Quantification of land use and land cover in a Monsoon agricultural mosaic from space

Land use and land cover (LULC) are fundamental elements of the global ecosystem and LULC changes are key aspects of global change. Information on LULC is essential in a wide range of research fields, including environmental science, ecosystem services, and environmental decision making. The quality of LULC information significantly impacts on the outcomes of these research applications. Hence, acquisition of appropriate LULC data is an important issue for research, especially in complex heterogeneous agricultural landscapes. Particularly in these types of landscapes, the existing global land cover (GLC) products are restricted in their thematic, spatial, and temporal resolution. Therefore, the use of the GLC products may lead to an inadequate representation of the actual landscape. For cultivated landscapes, methods able to retrieve detailed LULC data as well as improvements of GLC products are strongly desired. This dissertation focuses on enhancing LULC quantification in complex heterogeneous agricultural landscapes. Specifically, extraction of spatially and thematically detailed LULC information from existing, medium resolution, multi-spectral satellite products is pursued. Three main contri- butions to LULC quantification are presented: ground data collection, derivation of continuous LULC, and classification of multi-crop LULC. First, high-quality LULC observation data was collected over the study site Haean catchment, South Korea. The observed data illustrates the detailed LULC of the catchment for the three-year study period (2009 – 2011). A comparison with the MODerate Resolution Imaging Spectroradiometer (MODIS) land cover product (MCD12Q1) revealed limitations of this GLC product in spatial and thematic resolution. The limitations were due to the large cell size and the broadly defined cropland classes of the product. This result illustrates the difficulty in using GLC products to monitor LULC changes in complex heterogeneous landscapes. Second, estimation of continuous LULC was addressed. For the study site, a fractional LULC regression model was developed for 10 LULC classes based on a MODIS multi-spectral dataset (MODIS 13Q1) and Random Forests models. In order to allow for making informed decisions when choosing data-processing options, three key data-processing options were evaluated: selection of spectral predictor sets (NDVI, EVI, surface reflectance, and all combined), time interval (8-day vs. 16-day), and smoothing (no smoothing vs. Savitzky-Golay filter). The models successfully reproduced spatial distributions of the LULC fractions, thus illustrated the potential of existing, medium resolution satellite products for continuous LULC estimation. Third, a multi-crop LULC classification model was developed to improve thematic LULC representation. LULC data tends to be imbalanced as majority types dominate over minority types (e.g. un- equal distributions of LULC type labels in raster maps). This imbalance is partly a cause of the under-development of multi-crop LULC products. Here, a synthetic sampling method was used to alleviate the problem of data imbalance in the LULC observation data for the study site. Artificial balancing of the training data substantially increased the classification performance of some minority LULC types. However, other minority LULC types remained difficult to classify due to substantial class overlaps (i.e. spectral similarities between LULC types). For ecosystem research and decision making, continuous representations of LULC and multi-crop LULC are key information sources. In this dissertation, approaches connecting extensive field work, remote sensing and state-of-the-art analysis methods (e.g. Random Forests) are proposed and evaluated. It is shown that a judicious choice of data processing options (e.g. avoiding exces- sive data smoothing) and synthetic resampling methods can be useful to achieve better LULC presentations from medium resolution remote sensing data in complex cultivated landscapes. The data analysis approach presented in the dissertation was designed to be transferable to other landscapes. The methods can help analysing publicly available remote sensing data for creating detailed spatial and thematic representations of LULC types such as cultivated crops, and enhancing existing global land use and land cover products.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bayreuth

Year

2015

Author dc:creator

• Seo, Bumsuk

Contributors dc:contributor

• Björn, Reineking

Repository record source_url

https://epub.uni-bayreuth.de/id/eprint/2647/

OAI identifier oai:identifier

oai:epub.uni-bayreuth.de:2647