Conference Agenda

Forest height is one of the most important forest structure parameters due to its close relation with biomass, terrestrial carbon storage and ecosystem dynamics. Among the forest height observation methods such as passive optical, lidar and manual in-situ measurement, InSAR has unique advantages, including its height sensitivity as well as all-weather, day-and-night, and wall-to-wall mapping capability, which makes it a powerful tool for large-scale forest height observation. As the majority of spaceborne InSAR coverage, single-baseline single-polarization InSAR data was used in this research due to its data accessibility.

The accuracy of physical scattering model-based forest height inversion with single-baseline single-polarization InSAR is restricted by the limited number of solvable unknown parameters, especially in lower frequency like L-band where the ground-associated contributions become significant. In this paper, we propose a method that exploits a backscatter model with the help of GEDI and ATLAS/ICESat-2 spaceborne lidar height products as auxiliary information to reduce the number of unknowns.

In particular, the Random Volume over Ground (RVoG) model was derived by incorporating ground-related contributions (both ground-surface backscattering and double-bounce) into the the Random Volume (RV) model, which describes the relation between InSAR complex coherence and randomly oriented and distributed scatterers within a volume. This process introduces two additional unknowns rendering the forest height inversion problem underdetermined. To address this problem, the backscatter RVoG model is adopted in this work. The two above-mentioned parameters are estimated as global constants via backscatter regression, utilizing Lutan-1 backscatter and spaceborne lidar forest height at each lidar footprint. Global parameters estimated in this step are used in the pixelwise wall-to-wall mapping from Lutan-1 InSAR complex coherence to forest height through using the interferometric RVoG model.

Lutan-1 bistatic data with HH polarization at Hainan Tropical Rainforest National Park was selected to validate this method for complicated tropical forest. Standard SAR and InSAR processing were performed on the Lutan-1 data to obtain backscatter power, corrected coherence amplitude and interferometric phase. The underlying digital terrain model (DTM) was obtained by using the few-look InSAR phase height histogram method developed in our previous work. The physical scattering model-based forest height inversion method described above was then applied on a pixel-by-pixel basis to estimate forest height.

Validation against ALS lidar measurements demonstrates the proposed method's strength. It effectively corrects the significant global underestimation seen with the RV model, achieving a negligible bias (0.39 m) and a RMSE of 5.65 m in contrast to the -2.48 m bias and 5.76 m RMSE of RV model. Performance, however, varies by vegetation height, which was analyzed in three distinct height ranges. The model shows high linearity and accuracy for medium-height trees (15–30 m), as the global parameters are optimized around this average. However, it tends to overestimate short vegetation (<15 m). This is attributed to the model's increased sensitivity to decorrelation and potential contamination from non-forest targets like crops, where the applied forest SAR/InSAR scattering model is invalid. For taller trees (≥30 m), the method underestimates height. This is identified as an intrinsic property of the L-band's deep penetration capability when observing the sparse upper canopies, resulting in a loss of sensitivity of the complex coherence to height. As a result, tall trees are indistinguishable from medium ones.