Epsilon Archive for Student Projects

Abstract

Despite improved spatial and spectral characteristics of satellite and aerial imaging systems, land-cover classification is still challenged by a continuously evolving and complex rural and urban landscape conditions resulting from diverse land-use scenarios. Sizes and material composition of impervious surfaces changes greatly from urban to rural areas, leading to varying spectral signatures and ultimately misclassification. This creates a challenge in choosing suitable classification algorithms and image processing methods. In this study, the influence of spatial resolution and land-cover spectral and spatial heterogeneity on accuracy of land-cover classification at a rural-urban interface was examined alongside comparison of Random Forest (RF) and Support Vector Machine (SVM) classification algorithms. Further, the performance of spectral unmixing strategies was tested against standard feature extraction methods, namely, NAPCA and PCA. The results showed a 10 % improvement in classification accuracy from 30 m to 10 m spatial resolution for both overall accuracy and Kappa coefficients, however, relatively high per-pixel class disagreement (39 %) was recorded between the different resolution maps, pointing to the fact that overall accuracy or Kappa coefficients may not capture the spatial resolution effects on classification accuracy results in its entirety. SVM classifier proved superior to the RF classifier with even a relatively bigger margin at the coarser spatial resolution (i.e. 4.9 % and 5.7 % higher accuracy at 10 m and 30 m spatial resolution respectively). Higher classification accuracies were observed for partial unmixing and sum-to-unity unmixing feature extraction strategies at both spatial resolutions relative to the results from PCA, NAPCA and original image data (i.e. 62 %, 61 %, 51 %, 61 % and 59 % respectively for 30 m resolution, and, 67 %, 67 %, 62 %, 65 % and 66 % respectively for 10 m resolution image). It was found that the dominance of unmixing-based feature extraction methods reduced while the standard dimensionality reduction approaches (NAPCA and PCA) made a zero contribution to improving classification accuracy at finer spatial resolution (i.e. 10 m). According to the results of land-cover heterogeneity assessment, more fragmented and spatially diverse landscapes were comparably more spectrally diverse along the rural-urban gradient. A high degree of landscape heterogeneity and lowest classification accuracy was observed in the peri-urban region at approximately 11 kilometers from the very urban area. The findings indicate that landscapes with high PD, LSI, SHDI and low CONTAG have lower accuracy while homogeneous and less fragmented landscapes have higher accuracy. The findings from the study will provide a basis for more accurate time series analysis of land-use dynamics at the rural-urban interface.