Peri-Urban Growth into Natural Hazard-Prone Areas. Mapping Exposure Transformation of the Built Environment in Nairobi and Nyeri, Kenya from 1948 to Today

Kenya experiences massive urban growth, also into natural hazard-prone areas, exposing settlements and the natural environment to riverine and pluvial oods and other natural hazards. While Nairobi as the capital and principal city has been extensively analysed regarding urban growth and ood hazard in some central parts, awareness of growing peri-urban areas has not been studied as much. The results are of interest to other locations in Kenya and worldwide, too, since the current research and disaster risk practice focus is still too much on megacities and city centres. Therefore, the study compares urban growth into hazard areas in urban rims of Nairobi and Nyeri, Kenya. A change assessment from 1948 to 2020 is conducted by aerial images, declassied satellite images, and recent data. Urban growth rates are 10 to 20-fold, while growth into ood exposed areas ranges from 3 to 100-fold. This study reveals unused opportunities for expanding existing land-use change analysis back to the 1940s in data-scarce environments.


Introduction
Demands for risk monitoring and risk-informed decision-making are expressed by the Sendai Framework (United Nations 2015). Similarly, the needs for integrative assessments, including basic infrastructure, urban features and disaster risk, are expressed in the New Urban Agenda (UN/HABITAT 2017). However, while disaster risk studies have covered major cities and national capitals, studies on smaller cities or peri-urban areas are lacking widely (Birkmann et al. 2016).
Peri-urban areas are often among the fastest-growing urban areas when close to a major city or capital like Nairobi. Urban growth also expands the ratio of built-up areas potentially exposed to natural hazards such as oods or droughts.
Monitoring of disaster risk needs to capture static pictures of risk and include the dynamics to better understand risk and resilience development and identify pathways to govern urban areas towards better resilience in coping with shocks and disasters (Simonovic and Peck 2013). This notion of the importance of dynamic risk assessments is expressed by recent research on aspects of the transformation of cities (Pelling et al. 2015), pathways in coping with climate change (Gibson et al. 2016), or transition (Solecki et al. 2017). However, this leads to a demand for data available for long periods to analyse changes.
Limitations of static risk assessments have long been acknowledged (King 2001), but few studies provide change assessments of disaster risk. This is often due to the lack of data for comparison or awareness about existing sources of data. While this is a challenge existing worldwide, availability and access to data are a major constraint for many countries in Africa (Osuteye et al. 2017). It seems paramount to improve data usability for risk assessments, and this paper points to data sources not used widely yet.
There is a wealth of studies utilising remote sensing data and urban area classi cation approaches (Bhatta 2010; Patino and Duque 2013). Multi-temporal and time-series analyses have increasingly been applied in this area of urban and landuse mapping (Acevedo and Masuoka 1997;Li et al. 2018), or peri-urban growth (Shaw and Das 2018). Mapping transformation using remote sensing increasingly starts applying multi-spatial-resolution change detection (Zhang et al. 2016). However, many approaches in remote sensing studies focus on the technical advancements and are less applied for the operational mapping of urban transformation yet. And while urban growth studies are common (Bagan and Yamagata 2012; Masek et al. 2000;Song et al. 2016), fewer studies combine it with a spatial detection of exposure to natural hazards, especially over a range over more than 20-30 years in comparison. Even fewer studies have used satellite data before the 1970s, such as the declassi ed satellite data of the USA (Day et al. 1998). The USA has disclosed declassi ed former espionage imagery from the 1960s onwards that are hardly used for risk assessments so far (Fekete 2020). Declassi ed satellite data from the CORONA, GAMBIT or HEXAGON series have been applied rather selectively and not very often yet, to elds such as archaeology (Ur 2003 There is an increasing demand for multi-modal, multi-spectral and multi-temporal remote sensing approaches (Brito and Quintanilha 2012;Qiming 2011;Stiller et al. 2021), and making use of a combination of aerial and satellite images can offer opportunities for that.
Aerial imagery and old declassi ed former espionage reconnaissance satellite imagery are major data sources to capture previous land cover back to the 1940s. Kenya has been covered by UK Royal Airforce aerial imagery campaigns due to the former colonial history. Access to such data is still restricted; it is on data servers not known much outside remote sensing communities. A lot of this data is open access, but sometimes, fees have to be paid, and although they are relatively low, it can still mean a restriction for researchers or local institutions. Another data access challenge is nding hazard data. This is the case for industrialised countries as well as for many other countries too.
Geospatial assessments are used in Kenya for monitoring land-use change or property rights (Koeva et al. 2017). Aerial imagery is also used for monitoring the sustainable development of agriculture (Ekbom et al. 2001). Challenges in de ning urban boundaries and overestimating urban areas by single criteria such as population density are analysed using aerial imagery (Potts 2017). Hazards analysed using aerial images cover many hazards, including erosion and contamination, but also deal with lake basin (Khan et al. 2011;Habib et al. 2009)  Therefore, the objective of this paper is to expand existing land cover change assessments back to the 1940s using aerial imagery. This enables capturing urban growth outside the national capital, Nairobi, along urban rims and in expanding towns in Kenya. A transferrable mapping approach is demonstrated, and it is indicated which natural hazard data could be used to analyse urban growth into potential hazard zones over time. Natural hazards include river oods in this case, and exposure of settlement area to be ooded is the main focus here. Using heterogeneous remote sensing data from different satellite missions and with different sensors and resolutions, a temporal comparison is made possible that exceeds previous approaches. In addition, this approach can analyse data-poor areas and areas not covered yet by research on natural hazards and exposure, which may inspire similar research in other areas.

Method
Geospatial assessments are useful to cover large areas of land by one comparable data source. It needs to be coupled with ground-truthing and thereby often is used in combination with other modes of assessment, such as local eld visits, statistical data or others. However, geospatial data provides another form of multimodal characterisation, which is the special issue's focus. This case is a multi-modal selection of greyscale aerial photos, satellite-based greyscale imagery, multi-spectral recent satellite data, statistical data, and on-the-ground photos and eld visits. Geospatial data is bene cial for analysing changes over time, using the same data source, or in combination with other modes of satellite, aerial remote sensing, ground-based imagery or many other forms of data integrated into a geographic information system (GIS).

Case study areas
The case study areas were selected in a co-design approach with the local experts and co-authors in joint workshops in Kenya on Sept 5 2019, and online, 17 Feb 2021. The two areas have been selected for demonstrating and testing the data availability and GIS-assessment opportunities of a combination of old aerial, former satellite espionage and recent satellite and OSM data. One area is the peri-urban rim of Nairobi, the nation's capital and largest city (Fig. 1). The area selected in the North-East is where urban sprawl and growth occur to a great extent; it is Ruiru town in Nairobi East along the Thika Road. The other area selected also experiences urban growth, the city of Nyeri, located North of Nairobi, close to the Aberdare and Mt. Kenya mountain ranges. Both case study locations ( Fig. 1) experience major urban growth but are prone to natural hazards such as oods and droughts.

GIS Analysis
The analysis is a GIS assessment using remote sensing data in combination with statistical data. QGIS is used as the GIS software since it is cost-free, open-source and enables the usage with different computer operating systems; so, it can be used by partners in Kenya and Germany and other countries, in teaching as well as in further research. The projection used is WGS84 UTM Zone 37N, so metric analyses can be conducted.
First, urban growth is analysed by change detection using aerial images from 1948, declassi ed CORONA mission (Day, et al. 1998), related data from 1967 and 1977 and Sentinel 2 data from 2020. The earliest available data and su cient spatial resolution, coverage of the case study area and cloud coverage, and costs were selection criteria for the data (Table 1 provides details on remote sensing data ordered and used). The aerial images were ordered from the National Collection of Aerial Photography from the UK. The declassi ed CORONA, HEXAGON and GAMBIT mission satellite images were obtained from the USGS Earth Explorer portal; three images were already available for free download. One additional image had been ordered. The OSM and Sentinel-2 data were openly available on OSM and EOX platforms. The aerial photos and CORONA satellite images were georecti ed based on OSM and Sentinel-2 data as base maps. Urban features were manually extracted and digitised from the 1948 aerial photos and the CORONA satellite images at a viewing scale of 1:2.500 to 1:5.000 in QGIS. Sentinel-2 data were digitised at spatial scales between 1:5000 and 1:10.000.
To capture built-up areas, mainly houses were mapped since they were the easiest to identify by rectangular shapes as distinct urban areas. Houses are identi ed mainly by the rectangular shape of the roofs. In addition, roads and paved (/impervious/sealed) areas were mapped when connected to buildings but not gardens or other partly natural surfaces. Special care was taken along rivers to capture houses and urban features. Vegetation land cover was mapped especially for human-shaped irrigated land use when it was identi able by rectangular features such as roads or borders to other vegetation.
For training purposes, 100x100 m grids were created. An urban area is de ned when adjacent urban features cover an area of at least 100 m 2 . Roads were only mapped when directly connected to buildings, up to 100 m to the buildings. Open spaces within a city were mapped as urban when surrounded by a built-up area. Still, they were excluded from the are to be de ned as 'exposed' to oods, when those open spaces consisted of vegetation and non-sealed surfaces close to a river within the city. The mapped areas were then compared to recent land cover data from Sentinel-2 and Open Street Map (OSM) data to capture urban growth and land-use change, but also in some cases to check the plausibility of urban features still existing at the same spot.
Flood hazard features were analysed using water areas and bodies, river course and other hydrological data available from different data providers (OSM, Regional Centre for Mapping of Resources for Development, SRTM, World Bank Data Catalogue, World Resources Institute). Since no ood hazard zonation maps were available and no DEM in su cient resolution, we have estimated a buffer zone around rivers in areas with rather even topography, crossing the city quarter of Nairobi North-East and Ruiru according to the average width of wetlands in data we found. We also checked it against a visual interpretation based on Sentinel-2 data from 2020. Since the wetlands had a buffer around the river of around 150-200 m in width in the case study area as measured in the data set, we have mapped ood exposed areas by the length of settlement perimeters within a 150 m buffer zone. 150 m have been selected to remain more on the cautious side of estimations. Since topography consists of relatively at basins in this area, the approach can be used, but of course, is not precise.

Results
Both case study areas, in Nairobi North-East and Nyeri, exhibit massive urban growth and growth into ood exposed areas and erosion and mass movement risk.

Nairobi -Ruiru
The mapping of the aerial image from 1948, the CORONA satellite data from 1977 and the Sentinel-2 data from 2020 reveals massive urban growth in north-eastern Nairobi (Fig. 2). The growth between 1948 and 1977 already is signi cant. It has to be noted that some of the settlement areas from 1948 as mapped from the aerial photo have not been mapped on the 1977 CORONA satellite image since the resolution was coarser. While some of the features, such as rectangular shapes that could be houses also are visible on the 1977 CORONA image, they were not mapped when the signal was not clear enough to retain a consistent mapping approach.
The assessment in Nairobi reveals a massive urban growth, as well as a growth of built-up area into potentially ood hazard-exposed zones, too (Fig. 3). To estimate ood exposure, buffer zones were created at 150 m width und have been checked against the topography as derived from a 30 m resolution SRTM model, as well as against pre-existing wetland data sets (Fig. 4). The results show that the 10m contour lines and computed slope degrees match the 150m buffer very well. The terrain is rather even, and therefore, buffers extend to both sides of the river course quite evenly.
Using GIS, the area sizes of the built-up areas are calculated, and the results show that for Ruiru, the built-up area had grown 20-fold from 1948 to 1977 and 10-fold from 1977 to 2020 (Table 2). At the same time, the potentially ood exposed area had grown 7 times from 1948 to 1977 and 3 times from 1977 to 2020. The analysis also reveals that Ruiru had grown much more than the adjacent South-Eastern part of the outskirts of Nairobi between 1977 to 2020. And the ood exposure of Ruiru had grown even much more by comparison to the built-up area. The growth rate of Nyeri between 1967 to 2020 is quite comparable to the area analysed in Nairobi. However, since satellite data with su cient resolution was not available from the same years, a direct comparison is not possible; it only can indicate similarities. Flood exposure also has increased in Nyeri, as the perimeter of the built-up area also has grown along the river Chanya. Land-use changes are also observable within the Ruiru area as well as in Nyeri. In Ruiru, some areas had already been visibly transformed from natural environments to agricultural usage (Fig. 5). Interestingly, many areas that were used intensively in 1948 (orange borders on the map) were not used in the same way in 1977. In 1948, the signature of the aerial image revealed darker colours and regular shapes of both roads crossing it and regular parallel linear structures as common for crops. In 1977, these areas revealed a speckled and heterogeneous re ection with lighter and mixed greyscales. On the other hand, some areas were intensively used in 1977, which had not been used in this way in 1948 (red borders on the map). Some other areas had been used both in 1948 and 1977, presumably for agriculture (not mapped).
These land-use changes of some areas are important for mapping urban growth, too, since they show the landscape before settlement. In the southern part of the aerial image, one larger area is mapped (Fig. 6). It can be important to monitor such previous land usages for loss of biodiversity or for revealing settlements that have overgrown previous wetlands. It also shows up opportunities for drought monitoring and understanding long-term changes.

Nyeri
In Nyeri, a similar strong urban development took place between 1967 to 2020 (Fig. 7). Aerial images from before the 1960s could not be obtained, so no comparison to the 1940s was feasible. The location of Nyeri along the Chanya river indicates a potential ood risk, con rmed by local experts. Urban growth reveals an expansion of the potentially oodexposed built-up areas. Of course, local topography must be observed, which is quite steep in this terrain and means that a river ood may not affect both sides equally.
In Nyeri, the terrain is steeper than in Nairobi-Ruiru. However, the 150 m buffer zone still ts quite well to the general shape of the river Chanya valley, as derived from the STRM digital elevation model (DEM) (Fig. 8). But of course, due to the steep terrain, a much more accurate DEM data set and ground observation are necessary to evaluate the ood exposure risk.
The SRTM data set exhibits some data processing aws, as visible by large rectangular shapes of the slopes in the upper left part of Figure 8. The high degrees of inclination also point out landslides and erosion as natural hazards and related exposure risks. The circles in Figure 8 indicate areas for further observation on-ground, where bridges and buildings could be at risk of ooding or mass movement.

Discussion
The case study analysis reveals some potential usage of GIS in combination with remote sensing data for monitoring land-use change over several decades. Signi cantly, aerial imagery and older declassi ed satellite data can expand the time range covered.

Data and methodology constraints
It has become evident that many constraints exist, such as in spatial resolution and temporal availability. Satellite and aerial imagery missions in the 1980s and before were limited in temporal coverage, spatial resolution and also, not all areas have been covered in a country. Typically, larger urban areas or other areas of strategic military interest often become mapped primarily (Day, et al. 1998). For the case studies, it was more di cult obtaining images for Nyeri than for Nairobi. But even for Nairobi, older missions have mainly mapped the urban area at that time. At urban rims such as Ruiru, the availability of imagery was constrained. For example, the high-resolution image from 1964 did not cover these outskirts of Nairobi anymore. resolution is also available for the built environment, and 2 m for 2015 data. Comparing the machine learning approaches with manual mapping, the manually extracted results still are more precise in this spatial resolution (Fig. 9). But of course, the manual mapping is also constrained by original data resolution and mapping skills.  (Soergel 2010). This is promising for recent satellite data but has limitations when applied to greyscale older satellite images with often medium or rather coarse spatial resolution.

Multimodal enrichment opportunities
The presented GIS assessment can be extended from land-use change mapping to enrichment by additional data and interpretations. For example, when comparing with recent OSM data, features such as previous river beds that now have been overgrown by settlement areas could be added to OSM (Fig. 10).
Other options would be to add other sources of VGI, such as photos of locals after ood events or Twitter messages (Taylor 2016). Another option is to add community efforts in preparing ood response measures, as mapped in Kibera; removing garbage, widening rain channels, diverting oods with sandbags etc. (Nelson 2016). Another important area is adding vulnerability and resilience information. For example, about the informal settlements which cover many areas in Nairobi. Certain slums in the city centre have been covered by studies (Corburn and Karanja 2014), but peri-urban areas lack coverage so far. The potential of remote sensing data, however has been demonstrated to capture such informal settlements and could show an important method to cover such areas (Kraff et al. 2020;Taubenböck et al. 2018) where data availability and local accessibility often are a challenge, and ethical aspects must be observed, too. The same applies to covering other features of urban composition, too, such as central business districts that carry another type of economic risk (Taubenböck et al. 2013). But next to social vulnerability, dependency of urban areas from critical or baseline infrastructure is an upcoming topic, for example, as demanded by the United Nations (United Nations 2015). Urban populations depend on piped water, roads, food, and many other urban services that should be integrated into holistic risk assessments (Ledant 2013). But also, the natural environment needs to be included in such holistic risk assessments. For example, the ecosystem services, the risk, the value of nature-based solutions and investments into sustainable land use (Vogl et al. 2017). And while this article has shown applications to natural hazards only to ood hazards, it also indicated land-use change monitoring for other hazards such as droughts. It could also be applied to erosion and landslide risk.

Conclusion
The article demonstrates how urban growth into ood-prone areas can be mapped using aerial and satellite images over several decades. Since current international bodies and concepts such as the Sendai Framework demand more riskinformed decision making (United Nations 2015), the need for data, monitoring and data visualisation for decision-makers and the public have grown. Social inequalities, vulnerabilities and countries in the global South are a focus area due to natural hazard events and fewer resources to cope with damages and losses (Warner and Zakieldeen 2012; Wrathall et al.

2015)
. However, large areas, especially in the global South, are largely uncovered for even basic risk monitoring of urban growth or spatial exposure to common natural hazards such as river oods. Data scarcity adds only to the problem and underlines the knowledge gaps persisting (Osuteye, et al. 2017). Overall, this may point towards a distorted understanding of the risk of societies, where the disaster risk could be less known or understood in some countries less covered by data and respective risk assessments. Hitherto rather unexplored data and respective data repositories such as aerial imagery captured by former colonial governments or espionage reconnaissance missions could provide key innovations to expand the current limited understanding of the dynamic spatial development of risk. The article also shows caveats of limited availability and usefulness of old data when constrained by accessibility, spatial and spectral resolution. Visual mapping is time-consuming and limited to detecting certain types of natural hazards such as oods or landslides, droughts or wild res. But it is a good starting point for risk mapping and expanding the application of remote sensing data.
Such temporal monitoring of exposure to natural hazards can help inform not only a static risk but also the dynamics of the opportunities existing for measuring changes using remote sensing data (Kraff, et al. 2020 fundamental footprint of human activity and the result of many decisions taken in combination, such as economic, legal and environmental decisions. It is often di cult to capture how to avoid risks in context to on-going transformations and one example could be to better guide "build back better" programmes (Gibson, et al. 2016;Kennedy et al. 2008) on how to avoid reconstruction in areas where settlement have overgrown former river beds. Mapping spatial exposure transformation can therefore be understood as a tool helping to document and understand a larger transformation. In context to climate change, more documentation of the hazard trends and human interaction with natural hazard-prone areas is necessary, especially in countries underresearched and hampered by data scarcity. Figure 1 The selected case study cities, Nairobi and Nyeri, are within the most populated areas in Kenya (data: left tile: population density in 1989, data by WRI 2007, neighbouring countries by OSM 2021 data. Right tiles: GHSL population grid by ), urban boundaries mapped using Sentinel-2 data) Figure 2 Mapped area North-East of Nairobi, along Thika Road, at Ruiru town. Urban built-up area from 1948, 1977 and 2020. (Data: see Table 1 for the aerial image, CORONA and Sentinel-2 data)

Figure 3
Urban growth and 150 m buffer river zone at Nairobi Thika Road NE, at Ruiru town.

Figure 4
Page 14/16 Contour lines in 10 m intervals and slope degrees (from SRTM data) overlayed on the 150 m buffer zone created Land-use changes in Ruiru from 1948 to 1977. (Data: see Table 1 for the aerial image and CORONA data) Figure 6 Development from intense agricultural use in 1948 to mixed-use in 1977 and the rather dense built-up area in 2020 in Ruiru. (Data: see Table 1 for the aerial image, CORONA and Sentinel-2 data) Figure 7 Urban growth in Nyeri from 1967-2020 Figure 8 Urban boundaries of Nyeri in 2020 overlayed on digital elevation model (SRTM), with Chanya river 150m buffer zone. Annex.docx