Utilization of mid-resolution imagery and remote sensing techniques to characterize urban growth over the Philippi Horticultural Area

Abstract

The Philippi Horticultural Area (PHA) is a peri-urban area with a long history of food production dating back to the mid-1800s. The total area of the PHA comprises over 3000 hectares, of which 1200 hectares are suitable for food production. However, farming within the PHA has been affected by increased development in the area. Thus, the need for a spatial understanding of urban growth in the area is imperative. This study aims to utilize satellite imagery combined with remote sensing to identify vegetation and urban indices and to analyse land use through change detection mapping. Part of the process was to determine the change detection through the classification of satellite images and the insertion of vegetation indices in order to produce accurate land cover changes on a map. Limitations to the study were restriction to sensors being used, the years available of images for the study area. Certain population statistics were not available for the years in question and only three classifiers could be used. The study was done between the years 2015 and 2021. This study used image classification and PlanetScope digital images data which were essential in the application that was used in this study, namely machine learning (ML). The need to provide analysis and decision-making for the PHA has been a challenge, but currently, remote sensing has globally been applied with the use of current advanced satellite systems and sensors. In terms of classifying satellite imagery, three machine learning techniques have been applied to this research, namely Random Forest (RF), K Nearest Neighbour (K-NN), and Support Vector Machines (SVM). The best-performing classifier was used to classify the images into six classes namely urban fabric, water, vegetation (agricultural and natural), and bare ground (sand and bare ground). Change detection was done on images of consecutive years by displaying differences over time and then by mapping the trajectory of urban growth. The main findings in this study was the growth in urban fabric. Urban fabric started at 30% during 2015 and 2016, it increased by 2% during 2016 and 2017, and a further 3%% between 2017 and 2018, there was a slight decrease by 1% between 2018 and 2019, it increased slightly by 1% in 2019 and 2020 and further increased by another 1% between 2020 and 2021. In contrast, farming area started off at 10% for 2015 and 2016, it increased by 9% during 2016 and 2017, reduced by 9% again during 2017 and 2018, there was another increase by 7% between 2018 and 2019, a slight decrease of 2% between 2019 and 2020 and further decreased by a 3% between 2020 and 2021. The vegetation indices resulted in the following being found with the overall classification accuracy showing improvement with the inclusion of indices for each year. In 2015 the accuracy was found to increase by .04%, in 2016 the accuracy improved by 6.6%, in 2017 the accuracy improved by 27.5%, in 2018 the accuracy increased by 10.8%, in 2019 the accuracy increased by 1.6%, in 2020 the accuracy improved by 18.6% and in 2021 the accuracy increased by 7.8%. The accuracy results for two of the three classifiers—Random Forest (RF) and Support Machine Vector—were comparable. Despite the great accuracy of the two classifiers, Random Forest (RF) consistently outperformed Support Vector Machines. The findings indicate that high accuracy classifications for mapping the agricultural and urban edge are possible using both the SVM and RF classifiers. It is clear that the study region influences how well the classifiers function.