The principal goal of this thesis was to map and monitor slums using geoinformation technologies. The partial aims were to develop a test methodology for mapping and monitoring slums using open data, to leverage satellite imagery, geophysical datasets and complementary data for slum detection within different cities, to evaluate the functionalities of different algorithms and determine which algorithm provides the best result and to monitor the slum growths in the context of curbing disease outbreak.
Two study areas with different slum morphology and relief characteristics were chosen for this thesis to test whether the algorithms can map and monitor slums with different slum characteristics. The first study area was Lagos Mainland local government in Lagos, Nigeria in which Makoko slum (majorly built on water (Lagos lagoon) with sticks and woods shelters) is one of the settlements in this region. The second study area was Vila Andrade district in Sao Paulo, Brazil, which housed the Paraisopolis slum (one of the most populated slums in Brazil). In this study, three imagery were used: the Sentinel-2 imagery being the major dataset, the drone imagery, and orthophoto.
To achieve the aim of this thesis, the workflow was categorized into data acquisition, processing and analysis and finally, web application development. The data were acquired from their available platforms. Sentinel-2 imagery of the study areas were acquired from the Copernicus open access hub web portal, DJI Mavic Pro was leveraged to acquire the drone imagery of the first study area, the orthophoto of the second study area was downloaded in tiles from the Sao Paulo open access Geosampa web page and lastly, the administrative boundaries of the study areas were downloaded from GADM open portal.
The data processing and analysis involved image mosaicing, image classification, and slum change detection using the ArcGIS Pro software due to its robustness and capability for numerous GIS applications. The supervised image classification in the classification wizard of the ArcGIS Pro which housed the classification methods (pixel-based and object-based) and classification algorithms (maximum likelihood (ML), random trees (RT) and support vector machine (SVM)) was used for the image classification process.
Before the classification proper, the drone images were mosaiced and also the orthophoto since they were captured in tiles. Classification schema was created with five (5) classes (slum (buildings), non-slum (buildings), water, vegetation and roads) and training samples (polygon) were selected from the imagery, respectively. The algorithms were trained with the training sample, the trained algorithms were applied to classify the imagery and the classes were reclassed into three (3) i.e. slums (buildings), non-slums (buildings) and others (the water, vegetation and roads were combined).
Foremost, the analysis was achieved on the Sentinel-2 imagery of the study areas with the pixel-based method. At the same time, the algorithms (ML, RT, SVM) were applied respectively for the classification. Afterwards, the same training samples were used to classify the Sentinel-2 imagery with the object-based method using the same algorithms as that of the pixel-based method.
To determine whether the algorithms can map slums from different datasets (imagery) while checking their performances, pixel and object-based classification were carried out on the drone imagery and the orthophoto individually with the same training datasets as explained above.
The deep learning (DL) classification was the third method and was carried out on the drone imagery and the orthophoto. DL classification has a separate workflow in ArcGIS Pro which followed the sequence of preparing the training data, training the deep learning model and lastly classify the imagery with the trained model. Using the same training samples from the supervised classification, the deep learning model (U-Net model) was trained and was applied to classify the drone imagery. The orthophoto was classified with the same process.
The finding of this study showed that the pixel-based and object-based SVM algorithms outperformed other algorithms for all datasets however the object-based SVM had good prediction with an overall accuracy of 68% over the pixel-based SVM (63.1%). The RT algorithm for both pixel and object-based methods had accuracies of 58.4% and 52.8% respectively followed by the ML algorithm with overall accuracies of 49.8% and 38.7% for both methods. The deep learning classification had an overall accuracy of 60%.
This thesis demonstrated the capability of Sentinel-2 imagery for mapping and monitoring slums due to its free access and global coverage. Also, VHR imagery acquired from drone and LIDAR platforms are a very reliable source of information for slum mapping as proved in this thesis. This study can be used as a test methodology for mapping and monitoring slums in other locations however to improve the accuracy of the results, more input data (DEM, DSM, land cover data) are evident. Since slums is a global challenge, this study suggests that slum areas should be added as a global land cover class and be separated from urban areas (buildings) when generating an updated global land cover.