Mapping land use change in the Congo Basin: Machine learning approaches to land cover monitoring, projecting forest cover change under IPCC climate scenarios, and modelling large mammal habitat suitability
When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
Machine learning (ML) models have been developed as powerful computational tools for monitoring, mapping and quantifying land use and land cover (LULC) and its change over time. ML Models such as k-nearest neighbor (kNN), support vector machines (SVM), artificial neural networks (ANN), and random forests (RF) have been used effectively to classify LULC types at a range of geographical scales. However, ML models have not been widely applied and compared in African tropical regions, especially in the Congo Basin where LULCC mapping most often face methodological challenges that arise from relying on coarse-resolution satellite images, the only freely available option. Those studies that do exist have generally relied on applying only a single method, which in effect, can increase classification uncertainties relative to the use of multiple ML classifiers. There is therefore, a need to apply and compare the classification performance of several ML algorithms, so as to improve large scale and long-term land cover change mapping in the Congo Basin, as well as in other afro tropical forest areas underrepresented in the remote sensing literature. The Congo Basin is a global hotspot for forest fragmentation and natural resource degradation. The region loses approximately 1 million hectares (Mha) of forest per year as a result of socioeconomic, demographic and climate-driven disturbances. This forest loss contributes to LULCC problems, with effects expected to worsen in the future as human population grows and with increased global warming. Wildlife species, particularly endangered primates and large mammals are losing the suitable habitats required for their survival, and consultations from the International Union for the Conservation of Nature (IUCN) and the United Nations Educational, Scientific and Cultural Organization (UNESCO) suggest that species are likely to face extinction within the Dzanga Sangha Protected Areas (DSPA) of the Central African Republic (CAR).
These key issues are addressed in this dissertation research under three manuscript chapter-based topics. In Chapter 2, I applied statistical and machine learning approaches using four classification algorithms (ANN, kNN, SVM, and RF) to effectively map and quantify LULCC within a tropical rainforest region in Central Africa (the Mayo Rey department of northern Cameroon), thereby providing a first attempt in comparing these algorithms within such a setting. I found that all four classification algorithms produced relatively high accuracy (overall classification accuracy > 80%), with the Random Forest (RF) model (> 90% classification accuracy) outperforming the kNN, SVM, and ANN models. These results thus suggest that the RF model is likely the best ML classifier for large scale LULC mapping within Afro-tropical forest settings.
In Chapter 3, I used the RF model to map decadal changes in LULC patterns for the Congo Basin for the period 1990 – 2020, and further project these changes to the year 2050, under various scenarios of climate change and socioeconomic impacts. I found that dense forest cover decreased by over 5.2 percentage points (pp) (215938 km2), 1.2 pp (50046 km2), and 2.1 pp (86658 km2) in the Congo Basin between 1990 – 2000, 2000 – 2010, and 2010 – 2020 respectively, accounting for approximately 8.5 pp (352642 km2) loss in dense forest cover between 1990 - 2020. For the period 2020 – 2050, we project a 3.7-4 pp (174859.6 - 204161 km2) loss in dense forest cover under all three climate change scenarios, suggesting that approximately 12.3 pp (556803 km2) of forest cover will be lost in this region over a 60-year period (1990-2050). I provide evidence that forest cover loss in this region is highly influenced by logging and forest clearing, wildland fires, human population density, distance to built-up areas, and temperature increases. I also show that the Congo Basin is experiencing large-scale expansion of croplands and built-up areas, with an anticipated two-fold expansion projected by the year 2050 relative to the period from 1990-2020, leading to continued loss in forest cover in this region. These results fill a critical knowledge gap about the current and future states of land cover in the Congo Basin resulting from the impacts of human activities and climate change.
In Chapter 4, I produced spatially explicit species distribution models to map the spatial variability and changes in ape and elephant habitat suitability within the DSPA between two survey years (2015 and 2020), using spatial datasets of eco-guard patrol activities, habitat fragmentation, land cover, human pressure, topography, and climatic variability as model predictors. I identified priority habitats, as well as key factors affecting species distribution. I found that priority chimpanzee habitats covered about 1383 km2 (30 %) of the entire DSPA in the year 2015, while priority gorilla and elephant habitats covered approximately 2569 km2 (56 %) and 3075 km2 (67 %) respectively. Priority habitat area for the three species declined by 4, 4.5 and 9.8 percentage points respectively between 2015 and 2020, mostly due to increased human pressures. I further provide evidence that the DSPA National Parks represent regions of higher priority habitat for all three species owing to the reduced human pressure that has resulted from higher eco-guard patrol efforts. The findings from this chapter contribute to our understanding of great ape conservation biology, and provide tools and data to help secure their future within the DSPA.