Mapping Eucalyptus trees in Johannesburg city using high resolution multispectral image

Abstract

Invasive alien plants are considered as a major threat to ecological and socioeconomic systems. Nevertheless, because of the socioeconomic benefits some alien plants provide, their management is often complicated by controversies. Thus, understanding their spatial distribution and abundance facilitates management decision making processes of invasive alien species. Mapping plant species in a heterogeneous environment such as highly urbanized areas is often complicated by high spectral confusion between species. This study investigated the utility of new generation WorldView-2 (WV2) satellite imagery with both high spectral and spatial resolution in mapping eucalyptus in the historical mining area located south of Johannesburg city. It also evaluated if the medium spatial resolution satellite image SPOT-7 could be used as a cheaper alternative to map eucalyptus trees in an urban environment. Furthermore, the performances of Random Forest (RF) and Support Vector Machines (SVM) were compared to determine the most effective classification algorithms between the two methods. Both WV-2 image and SPOT-7 image attained satisfactory overall accuracies, although the WV-2 performed better than the SPOT-7 imagery. WV-2 attained accuracies of 81.67% (0.78 kappa) for RF algorithm and 80% (0.76 Kappa) for SVM algorithm, whilst SPOT-7 had overall accuracies of 72.78% (0.67 kappa) for RF and 71.11% (0.65 Kappa) for SVM. Although the overall accuracies for SPOT image was satisfactory, the user’s accuracies for the eucalyptus class was very low (60% and 56.67% for RF and SVM algorithms, respectively). This suggests that WV-2, with higher user’s accuracies for the eucalyptus class (73.33% and 70% for RF and SVM algorithms, respectively), is more suitable for mapping eucalyptus trees in an urban area than the SPOT data. The two classification algorithms showed high accuracy levels for both satellite data, although RF had slightly higher accuracies than SVM. The combination of WV-2 image and RF produced a more accurate map of the eucalyptus trees in the study areas. The overall accuracy was 81.67% and a kappa coefficient of 0.78 and eucalyptus class attained user’s and producer’s accuracies of 73.33% and 75.86%, respectively.