Optical sensors are, at this stage, still the most effective in capturing the variations in AGB. One way to combat the saturation problem is to include longer wavelength sensors (i.e., Radar) that can penetrate the canopy and capture structural variability, thus not being as prone to saturation effects. To this end, the longer the wavelength, the further it can penetrate the vegetation, and currently this is JAXA's L-band ALOS/PALSAR-2 satellite (ESA's P-band BIOMASS satellite is set to launch early next year).
With a machine learning algorithm using annual percentiles of Sentinel-2 data and a mean composite ALOS/PALSAR-2 data, we were able to model the forest biomass of the Zambezia province at 10m resolution. This first version (Mark I) is a reflection of the current capabilities of purely EO-based solutions. The forests in this region proved to be difficult to map due to the spatial variability of Miombo woodland communities.
There are a wealth of published AGB products freely available, and these are most often at a global scale and therefore have relatively coarse resolutions and high uncertainties at local or even national scales. It is best practice, therefore, to calibrate maps that will be used for MRV using local ‘ground-truth’ data. When comparing the Mark I product to a selection of such global products, we can see the accuracy improvement, not to mention higher spatial resolution.
The project now continues with Mark II, the improvement component. As concluded from an investigation into state-of-the-art and evolving technologies by the WorldBank, this is to (among others) incorporate LiDAR into a multi-step multi-scale approach. More precisely, this aims to 1) improve the ‘ground-truth’ measurements using Terrestrial Laser Scanning (TLS) and, 2) provide an intermediate step of extrapolating AGB using Aerial Laser Scanning (ALS) to a sub-region before using these drastically improved estimates (in accuracy, spatial resolution and amount of data sampling) to map the whole province using satellite EO inputs.
The extent and continuous format of this intermediate LiDAR dataset is to further improve upon methods by implementing a U-net regression algorithm which will take into account the spatial variability of the forest.