Chapters published as articles

This thesis includes one publication in a scientific journal (Chapter 2), one accepted for its publication in September 2018 (Chapter 3), and one submitted to consider its publication (Chapter 4). Minor changes have been made to the first publication to fit this thesis. A synthesis of main findings is presented in Chapter 5, targeting the research objectives and questions formulated, finalizing with conclusions and outlook.

Santos F., Dubovyk O. & Menz G., 2017. Monitoring forest dynamics in the Andean Amazon: The applicability of breakpoint detection methods using Landsat time-series and genetic algorithms. Remote Sensing, 9(1).

Abstract: The Andean Amazon is an endangered biodiversity hotspot but its forest dynamics are less studied than those of the Amazon lowland and forests in middle or high latitudes. This is because landscape variability, complex topography and cloudy conditions constitute a challenging environment for any remote-sensing assessment. Breakpoint detection with Landsat time-series data is an established robust approach for monitoring forest dynamics around the globe but has not been properly evaluated for implementation in the Andean Amazon. We analyzed breakpoint detection- generated forest dynamics in order to determine limitations when applied to three

12

different study areas located along an altitude gradient in the Andean Amazon in Ecuador. Using the available Landsat imagery for the period 1997–2016, we evaluated different pre-processing approaches, noise reduction techniques, and breakpoint detection algorithms. These procedures were integrated into a complex function called the processing chain generator. Calibration was not straightforward since it required defining values for 24 parameters. To solve this problem, we implemented a novel approach using genetic algorithms. We calibrated the processing chain generator by applying a stratified training sampling and a reference dataset based on high resolution imagery. After the best calibration solution was found and the processing chain generator executed, we assessed accuracy and found that data gaps, inaccurate co- registration, radiometric variability in sensor calibration, unmasked cloud, and shadows can drastically affect the results, compromising the application of breakpoint detection in mountainous areas of the Andean Amazon. Moreover, since breakpoint detection analysis of landscape variability in the Andean Amazon requires a unique calibration of algorithms, the time required to optimize analysis could complicate its proper implementation and undermine its application for large-scale projects. In exceptional cases when data quality and quantity are adequate, we recommend the pre-processing approaches, noise reduction algorithms and breakpoint detection algorithms procedures that can enhance results. Finally, we include recommendations for achieving a faster and more accurate calibration of complex functions applied to remote sensing using genetic algorithms.

Santos F., Meneses P. & Hostert P., 2018. Monitoring Long-Term Forest Dynamics with Scarce Data: A Multi-Date Classification Implementation in the Ecuadorian Amazon. European Journal of Remote Sensing (Manuscript accepted for publication).

Abstract: Monitoring long-term forest dynamics is essential for assessing human- induced land-cover changes, and related studies are often based on the multi-decadal Landsat archive. However, in areas such as the Tropical Andes, scarce data and the resulting poor signal-to-noise ratio in time-series data renders the implementation of

13

automated time-series analysis algorithms difficult. The aim of this research was to investigate a novel approach that combines image compositing, multi-sensor data fusion, and post-classification change detection that is applicable in data-scarce regions of the Tropical Andes, exemplified for a case study in Ecuador. We derived biennial deforestation and reforestation patterns for the period from 1992 to 2014, achieving accuracies of 82 ± 3% for deforestation and 71 ± 3% for reforestation mapping. Our research demonstrates that an adapted methodology allowed us to derive the forest dynamics from the Landsat time-series despite the abundant regional data gaps in the archive, namely across the Tropical Andes. This study therefore presents a novel methodology in support of monitoring long-term forest dynamics in areas with limited historical data availability.

Santos F. & Graw V., 2019. Analyzing Underlying Causes of Deforestation and Reforestation in the Central Ecuadorian Amazon: A Geographically Weighted Ridge Regression Approach. PLOS ONE (Manuscript submitted for publication).

Abstract: The Tropical Andes region encompasses endangered biodiversity hotspots with high conservation priority. Deforestation due to population growth and agriculture expansion is therefore one of the main threats to this region and thus highlights the importance of understanding the drivers of this process on multiple scales. On the other hand, the drivers of reforestation and their role in forest recovery are less known. Therefore, we propose an interdisciplinary approach to analyze both deforestation and reforestation drivers by applying geographically weighted ridge regression. This method evaluates spatial non-stationarity and provides surface representations of local parameter estimates to identify regions where drivers show higher significance for either deforestation or reforestation. Our analysis includes nine different variable groups and two predictors using socio-economic data from population censuses, accessibility models and biophysical features. Information on deforestation and reforestation were based on remote sensing input data. We used dasymetric mapping in conjunction with land-cover maps to downscale areal-based data and improve the

14

spatial resolution of our analysis. We conducted our research in the Tropical Andes of the Ecuadorian Amazon, a highly heterogeneous region, within the time period 2000 - 2010. Areas were highlighted where improved accessibility to palm oil, coffee, cacao and milk production facilities motivated deforestation, while reforestation seems to follow the opposite trend. Moreover, gender, ethnicity and household structure showed a high influence on untangled population dynamics and their relationship with forest change. This approach demonstrates the benefits of integrating remote sensing derived products and socio-economic data for understanding coupled socio-ecological systems from local to global scales.

15