PhD defence: Jaime Caballer Revenga

Jaime Caballer Revenga defends his thesis:

UAV-LiDAR and machine learning methods for above-ground biomass and carbon stock prediction in forests and croplands

Supervisor:
Professor Thomas Friborg
Assistant Professor Katerina Trepekli, IGN
Assistant Professor Stefan Oehmcke, DIKU
Professor Christian Igel, DIKU
Fabian Gieseke, WW-Universität Münster – Germany

Assessment committee:
Professor Signe Normand, Dept. of Biology, Aarhus University
Professor Ioannis N. Athanasiadis, Wageningen University - The Netherlands
Associate Professor Guy Schurgers (chair), IGN

Abstract:
Croplands and forests are key agents in the global carbon (C) cycle and their response to a warming climate is largely uncertain as different land management practices and plant functional types (PFT) may result in diverging feedback mechanisms to the local impacts of climate change (CC). Moreover, accurate quantification of C stocks and of turnover values across ecosystems and scales remains challenging beyond well-studied ecosystems that require ground-based instrumentation and long-term flux
monitoring, and can only provide point-based estimates, therefore missing the two-dimensional (2D) spatial variability.
In conventionally managed croplands, a short growing season determines the yearly C uptake so that the seasonal net primary productivity (NPP) quickly translates to built-up biomass within ca. 90-120 days. The photoassimilated C is then allocated to physiological and structural functions within the plant, therefore becoming the build up of biomass a decent proxy for the ecosystem sink efficiency. Being aware of it, here, we aimed at measuring the biomass contained in the standing part of cereal crops and, concurrently, surveying the three-dimensional (3D) structure of the ecosystem (Paper I). To that end, we employed an unmanned aerial vehicle (UAV) carrying a light detection and ranging (LiDAR) sensor. We then examined to what extent the total plant-mediated C accounts for the seasonal NPP at the ecosystem scale, where the preliminary findings revealed favorable levels of agreement between the LiDAR-based method and the eddy-covariance technique (Paper III). 
In forests, the growing season spans a longer time (April-November), but the yearly NPP does not generally translate into structural growth that is measurable by discrete-beam LiDAR sensors. Therefore, when investigating forests (Paper II), our approach focused on (i) understanding which community processes mediate individual tree growth and (ii) testing whether a machine learning framework could quantify how (and how much) forest structure partially conditions individual tree growth. To test this, we conducted supervised regression experiments to predict individual tree biomass, in which we introduced information about the spatial context of the individual tree subject as a distinguishing factor between pairs of regression experiments. The results showed that context-aware regressions consistently outperformed context-unaware regressions by up to 9.1% of root-mean-squared error (RMSE). Remarkably, we encoded context using entirely dataset-native methods—i.e., without using ancillary data streams or expert-based assessments—so that the methods developed can potentially be applied to other regions and scales.
As our methods draw on near-field remote sensing (RS) techniques, we have focused solely on the above-ground component of biomass (i.e., AGB) and have not considered the below-ground aspects. Therefore, developing supervised regression methods for AGB prediction has been our main objective. Fortunately, in the case of conventional croplands, the below-ground biomass (BGB) is largely determined by management practices. However, for forests, the systematic observation and prediction of BGB is beyond our scope due to its inherent complexity. 
At the national scale (Paper IV), the availability of large datasets allowed us to test Deep Learning (DL) regression based on 3D convolutions, using airborne laser scanning(ALS) point cloud data (PCD) scenes, to predict forest biomass across the Danish territory without the need for intermediate data representations or feature engineering procedures. Results showed that prediction accuracy in test sets outperformed state-of-the-art methods without overfitting the models, thereby indicating a significant methodological shift in AGB mapping initiatives at large scales. 
In conclusion, the results and the methods we developed illustrate how above-ground carbon (AGC) can be accurately estimated from mobile platforms and LiDAR sensors by leveraging machine and deep learning regression methods at both the ecosystem and regional scales, thereby facilitating applications to more accurate C cycling assessments, precision agriculture and quantitative ecology.

A digital version of the PhD thesis can be obtained from the PhD secretary phd@ign.ku.dk