Bi-Temporal Analysis of High-Resolution Satellite Imagery in Support of a Forest Conservation Program in Western Uganda

Friday, 19 December 2014
Nancy Thomas1, Eric Lambin1, Robin Audy2, Byamukama Biryahwaho3, Joost de Laat4 and Seema Jayachandran5, (1)Stanford University, Stanford, CA, United States, (2)Innovations for Poverty Action, New Haven, CT, United States, (3)Nature Harness Initiatives, Kampala, Uganda, (4)World Bank, Washington, DC, United States, (5)Northwestern University, Evanston, IL, United States
Recent studies in land use sustainability have shown the conservation value of even small forest fragments in tropical smallholder agricultural regions. Forest patches provide important ecosystem services, wildlife habitat, and support human livelihoods. Our study incorporates multiple dates of high-resolution Quickbird imagery to map forest disturbance and regrowth in a smallholder agricultural landscape in western Uganda. This work is in support of a payments for ecosystem services (PES) project which uses a randomized controlled trial to assess the efficacy of PES for enhancing forest conservation. The research presented here details the remote sensing phase of this project. We developed an object-based methodology for detecting forest change from high-resolution imagery that calculates per class image reflectance and change statistics to determine persistent forest, non-forest, forest gain, and forest loss classes. The large study area (~ 2,400 km2) necessitated using a combination of 10 different image pairs of varying seasonality, sun angle, and viewing angle. We discuss the impact of these factors on mapping results. Reflectance data was used in conjunction with texture measures and knowledge-driven modeling to derive forest change maps. First, baseline Quickbird images were mapped into tree cover and non-tree categories based on segmented image objects and field inventory data, applied through a classification and regression tree (CART) classifier. Then a bi-temporal segmentation layer was generated and a series of object metrics from both image dates were extracted. A sample set of persistent forest objects that remained undisturbed was derived from the tree cover map and the red band (B3) change values. We calculated a variety of statistical indices for these persistent tree cover objects from the post- survey imagery to create maps of both forest cover loss and forest cover gain. These results are compared to visually assessed image objects in addition to a Landsat-time series based global forest change map (Hansen et al. 2013) to evaluate the change detection approach. We present the forest change rates and assess the impact of the PES project at both the village and individual forest plot levels.