Browsing by Author "Adesuyi, Ayodeji Steve"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Automating land cover classification using time series NDVI : a case study in the Berg River Catchment Area
    (Stellenbosch : Stellenbosch University, 2016-12) Adesuyi, Ayodeji Steve; Munch, Zahn; Stellenbosch University. Faculty of Arts and Social Sciences. Dept of Geography and Environmental Studies.
    ENGLISH ABSTRACT: The processing of large volumes of geographic information system (GIS) and remote sensing (RS) data necessitates the development of automated techniques which are cost-effective, faster and user-friendly in order to aid spatial decision making. In this study, an automated technique for identifying agricultural land cover was developed using a custom tool. Multiple ensemble classifiers in ArcGIS workflow automation tool (MEAWAT) was tested on time-series MODIS normalised difference vegetation (NDVI) data using the Berg River catchment area of Western Cape, South Africa as a case study. Although the tool was developed to perform agricultural land cover classification using MODIS input data, the tool was subsequently applied to Landsat NDVI data of the same study extent. A few modifications to the tool were implemented to accommodate the different satellite imagery. The tool was built on an ArcGIS/Python platform, and various GIS & RS functions usually performed in a variety of different software packages were integrated, including study area selection, reprojection, classification and accuracy assessment. The NDVI phenology curve was used to create training data for the classification. Different parameters were tested which allow users to engage with different rules and derive a suitable land cover map for their purpose. MEAWAT uses decision tree and ensemble classifiers such as random forest and extra-tree as well as boosting using a meta-estimator (AdaBoost). Classification accuracies of 70.5%, 75.5%, 76.3% and 78.7% were achieved respectively with MODIS data, while an accuracy of 89% was achieved using the boosted random forest classifier on the Landsat data. It was observed that a better classification output can be derived using MEAWAT on higher resolution satellite imagery provided good training data are available. These findings highlight the potential of MEAWAT for large dataset land cover classification using different satellite imagery. In addition, it exposed limitations of the tool, indicating that various adjustments will be needed on the tool when working with other satellite imagery different from MODIS and Landsat.