Multidimensional remote sensing based mapping of tropical forests and their dynamics

Tropical forests concentrate a large part of the terrestrial biodiversity, provide important resources, and deliver many ecosystem services such as climate regulation, carbon sequestration, and hence climate change mitigation. While in the current context of anthropogenic pressure these forests are threatened by deforestation, forest degradation and climate change, they also have shown to be, in certain cases, highly resilient and able to recover from disturbances. Quantitative measures of forest resources and insights into their dynamics and functioning are therefore crucial in this context of climate and land use change. Sensors on-board satellites have been collecting a large variety of data about the surface of the earth in a systematic and objective way, making remote sensing a tool that holds tremendous potential for mapping and monitoring the earth. The main aim of this research is to explore the potential of remote sensing for mapping forest attributes and dynamics. Tropical South America, which contains the largest area of tropical forest on the planet, and is therefore of global significance, is the regional focus of the research. Different methods are developed and assessed to: (i) map forest attributes at national scale, (ii) detect forest cover loss, (iii) quantify land use intensity over shifting cultivation landscapes, and (iv) measure spectral recovery and resilience of regrowing forests. Remote sensing data are diverse and multidimensional; a constellation of satellite sensors collects data at various spatial, temporal and spectral resolutions, which can be used to inform on different components of forests and their dynamics. To better map and monitor ecological processes, which are inherently multidimensional, this thesis develops methods that combine multiple data sources, and integrate the spatial, temporal and spectral dimensions contained in remote sensing datasets. This is achieved for instance by assembling time-series to fully exploit the temporal signal contained in the data, or by working with multiple spectral channels as a way to better capture subtle ecological features and processes. After introducing the general objectives of the thesis in Chapter 1, Chapter 2 presents an approach for mapping forest attributes at national scale. In this chapter, 28 coarse resolution remote sensing predictors from diverse sources are used in combination with in-situ data from 220 forest inventory plots to predict nine forest attributes over lowland Bolivia. The attributes include traditional forest inventory variables such as forest structure, floristic properties, and abundance of life forms. Modelling is done using the random forest approach and reasonable prediction potential was found for variables related to floristic properties, while forest attributes relating to structure had a low prediction potential. This methodological development demonstrates the potential of coarse resolution remote sensing for scaling local in-situ ecological measurements to country-wide maps, thus providing information that is highly valuable for biodiversity conservation, resource use planning, and for understanding tropical forest functioning. Chapter 3 presents an approach to detect forest cover loss from remote sensing time-series. While change detection has been the object of many studies, the novel contribution of the present example concerns the capacity to detect change in environments with strong inter-annual variations, such as seasonally dry tropical forests. By combining Landsat with Moderate Resolution Imaging Spectroradiometer (MODIS) time-series in a change detection framework, the approach provides information at 30 m resolution on forest cover loss, while normalizing for the natural variability of the ecosystem that would otherwise be detected as change. The proposed approach of combining two data streams at different spatial resolutions provides the opportunity to distinguish anthropogenic disturbances from natural change in tropical forests. Chapter 4 introduces a new method to quantify land use intensity in swidden agriculture systems, using remote sensing time-series. Land use intensity — a parameter known for influencing forest resilience — is retrieved in this case by applying a temporal segmentation algorithm derived from the econometrics field and capable of identifying shifts in land dynamic regimes, to Landsat time-series. These shifts, or breakpoints, are then classified into the different events of the swidden agriculture cycle, which allows to quantify the number of cultivation cycles that has taken place for a given agricultural field. The method enables the production of objective and spatially continuous information on land use intensity for large areas, hence benefiting the study of spatio-temporal patterns of land use and the resulting forest resilience. The results were validated against an independent dataset of reported cultivation frequency and proved to be a reliable indicator of land use intensity. Chapter 5 further explores the concept of forest resilience. A framework to quantify spectral recovery time of forests that regrow after disturbance is developed, and applied to regrowing forests of the Amazon. Spatial patterns of spectral resilience as well as relations with environmental conditions are explored. Regrowing forests take on average 7.8 years to recover their spectral properties, and large variations in spectral recovery time occur at a local scale. This large local variability suggests that local factors, rather than climate, drive the spectral recovery of tropical forests. While spectral recovery times do not directly correspond to the time required for complete recovery of the biomass and species pool of tropical forests, they provide an indication on the kinetics of the early stages of forest regrowth. Chapter 6 summarizes the main findings of the thesis and provides additional reflections and prospects for future research. By predicting forest attributes country-wide or retrieving land use history over the 30 years time-span of the Landsat archive, the developed methods provide insights at spatial and temporal scales that are beyond the reach of ground based data collection methods. Remote sensing was therefore able to provide valuable information for better understanding, managing and conserving tropical forest ecosystems, and this was partly achieved by combining multiple sources of data and taking advantage of the available remote sensing dimensions. However, the work presented only explores a small part of the potential of remote sensing, so that future research should intensively focus on further exploiting the multiple dimensions and multi-scale nature of remote sensing data as a way to provide insights on complex multi-scale processes such as interactions between climate change, anthropogenic pressure, and ecological processes. Inspired by recent advances in operational forest monitoring, operationalization of scientific methods to retrieve ecological variables from remote sensing is also discussed. Such transfer of scientific advances to operational platforms that can automatically produce and update ecologically relevant variables globally would largely benefit ecological research, public awareness and the conservation and wise use of natural resources.