Mitchard, Ed

Ryan, Casey

Bollasina, Massimo

Dong, Wenquan

2024-11-13T15:23:51Z

2024-11-13T15:23:51Z

2024-11-13

Accurate quantification of forest aboveground biomass (AGB) and its changes is crucial for understanding the role of forests in the global carbon cycle and developing effective climate change mitigation strategies. This thesis presents a comprehensive assessment of forest AGB in China using a combination of field data, remote sensing observations, and machine learning algorithms. The research is divided into four chapters, each addressing specific aspects of forest AGB estimation and mapping. Chapter 2 used established methods, but novel focused datasets, to create the first high resolution map of China’s biomass for the mid-2000s. I generated a wall-to-wall AGB map of China for the year 2007 at a 50 m resolution using Ice, Cloud, and land Elevation Satellite (ICESAT) Geoscience Laser Altimeter System (GLAS) Lorey’s height data, L-band SAR from Advanced Land Observing Satellite (ALOS) Phased Array L-band SAR (PALSAR), C-band SAR from Environmental Satellite (Envisat) Advanced Synthetic Aperture Radar (ASAR), and optical satellite data from Landsat-5 and field data. I averaged GLAS data within 0.01 degrees × 0.01 degrees grid cells, obtaining 8,981 cells with at least two GLAS footprints. Lorey’s height was converted to AGB using allometric equations developed independently for northern and southern China using field data. Random forest (RF) regression was then employed to extrapolate AGB grid cells from GLAS data to a continuous map at 50 m resolution, using variables derived from Earth observation datasets and layers for training. Our estimates of total carbon stored in the forest in China was 9.52 Pg C, with an average forest AGB of 104 Mg ha⁻¹. Chapter 3 looked in detail at two regions where I conducted fieldwork, to assess more accurately how well satellite LiDAR from the GEDI sensor can estimate tree height and biomass, and how well these isolated footprints can be spatially extrapolated using other datasets. It involved generating 25 m resolution AGB maps for two 500 km x 500 km regions in northeastern and southwestern China for the year 2021, using GEDI data, field measurements, and Sentinel-1, ALOS-2 PALSAR-2, and Sentinel-2 data. We measured 26 plots (24 under GEDI footprints) in the northeastern region and 16 plots (12 under GEDI footprints) in the southwestern region. In the northeastern region, the closest relationship was observed between field AGB and RH98, while in the mountainous southwestern region, field AGB exhibited a stronger correlation with RH80. The fitted relationships were used to convert RH98 and RH80 to AGB for both regions, respectively. We found many GEDI footprints had errors not detected by their quality flags; filtering using remote sensing data to remove low-quality footprints improved results. In both the northeastern and southwestern regions, an inverse correlation between slope steepness and model accuracy was observed. Specifically, in the northeastern region, coefficient of determination (R²) values exhibited a decline from 0.93 in areas of minimal slope to 0.42 in locales exhibiting slopes greater than 30 degrees. Concurrently, the root mean square error (RMSE) values escalated from 14 to 38 Mg ha⁻¹. Similarly, in the southwestern region, R² values decreased from 0.75 in relatively flat terrains to 0.50 in areas with slopes exceeding 30 degrees, alongside an increase in RMSE values from 20 to 44 Mg ha⁻¹. In Chapter 4 I examine whether the latest deep learning methods can improve biomass estimation, beyond methods using relatively simple tree-based methods (Random Forest) as used in chapters 2 and 3. In particular I explore the use of the attention UNet (AU) deep learning model for estimating forest AGB in Guangdong Province. I converted GEDI relative height (RH) metrics to AGB using the allometric equations established in Chapter 3 and then employed attention UNet to extrapolate GEDI footprints to a wall-to-wall AGB map. The AU model demonstrated superior performance in biomass estimation accuracy compared to the traditional RF method, though computational requirements were considerable. I managed to produce a 2019 10-meter resolution AGB map for the whole of Guangdong using the AU model. This involved a novel approach to reduce boundary artifacts in patchbased predictions, which are prevalent in deep learning applications for image analysis. By overlapping patches and excluding edge pixels, the method improved spatial consistency and accuracy at the edges of predictions, leading to more dependable AGB estimates, without patch artefacts. Although AGB distributions from both AU and RF models were closely aligned, with similar mean values, the AU-derived AGB map offered more realistic spatial details upon visual assessment. Chapter 5 employs a multi-step approach to estimate forest AGB in China for 2021 and analyze its changes since 2007. It employs the approaches explored in Chapter 3, at the same scale as Chapter 2, and without using the advanced ML methods of Chapter 4 as the computational burden was not considered. First, GEDI L2A RH metrics were converted to AGB estimates using region-specific allometric equations developed in Chapter 2. Next, random forest models were trained for each first-level administrative unit in China to predict AGB using GEDI-derived AGB estimates and predictor variables from Sentinel-1, PALSAR-2, and Sentinel-2. The trained models were then applied to generate 25 m resolution wall-towall AGB maps for each first-level unit and then merged into a national-scale AGB map. Our results show that the total carbon stored in the forest in China for 2021 was 13.06 Pg C, with a mean AGB density of 108.75 Mg ha⁻¹. Comparing the 2021 AGB map with the 2007 map, we observed an overall increase in carbon storage, with a net gain of 3.54 Pg C and varying spatial patterns of AGB changes across the country. Spatial variations in AGB alterations were observed across the nation, exhibiting the most significant increments in the northeastern and south-central regions of China. The high-resolution AGB maps and quantified biomass changes generated in this thesis provide valuable insights into the spatial distribution and dynamics of forest biomass resources in China, supporting decision-making in forest management and climate change mitigation efforts.

https://hdl.handle.net/1842/42646

http://dx.doi.org/10.7488/era/5340

en

The University of Edinburgh

Chen, M., Dong, W., Yu, H., Woodhouse, I., Ryan, C. M., Liu, H., Georgiou, S. and Mitchard, E. T. (2023), ‘Multimodal deep learning for mapping forest dominant height by fusing gedi with earth observation data’, arXiv preprint arXiv:2311.11777

Dong, W., Mitchard, E. T. A., Santoro, M., Chen, M. and Wheeler, C. E. (2023), ‘2007 forest aboveground biomass map for china’. URL: https://doi.org/10.7488/ds/7480

Dong, W., Mitchard, E. T., Santoro, M., Chen, M. and Wheeler, C. E. (2024), ‘A new circa 2007 biomass map for china differs significantly from existing maps’, Scientific Data 11(1), 287

Dong, W., Mitchard, E. T., Yu, H., Hancock, S. and Ryan, C. M. (2023), ‘Forest aboveground biomass estimation using gedi and earth observation data through attention-based deep learning’, arXiv preprint arXiv:2311.03067

forest aboveground biomass

China

global carbon cycle

forests

Ice, Cloud, and land Elevation Satellite

Geoscience Laser Altimeter System

Advanced Land Observing Satellite

Environmental Satellite

Advanced Synthetic Aperture Radar

Random forest

Mapping forest aboveground biomass and its change in China

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy