Detecting small changes in tropical forests from space: experiments using synthetic aperture radar

Humid tropical forests are globally important for biodiversity and carbon sequestration. However, vast areas are being degraded, and so made vulnerable to deforestation, drought and fire, through logging that currently goes undetected by satellite monitoring. Nations and companies have pledged to meet climate and conservation targets through reductions in degradation, but methods of quantifying disturbances down to the scale of individual tree losses are required to verify progress towards these goals. Synthetic aperture radar (SAR) satellites are suited to this task as they can provide high resolution (1-10 m) images over large areas, are not affected by cloud cover, and are sensitive to vegetation structure. In this thesis, I test the ability of SAR data to quantify logging disturbances in terms of canopy area loss, aboveground biomass (AGB) loss, and canopy height changes. Reference data were obtained through the Tropical Forest Degradation Experiment (FODEX). At study sites in Peru and Gabon undergoing selective logging, tree diameters, wood volume from terrestrial laser scanning (TLS), and forest structure through unmanned aerial vehicle (UAV) LiDAR were measured at multiple time points over a period of four years. My aim was to improve SAR-based methods of detecting tropical forest disturbance. To do so, I analysed the effects of SAR imaging parameters (including wavelength, incidence angle and resolution), trialled different approaches to processing SAR data (such as the level of multilooking or length of time series used), and developed new change-detection algorithms. I focused on two sources of SAR data in particular: i) C-band SAR (wavelength of around 6 cm) from Sentinel-1 - a European Space Agency mission designed to provide consistent repeat imagery with an approximate resolution over land of 20 m; ii) and X-band SAR (wavelength around 3 cm) from TanDEM-X - a German Aerospace Center mission designed to measure elevation using twin satellites orbiting in formation, while providing experimental products for research at resolutions from 1-3 m. I present my work in three results chapters about (i) detection of canopy gaps through drops in Sentinel-1 intensity (Chapter 2); (ii) using TanDEM-X interferometry to detect biomass change in areas with steep slopes (Chapter 3); and (iii) modeling small canopy changes using both TanDEM-X, Sentinel-1, and new high resolution X-band sensors (Chapter 4). In Chapter 2, I develop an algorithm to detect sharp drops in C-band intensity in VV and VH polarisations from Sentinel-1. These SAR 'shadows’ were compared to canopy gaps delineated using UAV LiDAR spanning 1-year at the Gabon study site, and I show that disturbances as small as 0.02 ha were detected. Overall, 87% of canopy gaps larger than 0.01 ha were detected, with a 0.3% false alarm rate. I found a linear relationship (R2=0.74) between the area of Sentinel-1 shadow and the area of canopy gaps per hectare, which I applied to Sentinel-1 data across the country of Gabon to produce a national scale map of canopy cover loss in 2020. My map suggests that the total gross canopy cover loss was around an order of magnitude higher (1.3% of forested area) than the changes detected by previously published forest loss products based on optical or SAR imagery. In Chapter 3, I propose a method for monitoring degradation using X-band interferometric phase height, a variable that is closely related to mean canopy height and therefore expected to decrease where trees are removed. Previous research in flat terrain had not addressed how to deal with steep slopes, which are present in many remaining tropical forests. Here, I use eight TanDEM-X acquisitions over Gabon (four from ascending passes, four from descending passes) to show that topographic artefacts can be mitigated by selecting data from different pass directions on a pixel-by-pixel basis, determined by incidence angle and coherence. In addition, I demonstrate that minimising multilooking strengthens the relationship between phase height change and AGB change across four 1-ha plots. Finally, in Chapter 4, I compare three TanDEM-X variables (phase height, coherence, and intensity) from high-resolution spotlight images and Sentinel-1 intensity to canopy height changes over a two year period in the Peruvian study site. Models of canopy height change using each SAR variable were trained and tested on UAV LiDAR data. The strongest model used X-band intensity, followed by phase height, suggesting that high resolution X-band data is sensitive to even smaller disturbances than those detected by Sentinel-1. No SAR variable, however, showed sensitivity to the small amounts (1-2 m height increase) of forest growth present in the study area. In addition, commercial X-band imagery from ICEYE and Capella are presented, and I argue that these data provide new avenues for change detection in tropical forests at the finest scales. My results show that short wavelength SAR can quantify forest losses at the scale of individual canopy trees, most reliably by detecting decreases in backscatter along a dense time series. While X-band provides higher resolution, Sentinel-1 is the most promising tool for wide area mapping due to its routine acquisition program (normally 12 day repeats) and open data policy. Current global forest loss products do not capture the finest disturbances, meaning that logging and other forest dynamics are going undetected, potentially undermining efforts to protect biodiversity and slow climate change. Sentinel-1 mapping of tropical forest degradation should therefore be an urgent priority.