Author: Kunal Cheekhooree
Cheekhooree, Kunal, 2024 Canopy Height Assessment in South Australian Pinus Radiata Plantations using Sentinel-1: A Comparative Analysis between INSAR and Machine Learning Algorithms, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
In forestry applications, tree heights through their inclusion in specific allometric equations can be used to infer above-ground biomass and volume at both individual tree and plantation scales. However, the traditional field-based methods used, while accurate, are often cost-prohibitive due to extensive and time-consuming fieldwork. Remote sensing technologies, like airborne LIDAR, offer non-destructive alternatives by swiftly surveying large areas through the generation of Digital Surface Models (DSM) and Digital Ground Models (DGM), leading to the creation of a Canopy Height Model (CHM). Despite their efficacy, these surveys remain costly and often infrequent.
This study investigates the effectiveness of Synthetic Aperture Radar (SAR) to determine tree height using LIDAR data as the ground truth, focusing on Mount Panorama and Knot Hill Nature Forest Reserves in South Australia. These areas contain commercially planted Pinus radiata subjected to thinning cycles over their lifetimes (ranging from no thinning - T0, to one thinning - T1, two thinnings - T2, and three thinnings - T3). Initially, SAR VV data from the on-going C-band Sentinel-1 mission is used in interferometry (INSAR) to generate a corrected DSM from which a CHM is generated using the airborne LIDAR DGM. The second approach integrates Sentinel-1 SAR VV, VH, and their derived texture data with the multispectral bands from the Sentinel-2 mission and a DGM from the SRTM mission. These datasets feed into a random forest (RF) machine learning model (MLM), interpolating gaps in spaceborne LIDAR data collected by the GEDI instrument, resulting in another CHM.
Comparing the generated CHMs with the LIDAR ground truth CHM at the 90th height percentile revealed that only under T1 conditions, the INSAR method generated usable results with a 5.6% height underestimate and an RMSE of 4.2m. However, the RF MLM was appropriate under both T1 and T2 conditions, with an overestimated height of 9.3% and an RMSE of 2.3m, and a 4.4% height underestimate and an RMSE of 1.9m, respectively. Given the prevalence of T1 and T2 conditions in the study area, the RF MLM emerges as a potentially more suitable option due to its lower overall RMSE. Under T0 and T3 conditions, both the INSAR and RF MLM methods exhibit significant discrepancies from the ground truth, which can be attributed to the limited area in the study site under these conditions.
However, Sentinel-1 C-band data may not be optimal for INSAR over forests due to its upper canopy penetration and the current 12-day temporal resolution; these factors further limit its accuracy over forested areas which generally already exhibit low SAR coherence. Furthermore, the lack of GEDI input data over certain forested areas may have resulted in interpolation over larger data gaps than expected. Nevertheless, it was found that currently Sentinel-1 SAR data when used in combination with other optical and topographic data in MLMs results in more accurate canopy height predictions than when it is used directly through INSAR.
Further research using X-band SAR data that is better suited for INSAR, with its lower canopy penetration and higher temporal resolution, as well as the combination of SAR data with higher canopy penetration from the upcoming NISAR and BIOMASS missions in the RF MLM, could improve the accuracy of both methods for tree height estimations.
Keywords: SAR, INSAR, CHM, Canopy height, Pinus Radiata, RF, GEE
Subject: Engineering thesis
Thesis type: Masters
Completed: 2024
School: College of Science and Engineering
Supervisor: David Bruce