Abstract:
: Tropical highlands remain a challenging target for remote sensing due to their high heterogeneity of the landscape and frequent cloud cover, causing a shortage of high-quality and reliable comprehensive data on land use and land
cover on a local or regional scale. These, however, are urgently needed by local stakeholders and decisionmakers. This applies
for example to the Muringato sub-catchment in Nyeri County, Kenya, where acute water problems have been identified to
be usually directly related to specific land use and land cover. This article contributes to the understanding of tropical highlands from a remote sensing perspective by examining Sentinel-1, Sentinel-2 and Global Forest Canopy Height Model data
from the Global Ecosystem Dynamics Investigation, all provided by the Google Earth Engine. To do so, we assess classifiers
derived from these datasets for different land cover types, analyzing the performance of promising candidates identified
in the literature, using 2,800 samples extracted from high-resolution image data across Nyeri County. We also propose an
object-based classification strategy based on sequential masking. This strategy is adapted to very heterogeneous landscapes
by refining image objects after re-evaluating their homogeneity. Small buildings, which constitute a significant part of the settlement structure in the area, are particularly difficult to detect. To improve the recognition of these objects we additionally
consider the local contrast of the relevant classifier to identify potential candidates. Evaluating our sample data, we found
that especially optical indices like the Sentinel Water Index, the Enhanced Normalized Difference Impervious Surfaces
Index or specific Sentinel-2 bands combined with canopy height data are promising for water, built-up or tree cover detection. With these findings, our proposed object-based classification approach is applied to the Muringato sub-catchment as a
representative example of the Kenyan tropical highland region. We achieve a classification accuracy of approximately 88%
in the Muringato sub-catchment, outperforming existing products available for the study area. The knowledge gained in the
study will also be used for future remote sensing-based monitoring of the region.
