Earth Systems Lab 2024 - 3D SAR for Forest Biomass

Project Background

Synthetic Aperture Radar (SAR) is a remote sensing technology that uses radar signals to create detailed images of an area. Unlike optical sensors, which rely on visible light and are often obstructed by clouds or forest canopies, SAR operates in the microwave spectrum, allowing it to penetrate through dense vegetation, smoke, and atmospheric conditions. This capability makes SAR especially useful for monitoring forest environments where traditional optical systems struggle to capture accurate data beneath the canopy. SAR systems measure the backscatter of radar signals, which varies based on surface structure, moisture, and vegetation, providing valuable information for assessing forest biophysical parameters such as tree height and biomass.

The ESA Biomass satellite (illustrated below), launched in April 2025, has a 12m aperture wire-mesh reflector and will orbit Earth in a Sun-synchronous dawn-dusk orbit at an altitude of 666km. It is the cornerstone of an ambitious mission designed to globally measure forest biomass, using the first-ever space-borne P-band SAR sensor. This operates at a low frequency, enabling it to penetrate even deeper into forest canopies compared to other radar frequencies.

Project Results

For the heading performance comparison experiment, the orientation of the electromagnetic waves transmitted and received by the emitting device is a key metric. There are a number of options - HH Horizontal-Horizontal, where both transmitted and received waves are horizontally polarised; HV Horizontal-Vertical, where the transmitted wave is horizontal, but the received wave is vertical; and VV Vertical-Vertical, where both transmitted and received waves are vertically polarised. For the SE heading, models trained on data filtered to exclude areas below 5 meters in height showed significant improvements in MAE with increasing numbers of SLC inputs. Specifically, increasing from three to seven SLC inputs reduced the test MAE by an average of 8%, with further improvements observed when using the full 28-image stack. The vertical-vertical (VV) polarisation channel demonstrated superior performance across all SLC input scenarios, as illustrated opposite.

Reduction in MAE from 3 to 7 SLC images for VV was 17%, and this increased to 28% when 28 MLC images were used. Although this may have been expected given VV sensitivity to vertical structures such as tree trunks and echoes, it proved the VV data was best suited to the tomographic canopy height forest reconstruction and LiDAR analysis illustrated opposite.

When it came to tree classification, despite achieving an overall classification accuracy of 72%, the team's model exhibited a notable disparity between the performance of the majority class and the minority classes. The dominant class, Aspen, which constituted 64.3% of the samples, achieved a high F1-score (a measure of predictive performance) of 0.84. However, minority classes, such as Oak and Beech forests, showed much lower F1-scores, often below 0.30, suggesting the model may be overfitting to the majority class and struggling to learn from the underrepresented classes. This is graphically represented in the diagram opposite.

When compared to prior studies utilising full-stack tomographic SAR data, the ESL team's approach using SLC image stacks showed a competitive performance, albeit with a higher MAE in scenarios involving fewer input images.