BRIX-2 stands for Second Biomass Retrieval Inter-comparison eXercise and represents a joint effort between ESA and NASA to intercompare algorithms specifically for biomass mapping using current and future spaceborne missions. The exercise aims at using Synthetic Aperture Radar (SAR) at P-Band and L-band, and LIDAR datasets acquired as part of the ESA and NASA AfriSAR joint-campaign in support of the upcoming ESA’s BIOMASS [1] mission, of the upcoming NASA-ISRO SAR (NISAR) mission [2], and of the current NASA Global Ecosystem Dynamics Investigation (GEDI) mission [3].

The objectives of BRIX-2 are:
1. Provide an objective, standardized comparison and assessment of biomass retrieval algorithms developed for the BIOMASS, NISAR and GEDI missions, and fusion of these mission datasets.
2. Establish a forum to involve scientists in the development of retrievals that have so far not been part of the Biomass community.
3. The adoption of vetted validation standards and methods to compare biomass estimates to reference datasets (e.g. field plots or airborne lidar biomass maps).
4. Collect inputs from the biomass user and scientific community on data formats and characteristics towards the generation of Analysis Ready Data.

These objectives shall be achieved by making available standardized test cases (based on airborne campaign and spaceborne simulated data), inviting the scientific community to develop and apply retrieval algorithms based on this test case, and finally compare and evaluate the performance of submitted results [4].

For the purpose of an objective algorithm evaluation, the exercise was based on ESA-NASA Multi-Mission Algorithm and Analysis Platform (MAAP) [5]. This analysis platform is a virtual, open and collaborative environment for the processing, analysis and sharing of data and development and sharing of algorithms. The MAAP provides a common platform with computing capabilities co-located with data as well as a set of tools and algorithms developed to support this specific field of research.

Participants were invited to upload their code with a mandatory permissive open-source license to the MAAP (or develop it on the MAAP) and run it on the MAAP using the predefined campaign datasets.

The first results of the this inter-comparison exercise will be presented.


REFERENCES
[1] T. Le Toan, S. Quegan, M. Davidson, H. Balzter, P. Paillou, K. Papathanassiou, S. Plummer, F. Rocca, S. Saatchi, H. Shugart and L. Ulander, “The BIOMASS Mission: Mapping global forest biomass to better understand the terrestrial carbon cycle”, Remote Sensing of Environment, Vol. 115, No. 11, pp. 2850-2860, June 2011.

[2] P.A. Rosen, S. Hensley, S. Shaffer, L. Veilleux, M. Chakraborty, T. Misra, R. Bhan, V. Raju Sagi and R. Satish, "The NASA-ISRO SAR mission - An international space partnership for science and societal benefit", IEEE Radar Conference (RadarCon), pp. 1610-1613, 10-15 May 2015.

[3] https://science.nasa.gov/missions/gedi

[4] « Biomass Retrieval Inter-comparison eXercise #2: BRIX-2 Protocol », version 2.1, 6 August 2021, https://liferay.val.esa-maap.org/documents/portlet_file_entry/35530/BRIX-2+Protocol+V2.1.pdf/1d887f7c-5a15-9725-0ec8-3a50292f0010?download=true]

[5] Albinet, C., Whitehurst, A.S., Jewell, L.A. et al. A Joint ESA-NASA Multi-mission Algorithm and Analysis Platform (MAAP) for Biomass, NISAR, and GEDI. Surv Geophys 40, 1017–1027 (2019). https://doi.org/10.1007/s10712-019-09541-z

Accurate mapping of forest aboveground biomass (AGB) is critical for carbon budget accounting, sustainable forest management as well as for understanding the role of forest ecosystem in the climate change mitigation. In this study, spaceborne Global Ecosystem Dynamics Investigation (GEDI) LiDAR and Sentinel-2 multispectral data were used in combination with elevation and climate data to produce a wall-to-wall AGB map of Australia that is more accurate and with higher spatial and temporal resolution than what is possible with any one data source alone. Specifically, the AGB density map was produced that covers the whole extent of Australia at 200m spatial resolution for the Austral winter (June-August) of 2020. To produce this map Copernicus Sentinel-2 composite, GLO-90 Digital Elevation Model (DEM) and long-term climate variables were trained with samples from the GEDI Level 4A product.
From GEDI Level 4A data available within Australia between June – August 2020, all measurements not meeting the requirements of L4A product quality, and those with degraded state of pointing or positioning information and an estimated relative standard error in GEDI-derived AGB exceeding 50% were rejected. Seasonal Sentinel-2 composite was generated using a Sentinel-2 Global Mosaicking (S2GM) algorithm and was further used to calculate Normalized Difference Spectral Indices (NDSIs) from all spectral bands. Similarly, DEM was used to calculate aspect, roughness, slope, Topographic Position Index and Terrain Ruggedness Index. Finally, climate variables consisted of average precipitation, radiation as well as minimum and maximum temperatures calculated between 1970-2020.
The boosting tree machine learning model was applied to predict wall-to-wall AGB density map. For each 200m × 200m cell the number of available GEDI measurements was calculated and models were built based on average AGB density of cells containing > 5 GEDI measurements. Up to ≈62,000 cells, each 200m × 200m, were used to train predictive machine learning models of AGB density. The predictive performance of models based on both satellite imagery only (single-data source) and a fusion of satellite imagery with elevation and climate data (multi-data source) was compared. Bayesian hyperparameter optimization was used to identify the most accurate Light Gradient Boosting Machine (LightGBM) model using 5-fold cross-validation.
The multi-data source approach had a substantially higher accuracy (coefficient of determination (R2) increase of up to 0.1, root-mean-square error (RMSE) decrease of up to 7 Mg/ha and root-mean-square percentage error (RMSPE) decrease of up to 17%) as compared to the single-data source approach. The single-data source analysis based on only Sentinel-2 imagery resulted in AGB density predicted with the R2 of 0.68-0.75, RMSE of 40-46 Mg/ha and RMSPE of 47-69%. Model performance improved with the addition of DEM and climate information: AGB density prediction with R2 of 0.77-0.81, RMSE of 35-40 Mg/ha and RMSPE of 41-52%. Using a SHapley Additive exPlanations (SHAP) approach to explain the output of LightGBM models it was found that Sentinel-2 derived NDSIs using Red Edge and Short-wave Infrared bands were the most important in predicting seasonal AGB density.
Similar model performance is expected for annual prediction of AGB density at a finer resolution (e.g. 100m) due to higher density of GEDI measurements. This research highlights methodological opportunities for combining GEDI measurements with satellite imagery and other environmental data toward seasonal AGB mapping at the regional scale through data fusion.

TomoSense is an ESA-funded campaign over a temperate forest in support of future SAR mission concepts at P-, L- and C-band. This paper gives the first results from analysing the relationship between P-band tomographic SAR (TomoSAR) data and above-ground biomass (AGB) including ground slope effects. The results are important for the upcoming BIOMASS mission, in particular for temperate forest AGB estimation.

Airborne P-band SAR data over the Kermeter region in Germany was acquired and processed by MetaSensing. TomoSAR processing was performed by Politecnico di Milano by combining 56 SAR flight tracks (28 in each direction – NW and SE). The flights were performed in the mornings on the 22 and 23 July 2020. The SE track acquisitions were affected by RF interference, and only the NW tracks were used for this analysis. An AGB map based on airborne laser scanning (ALS) and 80 in-situ plots, each of a size of 500 m², was provided by CzechGlobe. In addition, a digital elevation model (DEM) over the area provided input for the assessment of topographical effects.

A mix of all forest biotypes (dominant species were beech and spruce, otherwise oak, pine and birch) showed an AGB RMSE of 15 % for VV polarization, 17 % for HV and 29 % for HH relative the AGB map. The results obtained were based on 397 areas of 0.5 ha size, limiting the surface slope angle to be less than 20°, and integrating the TomoSAR intensity from 20 m to 30 m in height. TomoSAR vertical profiles in this height interval were found to have the highest sensitivity to AGB. The corresponding AGB RMSE for integrating the intensity of the vertical profiles over the full height was 30 % in VV, 57 % in HV, whereas HH did not show a clear sensitivity to AGB.

Influence from topography on the AGB sensitivity was observed for increasing surface slope angles. Limiting the surface slope angle to below 20° improved the AGB RMSE from 22 % to 15 % in VV and from 24 % to 17 % in HV, when observing a mix of all forest biotypes. In the ground-range direction, negative slopes (facing away from the radar) decrease the TomoSAR intensity integrated from 20 m to 30 m whereas positive slopes (facing toward the radar) increase it. This effect is likely due to a combination of varying canopy attenuation and incidence angle dependent trunk scattering. A degradation of the AGB sensitivity was also observed for increasing slopes in the azimuth direction due to increased intensity variance per AGB interval.

These results show that TomoSAR, which will be available from BIOMASS, is a promising technique for estimating AGB in temperate mixed forests. Similar observations over the region at L- and C-band are currently being analysed.

BIOMASS [1] is ESA's seventh Earth Explorer, scheduled for launch in 2023. It will collect unprecedented information about forests thanks to the first spaceborne P-band Synthetic Aperture Radar (SAR), and featuring full polarimetry. The 435 MHz carrier frequency and 6 MHz bandwidth allow maximum sensitivity to the woody elements of the trees, while complying with ITU regulations and avoiding excessively strong ionospheric effects. Its acquisition cycle is designed to acquire globally (subject to Space Object Tracking Radars restricitions, excluding North America and Europe) and in a multi-baseline repeat pass interferometric configuration. The repeat cycle is set to 3 days, guaranteeing coherence for interferometric and tomographic processing.
After the initial Commissioning Phase, dedicated to instrument and antenna calibration, the experimental Tomographic Phase will achieve one global coverage in 14/16 months. This will allow mapping forests in 3 dimensions by collecting stacks of 7 acquisitions over 18 days for each location, in ascending and descending configurations. During Tomographic Phase it will be also possible to derive a sub-canopy Digital Terrain Model (DTM) [3]. This is fundamental to reject terrain contribution in interferometric processing [5], as terrain acts as a nuisance and disturbs biomass retrieval. The remainder of the mission is the main Interferometric Phase of about 5 years duration, with global coverage achieved in 7/9 months. In this case stacks are collected in dual baseline configuration over 6 days. Coverage will be built up successively, with the successive tomographic or interferometric stacks adding coverage to adjacent areas. Given the complex orbital pattern achieving global coverage over several months, significant environmental changes will occur that the estimation techniques must handle.
In this contribution we describe the Level-2 processing algorithms to estimate Forest Disturbance (FD), Forest Height (FH) and Above Ground Biomass (AGB) products from BIOMASS data [2]. The Level-2 processor requires phase calibrated stacks generated by the BIOMASS interferometric processor and the DTM estimated in the Tomographic Phase [6]. FD, FH and AGB products are generated at each global coverage during the mission lifetime mapping not only forest characteristics but also changes. The processing implements state-of-the-art polarimetric-interferometric techniques allowing to reject terrain signal and focus on canopy scattering. This is supported by strong evidence that in tropical forests the backscatter from the canopy region 25-35m above the ground is highly correlated with the total AGB, which can be exploited using the full power of tomography or interferometry. The actual Level-2 processor software implementation is also briefly presented, along with preliminary results on airborne campaign data.
In particular, AGB estimation results will be shown on BIOMASS-like acquisitions emulated from tropical forests campaign data. The AGB estimation performance is observed to depend on the AGB range and degrades when ground topography is significant. Good performance is achieved when the AGB interval is large (> 400 t/ha) and the average is in the interval 200–250 t/ha [4]. The algorithm is observed to be capable of achieving a relative RMSD of 20% with respect to in situ data using only few calibration points where reference AGB is available, although retrieval accuracy was observed to depend significantly on the quality of the available calibration points. Efforts are now focused on designing the global AGB estimation scheme for BIOMASS, especially with regards to calibration and validation AGB to be used.

[1] T. Le Toan, S. Quegan, M.W.J. Davidson, H. Balzter, P. Phaillou, K. Papathanassiou, S. Plummer, F. Rocca, S. Saatchi, H. Shugart, L. Ulander, “The BIOMASS mission: Mapping global forest biomass to better understand the terrestrial carbon cycle,” Remote Sensing of Environment, vol. 115, pp. 2850-2860, Jun. 2011
[2] Banda, F.; Giudici, D.; Le Toan, T.; Mariotti d’Alessandro, M.; Papathanassiou, K.; Quegan, S.; Riembauer, G.; Scipal, K.; Soja, M.; Tebaldini, S.; Ulander, L.; Villard, L. The BIOMASS Level 2 Prototype Processor: Design and Experimental Results of Above-Ground Biomass Estimation. Remote Sens. 2020, 12, 985.
[3] Mariotti D’Alessandro, M.; Tebaldini, S. “Digital Terrain Model Retrieval in Tropical Forests Through P-Band SAR Tomography” IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 9, pp. 6774-6781, Sept. 2019. doi: 10.1109/TGRS.2019.2908517
[4] Soja, M., Quegan, S., Mariotti d’Alessandro, M. Banda, F. Scipal, K., Tebaldini, S. Ulander, L.M.H. “Mapping above-ground biomass in tropical forests with ground-cancelled P-band SAR and limited reference data”, Remote Sens. Environ. Volume 253, February 2021, 112153
[5] M. Mariotti d’Alessandro, S. Tebaldini, S. Quegan, M. J. Soja, L. M. H. Ulander and K. Scipal, "Interferometric Ground Cancellation for Above Ground Biomass Estimation," in IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 9, pp. 6410-6419, Sept. 2020
[6] M. Pinheiro at al., “BIOMASS DEM Product Prototype Processor”, EUSAR 2021

Trees and shrubs (hereafter collectively referred to as trees) both inside and outside forests, are the basis for the functioning of tree-dominated ecosystems, and are regularly monitored at country scale via forest inventories. However, traditional inventories and large-scale forest mapping projects are expensive, labour-intensive and time-consuming, resulting in a trade-off between the details recorded, spatial coverage, accuracy, regularity of updates, and reproducibility. Also, forest inventories typically do not account for individual trees outside forests, although these trees play a vital role in sustaining communities through food supply, agricultural support, among other benefits. Moreover, the alarming rate of tree cover loss resulting from different natural and human-induced processes has brought both political and economic motives to attract efforts for landscape restoration especially in Africa. Nevertheless, currently, there is no accurate and regularly updated monitoring platform to track the progress and biophysical impact of such ongoing initiatives. Recent approaches counting trees in satellite images in Africa used very costly commercial images, were limited to isolated trees in savannas excluding small trees, and did not cover other complex and heterogeneous ecosystems such as forests. Here, we make use of novel deep learning techniques and publicly available aerial photographs, and introduce an accurate and rapid method to map the crown size, number of trees inside and outside forests, and corresponding carbon stock, regardless of tree size and ecosystem types in Rwanda. The applied deep learning model follows a UNet architecture and was trained using 67,088 manually labeled tree crowns. We mapped over 200 million individual trees in forests, farmlands, wetlands, grasslands, and urban areas, and found about 67.2% of the mapped trees outside forests. An average tree density of 94.6 and 70.8 trees per ha, and average crown size of 38.7 m2 and 15.2 m2 were mapped inside and outside forests, respectively. In savannas we found 64 trees per ha with an average crown size of 15.6 m2. In farmlands we found 79.6 trees per ha with an average crown size of 16.3 m2. We expect methods and results of this kind to become standard in the near future, enabling tree inventory reports to be of unprecedented accuracy.

The use of airborne laser scanning (ALS) data to estimate and map structure-related forest inventory variables, such as forest aboveground biomass, has strongly increased in the last decades and has even become operational in some countries. The development of new machine learning methods, including deep learning-based approaches, and the fine-tuning and validation of already existing methods for deriving these variables from ALS data require extensive datasets of forest inventory measurements on a single-tree level and corresponding ALS data.
Virtual laser scanning (VLS) is a time- and cost-efficient alternative to acquiring such data in the field. We present a framework for virtual laser scanning that combines forest inventory data with a tree point cloud database and an open-source laser scanning simulation framework.
Synthetic 3D representations of forest scenes are created based on forest stand information that can be derived from a forest inventory on a single-tree level or a forest growth simulator. For each tree in the forest stand, a 3D tree model of matching species, height, and crown diameter is inserted into the synthetic forest scene. Single-tree point clouds extracted from real ALS data are used as tree models. Laser scanning of the synthetic forest scene is simulated using the Heidelberg LiDAR Operations Simulator HELIOS++.
We investigate the performance of the VLS simulations using ALS and forest inventory data collected from six 1-ha plots in temperate forests in Southwest Germany. VLS is performed with the same acquisition settings as in the real ALS campaign. For a comparison of different tree model types, the synthetic forest stands are created using closely matching real tree point clouds and using two types of simplified tree models with cylindrical stems and spheroidal crowns. The simulated ALS point clouds are compared with the real ALS point clouds both qualitatively, based on their visual appearance in cross-sections, and quantitatively, based on the height distribution of the returns and several point cloud metrics. To assess the potential of the synthetic data in an application, they are used as training data for forest biomass models, which are then applied on the real ALS data.
Our validation confirms that the presented workflow can be used to generate synthetic ALS datasets, which are sufficiently realistic for many typical applications. The VLS approach can reproduce the relative height distribution of returns of real ALS data. The comparison of different tree model types reveals that the visual appearance of the synthetic forest scenes is much more realistic for the real tree point clouds than for the simplified tree models. However, the differences between the real tree point clouds and the simplified tree models are less pronounced in the quantitative analysis. Our findings suggest that for the wall-to-wall mapping of forest aboveground biomass, synthetic forest scenes composed of simplified tree models can be as suitable as synthetic forest scenes composed of single tree point clouds to create training datasets.