Integrating openstreetmap data and sentinel-2 Imagery for classifying and monitoring informal settlements

INTEGRATING OPENSTREETMAP DATA AND SENTINEL2 IMAGERY FOR CLASSIFYING AND MONITORING INFORMAL SETTLEMENTS Brenda Ayo ii INTEGRATING OPENSTREETMAP DATA AND SENTINEL-2 IMAGERY FOR CLASSIFYING AND MONITORING INFORMAL SETTLEMENTS Dissertation supervised by Joel Dinis Baptista Ferreira da Silva, PhD Instituto Superior de Estatística e Gestão de Informação, Universidade Nova de Lisboa Lisbon, Portugal Dissertation co-supervised by Prof. Dr. Hanna Meyer, Institute of Landscape Ecology, Heisenbergstr. 2, D-48149 Münster Dissertation co-supervised by Ignacio Guerrero Institute of New Imaging Technologies, Universitat Jaume I Castellón de la Plana, Spain February 2020 iii DECLARATION OF ORIGINALITY I declare that the work described in this document is my own and not from someone else. All the assistance I have received from other people is duly acknowledged and all the sources (published or not published) are referenced. This work has not been previously evaluated or submitted to NOVA Information Management School or elsewhere. Lisbon, Portugal, January 2020 Brenda Ayo iv ACKNOWLEDGMENTS Principally, I would like to thank my Supervisors, Joel Dinis Baptista Ferreira da Silva (PhD), Prof. Dr. Hanna Meyer and Ignacio Guerrero. You were instrumental from the inception, through the execution and end of this Thesis. I am extremely grateful and honored to have collaborated with you. I am extremely grateful to all the professors that helped through this master’s program especially professors Marco Painho and Christoph Brox for their guidance and support throughout the program duration. Thank you to my cohort and fellow MSc. GeoTech graduate students – especially Damien and Chamodi - you have been incredible resources, whose positive and hardworking spirit has made my Master’s experience so much more enjoyable and fun. Lastly, thanks to my friends and family who were of great support even if you were thousands of miles away. v Integrating Openstreetmap Data and Sentinel-2 Imagery for Classifying and Monitoring Informal Settlements ABSTRACT The identification and monitoring of informal settlements in urban areas is an important step in developing and implementing pro-poor urban policies. Understanding when, where and who lives inside informal settlements is critical to efforts to improve their resilience. This study aims at integrating OSM data and sentinel-2 imagery for classifying and monitoring the growth of informal settlements methods to map informal areas in Kampala (Uganda) and Dar es Salaam (Tanzania) and to monitor their growth in Kampala. Three building feature characteristics of size, shape and Distance to nearest Neighbour were derived and used to cluster and classify informal areas using Hotspot Cluster analysis and ML approach on OSM buildings data. The resultant informal regions in Kampala were used with Sentinel-2 image tiles to investigate the spatiotemporal changes in informal areas using Convolutional Neural Networks (CNNs). Results from Optimized Hot Spot Analysis and Random Forest Classification show that Informal regions can be mapped based on building outline characteristics. An accuracy of 90.3% was achieved when an optimally trained CNN was executed on a test set of 2019 satellite image tiles. Predictions of informality from new datasets for the years 2016 and 2017 provided promising results on combining different open source geospatial datasets to identify, classify and monitor informal settlements. vi KEYWORDS Informal Settlements; Remote Sensing; Urbanization; Machine Learning (ML); Random Forest (RF); Convolutional Neural Networks (CNN); OpenStreetMap (OSM) Sentinel-2 satellite imagery vii ACRONYMS CNN - Convolutional Neural Networks (CNNs) DEM - Digital elevation model DSM - Digital surface model EO – Earth Observation GEOBIA - Geographic object-based image analysis GPS - Global positioning system GSO - Global slum ontology HR - High Resolution Satellite Imagery ISDA - Informal settlement database atlas ISUP - Informal settlement upgrading Programme KOTAKU - Kota Tanpa Kumuh LiDAR - Light detection and ranging MGD - Millennium development goals MKL - Multiple Kernel Learning MLTs - Machine Learning Techniques OBIA - Object-based image analysis OHSA - Optimized Hotspot Analysis OOA - Object-oriented analysis OSM – Open Street Map PSUP - Participatory slum upgrading Programme RF - Random forest RGB – Red Green Blue SDI - Shack/Slum Dwellers International SVM - Support Vector Machine UAV - Unmanned Aerial Vehicle viii UN – United Nations UN-Habitat – United Nations Human Settlements Programme VGG - Visual Geometry Group VHR - Very high resolution VHR – Very High Resolution Satellite Imagery WDI - World Development Indicators ix INDEX OF THE TEXT DECLARATION OF ORIGINALITY .................................................................. iii ACKNOWLEDGMENTS ....................................................................................... iv ABSTRACT .............................................................................................................. v KEYWORDS ........................................................................................................... vi ACRONYMS ......................................................................................................... vii INDEX OF THE TEXT .......................................................................................... ix INDEX OF TABLES ............................................................................................... xi INDEX OF FIGURES ........................................................................................... xii 1. INTRODUCTION .......................................................................................... 1 1.1. Background and Motivation........................................................................... 1 1.2. Research Gap Identification ........................................................................... 3 1.3. Research Aim ................................................................................................ 4 1.4. Methodology overview .................................................................................. 5 1.5. Thesis Structure ............................................................................................. 5 2. LITERATURE REVIEW ............................................................................... 6 2.1. Characteristics of Informal Settlements .......................................................... 6 2.1.1. Building characteristics: ......................................................................... 6 2.1.2. Access Characteristics: ........................................................................... 7 2.1.3. Location and neighbourhood Characteristics: ......................................... 7 2.2. OpenStreetMap for Urban Analytics .............................................................. 7 2.3. Approaches for identifying and Mapping Informal Settlements ...................... 8 2.3.1. Survey and census-based approach ......................................................... 8 2.3.2. Participatory-based approach.................................................................. 9 2.3.3. GIS and Remote sensing-based approach ............................................... 9 2.4. Hot Spot Clustering Analysis ....................................................................... 11 2.5. An overview of Machine Learning Techniques ............................................ 11 2.5.1. Convolutional neural-networks.............................................................. 13 3. DATA AND CASE STUDY .......................................................................... 16 3.1. Case Study .................................................................................................. 16 3.2. Datasets and Pre-Processing ........................................................................ 18 3.2.1. OSM Data ............................................................................................. 18 3.2.2. Satellite Imagery ................................................................................... 19 3.2.3. Reference data....................................................................................... 20 4. RESEARCH METHODOLOGY ................................................................. 21 4.1. Data Enrichment and Computation of Parameters ........................................ 21 4 polygons, and at the same time identify informal regions across the entire city and not just particular places. There is also need to examine whether similar building polygons in informal settlements - similar in the sense of geometrical properties exhibit a spatial pattern. The informal settlement buildings characteristics coupled with Machine learning algorithms could help to predict where informal settlements could be and also if these characteristics are exhibited in varying geographical locations. This will help reduce the aggravation that comes up during their detection for inventory or planning issues and also provide more insight and understanding on the growth of these settlements. The other challenge with mapping informal settlements at city, regional or country level is the absence of adequate data for analysis at these scales. The availability of open source satellite imagery like Landsat and Sentinel data provide a wider coverage but the spatial resolutions they offer does not provide for adequate distinguishing between formal and informal settlements in cities. However, with accurate delineating of these settlements at a finer scale, these satellite images can be used to monitor the growth of informality over time since the integration of spatio-temporal analysis of informal settlements is still inadequate. 1.3. Research Aim The main aim of this research work is to integrate OSM data and sentinel-2 imagery for classifying and monitoring the growth of informal settlements. To achieve the main research Aim, the following sub questions are addressed: 1. How can we exploit the potential of using buildings outline characteristics such as size and shape to differentiate Informal settlements from Formal Settlements? 2. Is it possible to predict informal areas in a city by understanding housing informality in other cities of similar context using buildings outline characteristics and machine learning? 3. What is the most appropriate Machine Learning technique based on accuracy to predict informal areas in a city based on buildings outline characteristics? 4. How can we exploit the potential of freely available Sentinel-2 satellite imagery with advanced machine learning to estimate the growth of Informal settlements? 5 1.4. Methodology overview Based on the research questions, the following methodology workflow was adapted:  Preprocessing of OSM data and Satellite Imagery. This involved data cleaning and preparation suitable for Analysis.  OSM Building polygons enrichment with Feature geometric and topologic characteristics.  Hotspot clustering analysis on building features training dataset to cluster Informal settlement buildings from the formal ones.  Test and select the best classifier based on accuracy to predict informal areas in a city based on similarity context of buildings outline characteristics.  Sentinel-2 image classification. This stage involves slicing the images into tiles and using the predicted Informal settlement regions from the step above to generate training set on image tiles of corresponding date period. Thereafter, a Convolutional Neural Networks Model is trained to identify Image tiles that belong to Informal settlements. The trained model is then used to predict Informal settlement image tiles for other temporal datasets. 1.5. Thesis Structure The research is organized as follows:  Chapter 2 reviews the related works and theoretical background on informal settlements, approaches for mapping and identifying informal settlements, and Machine Learning Algorithms  Chapter 3 describes the study area, datasets and tools used for the research.  Chapter 4 describes the implementation details of the methodology and the experiments conducted  Chapter 5 presents the results;  Chapter 6 addresses the analysis and discussion of the results including the limitations of the research;  Chapter 7 presents a summary of the conclusions by answering the main re-search questions and recommendations for future works. 6 2. LITERATURE REVIEW This chapter gives an overview on the fundamental concepts related to this research and the current state of research in the field of mapping informal settlements It highlights the physical characteristics of informal settlements, a discussion on various approaches used to map informal settlements and the integration of GIS and informal settlements and Machine Learning Algorithms. 2.1. Characteristics of Informal Settlements The challenge with mapping and identify what an informal settlement is starts with the absence of a standard definition which has resulted into failure to incorporate these settlements into census and demographic surveys at national or even district level [23]. Based on literature, they take on names like Slum, Informal, Squatter, Spontaneous, Ghetto, Illegal, Irregular, or described by local names such as Favela, bidonville (French), mudun-safi (Arabic), bairros da lata (Portuguese) township or gecekondu (Turkish), to mention but a few [24]. The UN-Habitat qualitatively defined a slum as a household that is lacking either of improved water, improved sanitation, tenure security or overcrowded living environment is based on a measure of deprivation indicators. On the other hand, the definition of informal settlements differs from slum as those areas that developed through unauthorized occupation of land outside a legal, regulatory, planned and professional framework for example [26, 27]. Therefore, quantifying and classifying informal settlements/ slums in a universally acceptable way is particularly difficult because (a) What is referred to as a slum in one country may be a good quality of living and (b) over time and a settlement can get formalized through slum upgrading programs [27]. However, over the years, some features have been unique to informal settlements which include: 2.1.1. Building characteristics: Buildings in informal settlements tend to be smaller (10 to 40 m2), have simpler shapes that are mostly rectangular and heterogeneous orientation compared to those in formal settlements. The building density in slums is usually high since the spacing and gap between buildings in almost nonexistence [29, 30]. 7 2.1.2. Access Characteristics: Slums generally have irregular road access network with roads that vary in type, surface and width. [30]The roads are always short in length with many dead ends and dangles. The roads and access network in Informal settlements are usually narrow with limited usage by motor vehicles because they are either two narrow to let a car pass or can only allow access of on vehicle at a time. Slums are characterized with mostly footpaths. 2.1.3. Location and neighbourhood Characteristics: Slums are usually located in hazardous areas like near dumping sites, wetlands, along railways, sewerage canals. [31]This is because most of the land where these features are located is public land and in abandoned. They are usually close to employment opportunities for unskilled and low skilled jobs like manufacturing industries to ease access to employment. 2.2. OpenStreetMap for Urban Analytics Access to spatial data has changed rapidly over the years from traditionally prohibitive sources to openly licensed content and data due to changes in Information Technology and Communication (ITC) brought about by internet, social media and inexpensive portable GNSS devices like mobile phones and Handheld GPS [10, 32]. OSM is the most popular geospatial open data source platform containing billions of entries of VGI and is maintained by a massive community of mappers from around the world all working to-wards the goal of curating accurate and complete geospatial data. In the recent years, there has been an increased use of VGI data not only in GIScience but also in other fields like ecology, planning, computer science etc. [33,34,35]. for academic research and studies. OSM roads layers are associated with the highest completeness compared to buildings based on a recent study [36] that estimates that the OSM roads have reached more than 80% of completeness at a global scale. OSM datasets have been used to extract urban information and for analysis [37, 38]. These datasets are usually used as a supplementary source of information for land-use mapping with emphasis on classifying artificial surfaces [39], by combining OSM and satellite images to extract information related to the urban environment [39–42]. Few research studies exist on building footprints data enrichment for urban analysis [43, 44]. In these studies, automatic methods are developed by using the geometric and topological features in footprint data, in order to enhance the maps with the building usage information. Henn et al. [45] derived the architectural types 8 of buildings based on 3D coarse block models with vertical walls and flat roofs using Support Vector Machines (SVMs), whereby, geometric features such as length, width, area, and degree of perpendicularity of building footprints, types of buildings, as well as height information of buildings are required for the classification process. 2.3. Approaches for identifying and Mapping Informal Settlements Studies on the development and existence of informal settlements have been carried out not only by geographers but also social scientists who put emphasis on social and economic aspects of informal settlements. Due to their complexity and ever varying nature, informal settlements are usually inadequately represented on urban and city maps which leaves them vulnerable to neglect by concerned government agencies. There are broadly three methods used to collect data to inform characterization and classification of informal settlements: household surveys and census, Participatory approaches and Earth Observation imagery analysis using geospatial Techniques [46]. 2.3.1. Survey and census-based approach These approaches use data collected through census and surveys like data on demographics, population, social and economic aspects to map deprivation and poverty. This data maybe coupled with data from official sites like the world Bank is used to map slums based on social, economic and habitat/Infrastructure variables. In [47], an exploratory factor analysis to define the Slum Severity Index (SSI) Mexico City based on measuring the shelter deprivation levels of households from 1990 to 2010. The results showed that the SSI decreased significantly between 1990 and 2000 as a result of several policy reforms but increased between 2000 and 2010. Weeks et al. [48] quantitatively used census data for Accra, Ghana to create a slum index based on the UN slum indicators to measure the concentration of slums. High correlations were found between the slum index, the socio-economic characteristics of neighborhoods and certain land cover metrics derived from VHR satellite imagery. A fusion of open sources physical and socio-economic data in [9] was used to develop an indicator database for characterizing slum settlements by leveraging data mining techniques for mapping slums in Kenya’s major cities. It’s important to note that one of the major challenges is related to census under-coverage due to time constraint, inadequate quality assurance and inaccurate addresses. The absence of 9 censuses in informal settlements is largely due to the inaccessibility of make-shift structures due to political intolerance or general disorder [30]. 2.3.2. Participatory-based approach The participatory approaches involve the cooperation and participation of the informal settlement dwellers in order to generate both spatial and non-spatial information to profile informal settlements [30]. The participatory slum upgrading Programme (PSUP), established by the UN-Habitat and the Slum Dwellers International (SDI) was a participatory approach that encouraged and empowered communities to become active partners with stakeholders in devising strategies to plan sustainable informal settlement upgrades [49]. The Kota Tanpa Kumuh (KOTAKU) platform and program is responsible for the handling of informality in Indonesia through increasing the role of local government and community participation. Informality is defined by buildings, local streets, drinking water supply, community drainage, wastewater management, garbage management, and fire protection [50]. However, Participatory-based approaches are extremely time consuming and effort-intensive is characterized by limited spatial coverage making it difficult to cover larger areas like regional or district levels [30]. 2.3.3. GIS and Remote sensing-based approach The connection between informal settlements and geography has been a focus of study for a long time. Many research studies have investigated the value of using Imagebased identification of informal settlements through remote sensing and other GIS analysis tools for more than 2 decades [23,13,48,49, 50]. Three key criteria are often used in such studies: small grain size, high ground coverage and irregular access networks. Other factors such as proximity to hazardous areas, lack of vegetation and lowquality roofing are sometimes included. Such studies often conflate informality with slums. But while most such studies map a binary distinction between formal and informal morphologies, some adopt a more nuanced approach. Bradley et al [54] deploy the use of both HR and VHR satellite imagery to map informal settlements in developing countries using machine learning. They used the Canonical Correlation Forests (CCFs) to learn the spectral signal of informal settlements from HR satellite imagery. The second method used a CNN combined with VHR satellite imagery to extract finer grained features. Peter and Gulnaz studied the use of object based change detection and object tracking of informal settlements using remotely sensing data in Cape town (South 10 Africa) between the period 2000 to 2015. The object change detection provided for detection of thematic changes per object and to document the changes of their property values and potential movement. In [52], the authors discuss and evaluate various machine learning approaches and a combination of various features to detect slums. The classifications methods evaluated included multi-class and hierarchical to distinguish urban from other classes such as vegetation and water. [55] presented the use of CNNs for the detection of informal settlements from VHR i.e. 0.60m Quickbird satellite image for the city of Dar es Salaam, Tanzania acquired in 2007 and compared the results with state of the art classifiers that use handcrafted features. [56] evaluated and pointed out that Object Based Change Detection (OBCD) and OBIA are the most promising techniques for automatized identification of new buildings in informal settlements, but they need to work with VHR data rectified with a DSM of the same or compatible spatial resolution. [57] This paper aimed to document and understand basic infrastructural conditions and their changes over an eight-year period for the Kibera Slum using VHR satellite images by analyzing the dynamics of physical transformations of building sizes and heights, built-up densities, and building arrangements. A pioneering study by Hoffman [58] used OBIA to identify informal settlements from IKONOS imagery in the City of Cape Town. Informal settlement classification was undertaken using sub-classes that described settlement forms (dense, medium, new and bright) based on complex hierarchy and class descriptions such as textural and spectral features. The author found that the ability to detect informal settlements was dependent on the spatial resolution of the imagery. No quantitative results were presented as the findings of the study. This research was later improved by Hofmann et al. [59], who showed that several modifications were required when applying extraction methods to a Quickbird scene in Brazil. The adaptions included simplified and pruned class-hierarchies to make the chosen class descriptors in theory more transferable to comparable scenes. The results of this study demonstrated that the selection of a strategy for informal settlement segmentation and classification is data and context-specific. Shekhar [60] delineated informal settlements in Pune City, India, using Quickbird imagery. The study highlighted the efficacy of the developed methodology to discriminate informal regions by describing typical characteristics of these settlements. Fuzzy membership function of texture, geometry, and contextual information were used to achieve an overall accuracy of more than 87%. Kohli et al. [30] expanded upon the work of Hofmann et al. [59] and developed Generic Slum Ontology (GSO), which can be used as part of a conceptual classification OBIA schema. 11 However, techniques like OBIA require VHR Data usually with a resolution of under 5 meters to detect slum objects and is a very time consuming approach and requires an adoption of parameters and values to a AOI. This makes them difficult to transfer to other cities or countries. It should be noted that most of the recent approaches to identify and map informal settlements involve the fusion of two or more approaches. 2.4. Hot Spot Clustering Analysis Hotspot analysis is a spatial analysis and mapping technique interested in the identification of clustering of spatial phenomena into statistically significant hot spots and cold spots [61].These spatial phenomena are depicted as points in a map and refer to locations of events or objects. A hotspot is an area that has higher concentration of events compared to the expected number given a random distribution of events [62]. Hotspot detection has evolved from the study of point distributions or spatial arrangements of points in a space [63] to understand spatial patterns in time. The application of hotspot analysis within public health, epidemiological research and crime mapping and research has increased significantly in the past couple of decades mainly due to advancement in GIS-based software. Hotspot analysis usually involves the incident count of points in a location, Attributes that further describes points or Period of time i.e. Date or time of events. Very few studies are based on the use of hotspot analysis in the mapping of Informal settlements because it requires mostly vector data which is quite insufficient for slums compared to raster data. Mohamed et. al introduced a method for identifying and predicting informal settlements using street intersections data using Getis-Ord Gi Hot Spot Analysis and a machine learning approach[20]. 2.5. An overview of Machine Learning Techniques Machine learning techniques (MLTs) (Figure 2) generate knowledge in a learning phase by means of training data and transfer this to new data for prediction and generalization even with noise-contaminated and incomplete data [58,59]. MLTs have been successfully used in analyzing and modeling various complex environmental disciplines, including in medicine, financial markets, ecology, geography, biomedicine, and epidemiology [66–69]. The growing popularity of MLTs can be attributed to their abilities to approximate almost any complex non-linear functional relationship [70–72]. Machine 12 Learning algorithms are based on both supervised and unsupervised learning. Supervised Learning works with labeled input data, known as training data in a training process for predicting unknown data. Based on the known data, the training process thereby continues until the model achieves a desired level of accuracy. Once the model is trained it can be used to transfer this knowledge on new data by generalizing from the training data to unknown examples. Supervised learning problems can be further grouped into classification and regression problems. On the other hand, for Unsupervised Learning, the input data is not labeled and does not have a known result. The model tries to recognize a pattern in the input data to learn about it and extract rules. This can be done through a mathematical process to systematically reduce redundancy or by organizing the data by similarity like through Clustering and Association. Figure 2: Machine Learning Cheat Sheet [73] Some of the most common ML algorithms in Geospatial Sciences and Mapping of Informal settlements are Logistical Regression, Decision Trees (DT), self-organizing maps (SOM), random forests (RF), support vector machines (SVM) and artificial neural networks (ANN). It’s important to note that selecting the best algorithm for a task is quite challenging since no algorithm outperforms others in a given task. Machine learning techniques have been and are still being used to carry out geospatial analysis in various fields for clustering, classification 13 and prediction various environmental and socioeconomic phenomena using different types of data. 2.5.1. Convolutional neural-networks Convolutional neural-networks (CNNs) are a branch of artificial neural networks designed to recognize objects in images based on their ability to develop an internal representation of a two-dimensional image [74]. The convolutional neural network is composed mainly of three types of layers i.e. Convolutional Layers, Pooling Layers, Fully-Connected Layers. The Convolutional layers are comprised of filters which are the neurons of the layer and a feature map which is the output of one filter applied to the previous layer. The pooling layers downsample the previous layers feature map by following a sequence of one or more convolutional layers with the intention to consolidate the features learned and expressed in them. They are used to reduce the size of the data while maintaining the most important features. Fully connected layers are the normal flat feed-forward neural network layer used at the end of the network after feature extraction and consolidation has been performed by the convolutional and pooling layers, to create final non-linear combinations of features and for making predictions by the network [75]. CNNs are well suited for the task of image classification in the field of remote sensing, However, CNNs also pose challenges when applied to remote sensing as it is often difficult or impractical to obtain a large set of labelled images, the model performs poorly and tends to over fit [74]. Such challenges can be curbed using techniques like Data augmentation using rotations and flips to increase the total number of training images to further avoid overfitting problem [33]. 20 Informal settlements, though most informal settlement buildings are smaller than this spatial resolution of 10m*10m. The images of the dates in Table 2 were selected based on perceived visually similarity, quality and minimal cloud cover. Acquisition Date Bands Used 01-01-2016 4,3,2 23-08-2017 4,3,2 31-12-2019 4,3,2 Table 2: Details of Sentinel-2 satellite images 3.2.3. Reference data The map selected to validate regions mapped out as informal regions in Kampala (Figure 7) was prepared by AcTogether (Ug) in 2014, does not include the slum settlements which have recently come up in the city. However, it provides the locations of the slum settlements manually delineated through Fields surveys and Public Participation. The slums in Kampala city are not clustered at specific locations but have scattered presence all over the city with varying sizes and shapes. The settlements are located along the main roads in elongated and irregular shapes, along railway lines and near the lake. Unfortunately, recent Informality maps for Dar-es Salaam were difficult to come by. Figure 7: Distribution of Informal Settlements in Kampala based on Average Household Size [97] 21 4. RESEARCH METHODOLOGY The method consists of three main phases (Figure 8). First phase consists of enriching the building outline features with geometric and topologic parameters, The second phase entails hotspot clustering and classification of these characteristics to identify informal settlements in similar context cities, and the final phase involves spatio-temporal monitoring of informal settlements using sentinel-2 imagery. Figure 8: Methodology Flowchart 4.1. Data Enrichment and Computation of Parameters Before carrying the clustering and classification process, geometric and topological measures to allow grouping of features into different clusters were determined. The measures chosen for enriching the building features were size, shape and shortest distance to a neighboring building. A building’s shape was described by the number of vertices it has i.e. vertex number (VN). More complex buildings have many vertices and vis versa. The size was described by calculating its area (A) [98,99]. The Nearest Neighbor Distance (NND) and was used to describe the topological relations to neighboring features [100]. Buildings in informal 22 settlements tend to be very close to each other with limited gap between them. They are also small in size and have simple shapes like for rectangles or squares houses. The normalization of computed variables was carried out since they have varying scales and to avoid influence of some variables over others during use in the machine learning algorithms. The Min-max normalization approach was used where all the values were mapped between [0–1]. 𝑿󰆒= (𝑿𝑿𝐦𝐢𝐧) (𝑿𝐦𝐚𝐱𝑿𝐦𝐢𝐧) (Min-Max Normalization Equation) Where, X is an original value, X’ is the normalized value, and Xmin, Xmax are the minimum and maximum values of this particular property. 4.2. Informal Settlement Buildings Clustering and Classification In the first step, clusters are determined based on similarity of building feature characteristics using hotspot clustering analysis. The clusters are then reclassified to reduce the number of clusters. If a building belongs to the same cluster based on the features characteristics, its binned in one cluster and vis versa. This is done on the Kampala training partition. The second step involves using different classifiers to select one that suits the training partition data. The selected classifier is them used to determine informality regions in the Kampala Testing Partition and Dar-es-Salaam dataset. 4.2.1. Clustering of Informal Settlement Buildings As shown in chapter 2, informal settlement areas have building features characterised by simple shapes, small buildings and have limited gaps between them. In order to determine the pattern of spatial distribution and where low and high values of the building feature characteristics are grouped, hotspot analysis based on the use of Getis-Ord Gi* statistic was carried out on the Kampala training partition dataset to identify statistically significant hot and cold spots [61]. Since hotspot analysis requires the presence of clustering, there is need to test for the presence of clustering in the dataset by assessing spatial autocorrelation to identify clustering within the entire dataset. The Spatial Autocorrelation (Moran’s) tool in ArcGIS measures spatial autocorrelation by simultaneously measuring feature locations and attribute values. If features that are close together have similar values, 23 then that is said to be clustering, and vis versa. The Moran’s I returns values which include the z-score and p-value which will indicate if clustering is found in the data or not. The Optimized Hotspot Analysis (OHSA) tool from ArcGIS Pro used for the hotspot analysis is identifies statistically significant hot and cold spots for multiple testing and spatial dependence using the using the False Discovery Rate (FDR) correction method [101]. The features are grouped into seven major clusters i.e. features in the +/- 3 Gi_Bins were statistically significant at the 99% confidence level; features with 0 for the Gi_Bin field was not statistically significant, those in the +/2 bins reflected a 95% confidence level, and features in the +/1 bins reflected a 90% confidence level. Upon computing these GiZscores where cold spots in all the feature characteristics most likely represent informal settlement buildings, a cross validation test using Independent samples t-test on the hotspot analysis results in order to differentiate between formal and informal zones and to determine whether there is statistical evidence that the associated group means are significantly different between them [102]. The GiZscores for both groups; informal, and formal regions are compared and when the p-value is less than 0.001, the null hypothesis can be rejected and the results of GiZscore can be illustrated as statistically significant [11]. The features identified as cold spots were then reclassified and grouped into one class and all the others as another class, which produces new reclassified results for the feature characteristics of Size, shape and NND. The three reclassified results are combined and if a feature has all its feature characteristics as cold spots, its grouped as informal and if not, its clustered as formal. This produces one resultant cluster dataset for the building features which shows both informality and formality in the Kampala training partition dataset. These results (especially the cold spots) were compared to areas delineated as slums in Kampala generated by a collaboration between Slum Dwellers International (SDI) and other organizations [46, 89]. 4.2.2. Classification of Informal Settlements based on Building Characteristics To predict where informal settlement regions are based on building features characteristics of shape, size and NND, based on the clusters from the training partition dataset, supervised machine learning classification was carried out on data of the same city (Kampala Testing Partition) and on data of a different city (Dar-es Salaam). Firstly, the clustering results were 24 applied to train, validate and test several classifier algorithms in order to select one that suits these datasets. Models were built using the python Scikit-learn library, taking advantage of the capabilities and features of this library. These models included Logistical Regression, Decision Tree (DT), Multilayer Perceptron (MLP), Random Forest (RF) and Support Vector Machine (SVM). The choice of these classifiers in based on their use in previous studies to classifier buildings based on their characteristics and also their ability to provide good accuracy with reference to chapter 2. The script involved splitting the Training partition dataset into 70% for training and 30% validation. Classification was performed first on the training data and accuracy measures, including overall accuracy, precision, recall and f-measure were noted. This was carried out in order to select the best classifier for the datasets. The selected classifier is then used to predict which features belong to Informal settlements in Kampala Testing Partition and Dar-es Salaam. 4.3. Monitoring the growth of Informal Settlements The sentinel2 imagery after pre-processing and band composition were clipped to the Kampala Boundary. The acquired informal settlement regions from the classification phase above provides for polygons representing informality. We also assembled polygons for regions representing three categories: “built (without informal)”, “vegetation”, and “water”. All these regions were generated on the 2019 Image (Figure 9). This is based on the assumption that the OSM building outlines used to classify informality were as up-to-date as of October 2019. All the images were then sliced up into equal-sized tiles of 255*255 m to ensure uniformity in the datasets and placed in folders of “Tiles_2019”, “Tiles_2017” and “Tiles_2016” each with 5427 tiles. The tiles were saved as tiff files with attached geo-information which is required to plot the model predictions to the original locations in the map. 25 Figure 9: Informal and Formal Tiles The tiles in folder Tile_ 2019 were labelled based on the four categories created based on their intersection with the created polygons. All the tiles belonging to “built-up”, “vegetation”, and “water” were grouped into one folder as Formal and the Informal tiles placed in an Informal Folder. Regrouping the four categories into two was more accurate than training a binary classifier directly from the available data. The tiles were then split into training (80%) and testing sets (20%) for model-validation for use in the CNN model. A baseline CNN model based on general architectural principles of the Visual Geometry Group (VGG) on which other models can be compared is established. The CNN architecture involved stacking convolutional layers with small 3×3 filters followed by a max pooling layer. These layers form a block, and these blocks can be repeated where the number of filters in each block is increased with the depth of the network such as 32, 64, 128, 256 for the first four blocks of the model. Padding is used on the convolutional layers to ensure the height and width shapes of the output feature maps matches the inputs. This function can then be customized to define different baseline models, e.g. versions of the model with 1, 2, or 3 VGG style blocks. The model was fitted with stochastic gradient descent, a conservative learning rate of 0.001 and a momentum of 0.9. Since the problem was a binary classification task i.e Formal and Informal, 26 the prediction of one value of either 0 or 1 was required. An output layer with 1 node and a sigmoid activation was used and the model optimized using the binary cross-entropy loss function. One-block (32 filters), two-block (32 and 64 filters) and three-block (32,64 and 128 filters) VGG models were all compared on the dataset. Two approaches to address overfitting of the training dataset: dropout regularization and data augmentation were explored to select the best suited one. Upon selecting the best model with the adequate model parameters, Predictions were performed on the Independent tile datasets for “Tiles_2017” and “Tiles_2016” to automatically predict informal settlement tiles out of all the tiles. A polygon layer was then built for the new informality tiles. This prediction step leads to tiles that are either formal or not. Polygons of the image tile Boundaries are created and those that belong to informal group are overlaid onto the map for visual interpretation. 27 5. RESULTS This chapter describes the implementation details of the methodology and presents the results of this study. The source code of this implementation can be found at https://github.com/Bakked9/MastersThesis. 5.1. Clustering Analysis 5.1.1. Clustering Indices for Building Characteristics The Moran's Indices for the Area, shape and NND are all close to 0.2, and have p-values of 0.0 (Table 3). The z-score is a standard deviation which measures how many SDs away an element is from the mean, while the p-value is a measure of probability that a random process created the observed pattern. A small p-value is an indicator that the spatial pattern is not random and vis versa. A positive Moran’s I result indicates that neighboring areas or points are similar with respect to attribute values, while negative Moran’s I shows that nearby regions are less similar in attributes than one would expect in a random pattern [103]. This goes to show that clustering exists within the dataset with respect to the measurement variables and hence Hotspot analysis can be carried out on the data. Variable Moran's Index Expected Index Variance z-score p-value Nearest Distance 0.257656 -0.000017 0.000139 21.848243 0.000000 Area 0.278887 -0.000017 0.000139 23.696816 0.000000 Shape 1.336825 -0.000017 0.000139 113.285603 0.000000 Table 3: Spatial Autocorrelation (Moran's I) Results 5.1.2. Optimized Hotspot Analysis The optimized hotspot analysis resulted into Gi_Bin fields which identified the statistically significant hotspots, along with the non-significant and cold spots, indicating the type of clusters for the feature characteristics. In the case of the Training partition, the Gi_Bin field showed that patches spread around the selected area was characterized by cold spots, nonsignificant types, and hotspots types that are distributed allover (Figure 10). The areas identified as non-significant cluster types provide some sort of boundary between the cold and 28 hot spots. Whereas the cold spots were located in a small area mainly on the north and central parts, the hotspot clusters were spread out in on the north, south, east and west sides. Figure 10: OHSA Results Maps for Size, Shape and NND 5.1.3. The Association between the Cold spots and Informal settlements Table 4 illustrates the results of the t-test analysis, using a 95% confidence level for each feature characteristic. Since the Sig (2-Tailed) values i.e. the p-value for the GiZscore and Nneighbours predictors are very small i.e. p<0.001, we reject the null hypothesis of Levene's test and conclude that the variances in the GiZscore and Nneighbours predictors of Formal is significantly different from that of Informal. The Informal groups mean numbers of neighbors are larger than those of the formal areas. The mean GiZscores are the informal areas are also smaller than those of the formal areas. This shows that the informal regions are more likely to be cold spots with low values of Shape, size and NND. 29 SIZE(AREA) Group N Mean Std. Deviation Std. Error Mean Sig.(2 - tailed) GiZScore ColdSpot(Informal) 104884 -4.4855 2.0227 0.0062 0.000 HotSpot (Formal) 36868 3.6859 1.8595 0.0097 Nneighbors ColdSpot(Informal) 104884 958.00 302.407 .934 0.000 HotSpot (Formal) 36868 363.13 129.434 .674 NEAREST NEIGHBOUR DISTANCE (NND) Group N Mean Std. Deviation Std. Error Mean Sig.(2 - tailed) GiZScore ColdSpot(Informal) 123467 -5.6911 2.8813 0.0082 0.000 HotSpot (Formal) 41942 5.1848 3.3074 0.0161 Nneighbors ColdSpot(Informal) 123467 936.37 281.848 .802 0.000 HotSpot (Formal) 41942 344.97 119.569 .584 SHAPE (VERTEX NUMBER) Group N Mean Std. Deviation Std. Error Mean Sig.(2 - tailed) GiZScore ColdSpot(Informal) 129377 -4.8803 1.9626 0.0055 0.000 HotSpot (Formal) 34552 5.8729 4.6398 0.0250 Nneighbors ColdSpot(Informal) 129377 850.26 319.712 .889 0.000 HotSpot (Formal) 34552 512.44 286.337 1.540 Table 4: Independent samples t-test analysis Results Cold spots in this analysis represent areas that have low values of Area, Shape and NND. Informal settlements are characterised by buildings that have simple shape, small is size and are densely located in a particular location i.e. short NNDs between polygons (see figure 10). With this, the cold spots resulting from the OHSA are considered to be locations with informality based one the building typologies used to characterise them. To simplify the data for further analysis, all the OHSA results were classified to bin the data points into two classes i.e. “Informal” and “Formal”. This was done through reclassification of the Gi_bin results with a threshold of -2 or lower (95% cold spot confidence level) as Informality and the rest as formality. As shown in figure 11 and 12, the regions that identified as Informal share common locations across the map with respect to the building typologies. The results were combined to highlight only those areas that had been classified as informality in all the typologies. The resultant dataset was then used as a training set for the Classification step. 36 Figure 16: Change in Informal Settlement Regions in Kampala. This shows Informal tiles that existed in January 2016, new informal tiles that had developed by August 2017 and also by December 2019. 37 6. DISCUSSION 6.1. Discussion This study presented approaches to classify informality regions based on building outline characteristics and monitoring growth of informal settlements over time by executing CNNs on High resolution satellite imagery. Firstly, we highlight that OSM building feature characteristics i.e. size and shape of building outline, and distance to nearest building can be used to differentiate between formal and informal buildings. In the approach, the geometric and topologic measures presented in this research were clustered based on the values high or low. As an important aspect of the evaluation, the transferability of the obtained clusters from training set to test data was investigated. Firstly, the results were applied to the testing data of the same city and then to another city. 108,390 of 429,270 buildings (25.25%) were classified as informal settlement buildings in Kampala while 188,728 of 971,008 buildings (19.44%) in Dar-es Salaam were classified as informal settlement buildings (Figure 13). There was a decrease in Random Forest classifier model performance on the Kampala Test Partition and Dar-es Salaam may be due to differences in the characteristics of buildings in very urban areas, or the fact that more data were available in Dar-es -Salaam. It’s important to note that the accuracy values were consistently above 0.75, suggesting very good discriminatory performance irrespective of the city. The variation in building typologies used during this analysis (shape, size and spacing) played an important role during the classification of informality and formality even though these characteristics vary according to geographical location [93]. Though the morphological nature of informal settlements varies across the global, they still bare some similarities in characteristics. This could be the reason why the same model could be applied in a different city and produce distinct results. Compared to its areal coverage, there are fewer informal settlement buildings in Dar es Salaam. Over the years, a decrease in the population living in Slums in Tanzania has been decreasing with a sharp decrease observed from 2011 even with rapid population growth over time (Figure 17). This is probably due to strategies by the Tanzania government to shift from regularly demolished homes in informal settlements to awarding titles could be one of the reasons for a decrease in informality in the city. The MKURABITA Programme first launched in Dar es Salaam, aims at transforming property and businesses held in the informal sector into legal entities that are rooted firmly in the formal sector [104]. 38 Figure 17: Tanzania - Population Living In Slums (% Of Urban Population) [[105]] The experimentations of Using CNN on Sentinel-2 evaluates shows promising results on discriminating Informal from formal settlements over time. However, Seasonal variations between the images prevented the deep learning model from recognizing certain Informal Settlements. Testing the model on new multi-temporal datasets causes the model performance to decrease as we move away from the training set’s sensing time. The performance of the model is bound to degrade as we move further away from the training period. Upon comparing the development of the slums in Kampala city using the time slider in Google Earth from the year 2014 to the dates at which satellite images are downloaded, change observations can be made. It can be seen that since 2014, the built density of slums has increased at pre-existing locations and smaller new patches have emerged at suburban areas of the city. Compared to the larger areas of already existing slums, these new settlements are quite smaller and hence could be ignored during prediction on the image tiles due to mixed classes with vegetation and formal areas. To evaluate the results obtained, the prediction Accuracy obtained from the Baseline VGG3 + Data Augmentation CNN Model (90.3%) on the sentinel-2 imagery is a very good results based on the spatial resolution of the images. In [106], Federico .B and Damián .S also detect informal settlements using sentinel-2A image tiles. They employed the use of Remote Sensing Classifier to Predict tiles that belong to informal settlements. They recommend the use of deep learning techniques, like convolutional neural networks to improve the accuracy of the classification. This indicates that the CNN provides better classification accuracy especially 39 because it’s an image based classifier. On another note, the use of building outline characteristics like size, shape and Nearest Neighbour distance provides for a lot more insight on the nature and morphology of an informal settlement. Investigations on predicting informal settlements using road intersections from OSM data [20] showed that OSM data can be used in this scenario. However, some of the limitations to this study is that some informal settlements have roads that are not mapped on OSM which makes their detection difficult. This is where the use of Building feature characteristics, as shown in this research are useful. The characteristics help better delineate informal settlements. An even better alternative would be the combination of building feature characteristics and density of road intersections to identify these informal settlements. The use of Multilayer Perceptron (MLP) on hotspot analysis GiZscore and Nneighbours to predict informal settlements proved a good enough classifier, however in this research, though MLP performed well, Random Forest Classifier was a better classifier for Building feature characteristics to predict informality. 6.2. Limitations While the study produced good research findings, there are several limitations and opportunities for further research Additional vector data of land parcels, small administrative district units, city s functional zones, or land use data could be integrated to improve the creation of desired optimal shapes of image objects of compact homogeneous built-up areas for built-up density analysis. If available in accurate up-to-date form, vector layer of building footprints could be used for built-up density calculation, without the need of land cover classification and building extraction from the satellite image data. This layer would represent only buildings, without roads or other man-made urban structures. However, this data is not always freely available in accurate and up-to-date form, and this approach takes that into consideration. The lack of adequate data to test the model on other cities besides the two presented was experienced. This model worked for these two cities because there both that OSM data almost accurately available. This is not the case for most of the cities in the global south that are engulfed by informality. 40 7. CONCLUSION AND FUTURE WORKS The main research Aim of this research work was to integrate OSM data and sentinel-2 imagery for classifying and monitoring the growth of informal settlements. To achieve this, the research questions were investigated: 1. How can we exploit the potential of using buildings outline characteristics such as size and shape to differentiate Informal settlements from Formal Settlements? The results from the hotspot clustering analysis demonstrated that informal settlements can be differentiated from formal areas based on the use of Building feature characteristics that is shape, size and NND. The informal settlement buildings are characterized by low values (cold spots) for each characteristic. 2. Is it possible to predict informal areas in a city by understanding housing informality in other cities of similar context using buildings outline characteristics and machine learning? Also its observed that though morphologically different, slums share some characteristics like the nature of building outlines. This is shown by the transferability of the classifier trained on one dataset to predict informal settlements in new datasets in the same city and in a different city. 3. What is the most appropriate Machine Learning technique based on accuracy to predict informal areas in a city based on buildings outline characteristics? The most appropriate ML technique is the Random Forest classifier. It proved its effectiveness in a broad range of applications of Geospatial technologies. As a nonparametric method, the key property of Random Forest is its capability to handle different statistical distributions of features, which was one of the main challenges of this study. This make Random Forest a suitable algorithm for our binary classification and prediction problem. 4. How can we exploit the potential of freely available Sentinel-2 satellite imagery with advanced machine learning to estimate the growth of Informal settlements? 41 Sentinel 2 images, despite having a spatial resolution of 10m, can be used for informal settlement mapping when coupled with Advanced Machine Learning Algorithms like CNNs. We suggest that this research area could be explored more and future research could be done, building and improving on the results with emphasis on a better accurate representation of Informal settlements. Furthermore, other OSM features and parameters such as Roads and accessibility within the informal settlements or others could be used to better identify where the informality regions could be. This is because Informal settlements are usually characterized by narrow and short road segments this numerous dangles ant the end of the roads. Other parameters like proximity to hazardous areas like wetland and sewerage channels should be considered. This however, would however require much more detailed and complex land cover classification scheme as a source of information for this analysis, including identification of, vegetation types, difference between agricultural bare soil and bare soil areas in urban area or a construction sites and so on. If available, Digital Surface Model (DSM) could be used to more accurately identify and map informal settlements because inclusion of the third dimension of Elevation would help highlight more distinguishing characteristics of informal settlement buildings. Nevertheless, the contribution of the temporal dimension to informal settlement mapping investigated in this research requires further testing with larger datasets and longer time series. With this mind, we think the positive results achieved in this research are worth building upon in future studies to further investigate the temporal domain as an input to informal settlement mapping and classification models. 42 BIBLIOGRAPHIC REFERENCES 1. Africa’s Urban Transformation - Essays | Urban Age [Internet]. [cited 2020 Feb 6]. Available from: https://urbanage.lsecities.net/essays/africa-s-urban-transformation 2. 68% of the world population projected to live in urban areas by 2050, says UN | UN DESA | United Nations Department of Economic and Social Affairs [Internet]. [cited 2020 Feb 6]. Available from: https://www.un.org/development/desa/en/news/population/2018-revision-of-worldurbanization-prospects.html 3. Dar es Salaam is growing rapidly—and its planners are struggling to keep up [Internet]. [cited 2020 Feb 6]. Available from: https://www.nationalgeographic.com/environment/2019/04/tanzanian-city-may-soonbe-one-of-the-worlds-most-populous/ 4. About Slum Upgrading | Cities Alliance [Internet]. [cited 2020 Feb 6]. Available from: https://www.citiesalliance.org/about-slum-upgrading 5. Introduction to urban poverty | International Institute for Environment and Development [Internet]. [cited 2020 Feb 6]. Available from: https://www.iied.org/introduction-urban-poverty 6. Roy D, Lees MH, Palavalli B, Pfeffer K, Sloot MAP. The emergence of slums: A contemporary view on simulation models. Environ Model Softw [Internet]. 2014;59(2014):76–90. Available from: http://dx.doi.org/10.1016/j.envsoft.2014.05.004 7. File:Aerial view of Makoko Slum in Lagos Nigeria.png - Wikimedia Commons [Internet]. [cited 2020 Feb 6]. Available from: https://commons.wikimedia.org/wiki/File:Aerial_view_of_Makoko_Slum_in_Lagos_ Nigeria.png 8. The 20 Worst Slums in Africa [Internet]. [cited 2020 Feb 6]. Available from: http://www.africaranking.com/20-worst-slums-in-africa/6 9. Mahabir R, Agouris P, Stefanidis A, Croitoru A, Crooks AT. Detecting and mapping slums using open data: a case study in Kenya. Int J Digit Earth [Internet]. 2018;0(0):1– 25. Available from: https://doi.org/10.1080/17538947.2018.1554010 10. Suhartini N, Jones P. Urban Governance and Informal Settlements [Internet]. 2019. 235 p. Available from: https://www.springer.com/gp/book/9783030060930 43 11. Hofmann P, Taubenböck H, Werthmann C. Monitoring and modelling of informal settlements - A review on recent developments and challenges. 2015 Jt Urban Remote Sens Event, JURSE 2015. 2015;1–4. 12. Arimah BC. Slums as expression of social exclusion: Explaining the prevalence of slums in African countries. United Nations Hum Settlements Program [Internet]. 2001;1–33. Available from: http://www.oecd.org/dev/pgd/46837274.pdf 13. Kohli D, Sliuzas R, Stein A. Urban slum detection using texture and spatial metrics derived from satellite imagery. J Spat Sci [Internet]. 2016;61(2):405–26. Available from: http://dx.doi.org/10.1080/14498596.2016.1138247 14. Kohli D, Stein A, Sliuzas R. Uncertainty analysis for image interpretations of urban slums. Comput Environ Urban Syst [Internet]. 2016;60:37–49. Available from: http://dx.doi.org/10.1016/j.compenvurbsys.2016.07.010 15. Wurm M, Taubenböck H, Weigand M, Schmitt A. Slum mapping in polarimetric SAR data using spatial features. Vol. 194, Remote Sensing of Environment. 2017. 190–204 p. 16. Stasolla M, Gamba P. Exploiting Spatial Patterns for Informal Settlement Detection in Arid Environments. Int Arch Photogramm Remote Sens Spat Inf Sci. 2007;36:31–6. 17. Ranguelova E, Weel B, Roy D, Kuffer M, Pfeffer K, Lees M. Image based classification of slums, built-up and non-built-up areas in Kalyan and Bangalore, India. Eur J Remote Sens [Internet]. 2019;52(sup1):40–61. Available from: https://doi.org/10.1080/22797254.2018.1535838 18. Detecting Informal Settlements using Satellite Imagery and Convolutional Neural Networks [Internet]. [cited 2020 Feb 6]. Available from: https://blog.goodaudience.com/detecting-informal-settlements-using-satellite-imageryand-convolutional-neural-networks-d571a819bf44 19. Jean N, Xie M, Ermon S, Korea N. Combining Satellite Imagery and Machine Learning to Predict Poverty Inputs : Day # me satellite imagery Predic < ons : Economic indicators Learned features corresponding to buildings ,. 2015;4:2015. 20. Ibrahim MR, Titheridge H, Cheng T, Haworth J. predictSLUMS: A new model for identifying and predicting informal settlements and slums in cities from street intersections using machine learning. Comput Environ Urban Syst. 2019;76:31–56. 21. Amit Patel, Andrew Crooks and NK. Spatial Agent-based Modeling to Explore Slum 44 Formation Dynamics in Ahmedabad, India. 2018;37–54. Available from: http://link.springer.com/10.1007/978-3-319-59511-5 22. Crooks AT. Constructing and implementing an agent-based model of residential segregation through vector GIS. Int J Geogr Inf Sci. 2010;24(5):661–75. 23. United Nations Human Settlements Programme. The Challenge of Slums. The Challenge of Slums. 2012. 24. Ghani E. Urbanization and ( In ) Formalization. Urban Poverty Reduct Conf 2013 Bridg Rural Urban Perspect. 2013; 25. Bähr J, Mertins G. Marginalviertel in grossstadten der dritten welt. Geogr Rundsch. 2000; 26. The continuing challenge of informal settlements An introduction - Technische Informationsbibliothek (TIB) [Internet]. [cited 2020 Feb 6]. Available from: https://www.tib.eu/en/search/id/BLCP%3ACN063768801/The-continuing-challengeof-informal-settlements/ 27. Taubenböck H, Kraff NJ, Wurm M. The morphology of the Arrival City - A global categorization based on literature surveys and remotely sensed data. Appl Geogr. 2018;92(January):150–67. 28. Davis M. Planet of slums. New Left Review. 2004. 5–34 p. 29. Jamieson J. Shadow Cities: A Billion Squatters, a New Urban World. J Soc Polit Econ Stud. 2006; 30. Kohli D, Sliuzas R, Kerle N, Stein A. An ontology of slums for image-based classification. Comput Environ Urban Syst [Internet]. 2012;36(2):154–63. Available from: http://dx.doi.org/10.1016/j.compenvurbsys.2011.11.001 31. Kohli D. Identification and Characterization of Informal Settlements Using Satellite Images in Support of Land Administration. 32. Jokar Arsanjani J, Zipf A, Mooney P, Helbich M. OpenStreetMap in GIScience. OpenStreetMap GIScience Exp Res Appl. 2015; 33. Gervasoni L, Fenet S, Perrier R, Sturm P. Convolutional neural networks for disaggregated population mapping using open data. In: Proceedings - 2018 IEEE 5th International Conference on Data Science and Advanced Analytics, DSAA 2018. 2019. 34. Grippa T, Georganos S, Zarougui S, Bognounou P, Diboulo E, Forget Y, et al. 45 Mapping Urban Land Use at Street Block Level Using OpenStreetMap, Remote Sensing Data, and Spatial Metrics. ISPRS Int J Geo-Information. 2018; 35. Mahabir R, Croitoru A, Crooks A, Agouris P, Stefanidis A. News coverage, digital activism, and geographical saliency: A case study of refugee camps and volunteered geographical information. PLoS One. 2018; 36. Barrington-Leigh C, Millard-Ball A. The world’s user-generated road map is more than 80% complete. PLoS One. 2017; 37. Fonte: Using openstreetmap to create land use and... - Google Scholar [Internet]. [cited 2020 Feb 6]. Available from: https://scholar.google.com/scholar_lookup?title=Using+Openstreetmap+to+Create+La nd+Use+and+Land+Cover+Maps:+Development+of+an+Application&author=Fonte, +C.C.&author=Patriarca,+J.A.&author=Minghini,+M.&author=Antoniou,+V.&author =See,+L.&author=Brovelli,+M.A.&publication_year=2019&pages=1100–1123 38. OpenStreetMap: User-Generated Street Maps - IEEE Journals & Magazine [Internet]. [cited 2020 Feb 6]. Available from: https://ieeexplore.ieee.org/abstract/document/4653466 39. Johnson BA, Iizuka K. Integrating OpenStreetMap crowdsourced data and Landsat time-series imagery for rapid land use/land cover (LULC) mapping: Case study of the Laguna de Bay area of the Philippines. Appl Geogr [Internet]. 2016;67:140–9. Available from: http://dx.doi.org/10.1016/j.apgeog.2015.12.006 40. Yang D, Fu CS, Smith AC, Yu Q. Open land-use map: a regional land-use mapping strategy for incorporating OpenStreetMap with earth observations. Geo-Spatial Inf Sci [Internet]. 2017;20(3):269–81. Available from: http://doi.org/10.1080/10095020.2017.1371385 41. Kang L, Wang Q, Yan HW. Building extraction based on OpenStreetMap tags and very high spatial resolution image in urban area. Int Arch Photogramm Remote Sens Spat Inf Sci - ISPRS Arch. 2018;42(3):715–8. 42. Luo N, Wan T, Hao H, Lu Q. Fusing high-spatial-resolution remotely sensed imagery and OpenStreetMap data for land cover classification over urban areas. Remote Sens. 2019;11(1). 43. Huang H, Kieler B, Sester M. Urban building usage labeling by geometric and context analyses of the footprint data. Proceeding 26th Int Cartogr Conf. 2013; 52 8. ANNEXES 8.1. Independent Samples T-Tests On Hotspot Analysis GiZScores and Nneighbors 53 8.2. Plots for Accuracy and Loss for Some CNN Models - Train (blue) and Test (orange) datasets. 8.2.1. 1-VGG 8.2.2. 3-VGG 54 8.2.3. Baseline VGG3 + Data Augmentation 8.2.4. Pre-Trained VGG16 55 8.3. Informality Regions for Kampala in 2016 (top left), 2017 (top right) and 2019 (bottom) 01-01-2016 23-08-2017 31-12-2019 56