Pan-tropical modelling of land cover and land-use change trajectories for newly deforested areas

Abstract

Despite extensive efforts in modelling and monitoring forest cover and forest change, few advances in the study of land cover (LC) and land use (LU) dynamics following deforestation (so-called post-loss LC/LU change) exist. The current plethora of multi-source datasets with considerable high-frequency records and emerging technologies (e.g. cloud-computing) to process them, provide new possibilities in the study of long-term (>2 years) LC/LU dynamics over deforested areas. Building on these advances, this research aims to develop models for mapping and analysing post-loss LC/LU dynamics pantropically, based on time series analysis of earth observation and contextual data, for use in environmental and agricultural studies. The first part of this research conducts a state-of-art review which explores a rich body of theory, definitions and global policies relevant to the study of post-loss LC/LU change. Furthermore, a description of existing studies showing methodological advances at a large scale, from national to global, provide guidance for potential directions of work. As the majority of these advances rely on georeferenced data, the next step of this research investigates multiple data sources, from field records to coarse satellite imagery of over 120 sites detected as disturbed by Terra-i (a near real-time monitoring system for LC conversion). This local-scale but data-rich investigation helps to highlight opportunities and challenges in the study of long-term post-loss LC/LU change. By integrating the lessons learned above, the second part of the research deploys an end-to-end supervised deep learning architecture. This particular architecture is capable of ingesting and processing large volumes of MODIS satellite image time series in the generation of 19-years of LC predictions suited to the study of post-loss LC change trajectories. The scalability of the proposed model is evaluated through its implementation at macro-regional scale, i.e. the Amazon region, using big data principles and cloud-computing technologies. The implementation has brought some challenges, such as assessing the effectiveness of the model for different LC classification schemes derived from global LC maps. The results show the calibrated model outperforms conventional machine learning techniques used by other on-going initiatives mapping long-term post-loss LC/LU change. In addition, the feasibility of transferring the calibrated model to predict LC in other tropical areas in Latin America, Asia and Africa is successfully demonstrated. The final part of this thesis describes how post-loss LC change data generated for the Amazon can be characterised and grouped into different typologies. In doing so, this research introduces the minimum number of years for identifying these typologies and how they are linked to potential land-use types and impacts on carbon dynamics. The findings corroborate the need to go beyond simplistic assumptions, such as a full recovery or complete conversion to non-forest after disturbance, to enhance estimates of the impacts of deforestation. In addition, this research offers illustrative examples of the influence of 12 spatial layers representing inaccessibility, biophysical and policy factors on the observed post-loss LC change trajectories. It has been found some of these variables, such as proximity to farmland, proximity to urban/built-up, proximity to protected areas with sustainable use, and elevation, can explain human-related post-loss LC change. This PhD project is innovative in its use and deployment of novel methods (e.g. deep learning and sequence analysis) and technologies (e.g. cloud-computing) to analyse large volumes of high-frequency earth observation data and chronological post-loss LC change records. Furthermore, this research supports modellers of ecosystem services, biodiversity, and other deforestation-relevant topics by going beyond the immediate state of deforestation to understanding the typologies and dynamics of long-term post-loss LC/LU change trajectories, which will have impacts on these areas of research.

Date of Award	1 Sept 2020
Original language	English
Awarding Institution	King's College London
Supervisor	Mark Mulligan (Supervisor) & Thomas Smith (Supervisor)