Land cover classification in a heterogeneous environment : testing the perfomance of multispectral remote sensing data and the random forest ensemble algorithm.
Ndyamboti, Kuhle Siseko.
MetadataShow full item record
Land use/land cover (LULC) information is essential for a plethora of applications including environmental monitoring and natural resource management. Traditionally, field surveying techniques were the sole source of acquiring such information; however, these methods are labour intensive, costly and time consuming. With the advent of remote sensing, LULC information can be acquired in an economical, less tedious and non-time consuming manner at shorter temporal cycles and over larger areas. The aim of this study was to assess the utility of multispectral remote sensing data and the Random Forest (RF) algorithm to improve accuracy of LULC maps in heterogeneous ecosystems. The first part of this study used moderate resolution SPOT-5 data to compare the performance of the RF algorithm to that of the commonly used Maximum Likelihood (ML) classifier. Results indicated that RF performed significantly better than ML (66.1%) and yielded an overall accuracy of 80.2%. Moreover, RF variable importance measures were able to provide an insight on the bands that played a pivotal role in the classification process. Due to the fact that moderate resolution satellite data was used, both classifiers seemed to experience some difficulties in discriminating amongst classes that exhibited similar spectral responses such as Eucalyptus grandis and Pinus tree plantations, young sugarcane and mature sugarcane, as well as river and ocean water. In that regard, the next section attempted to address this shortfall. The second part of the study used high resolution multispectral data acquired from the WorldView-2 sensor to discriminate amongst six spectrally similar LULC classes using the advanced RF algorithm. Results suggested that the use of WorldView-2 data together with the RF ensemble algorithm is a robust and accurate method for separating classes exhibiting similar spectral responses. The classification process yielded an overall accuracy of 91.23% and also provided valuable insight into WorldView-2 bands that were most suitable for discriminating the LULC categories. Overall, the study concluded that: (i) multispectral remote sensing data is an effective tool for obtaining accurate and timely LULC information, (ii) moderate resolution multispectral data can be used to map broad LULC categories whereas high resolution multispectral data can be used to separate LULC at finer levels of detail, (iii) RF is a robust and effective tool for producing LULC maps that are less prone to error.