Revealing digitally invisible groups through a machine learning approach using multi-source data

Definition of digitally invisible groups

Data not being representative of the entire population is known as ‘data bias’ or the ‘digital divide’ (Doshi and Garschagen, 2023). Those individuals who are overlooked in digital data generation, collection and analysis processes are referred to as the ‘digitally invisible’ group (Taylor, 2017). Making extrapolations or assumptions about an entire population from datasets that under-sample digitally invisible groups can result in unjust policies and exacerbate social inequality. While some people actively opt for being digitally invisible due to privacy concerns, others find themselves in this position due to a lack of access to necessary infrastructure. The latter is the primary concern of this research and is labelled as unintentional non-participation (Lutz and Hoffmann, 2017), a situation where individuals are inadvertently excluded from data collection processes, and the issue has garnered attention from scholars studying digital inequality (Micheli et al., 2018). Although Doshi and Garschagen (2023) argued that other parties, including individuals, NGOs and civil society organisations in India, often step in to voice the concerns and experiences on behalf of digitally invisible groups, the challenge remains that no systematic research has been conducted to understand the extent to which these groups have been ignored and how this would affect data-driven decisions.

In the context of the Global South, the issue of digital invisibility is likely to be further exacerbated by the typical characteristics of inequality in developing cities. Digitally invisible groups, including lower-income, marginalised, and vulnerable populations, are systematically underrepresented in datasets due to limited access to technology, low smartphone usage, inconsistent internet connectivity and barriers like language, culture and privacy concerns (Taylor, 2017). For instance, people living in informal settlements are inherently more vulnerable than their counterparts in formal urban areas (Dodman et al., 2017), and are less likely to participate in sensor technology initiatives due to factors such as limited access to the internet and IT equipment or a lack of relevant knowledge (Wyche, 2015), which further exacerbates the digital invisibility issue (Yeboah et al., 2021). However, this group is projected to reach up to three billion worldwide by 2050 (Samper and Liao, 2023). The omission of such vulnerable groups reinforces inequalities, as the needs and experiences of these groups are less likely to be reflected in data-driven decisions.

Three dimensions of digital invisibility

In this study, we distinguish between digital invisibility as a condition of underrepresentation in digital data, and digitally invisible groups as the population segments affected by this condition. While data gaps have been widely discussed in geographical research, we emphasise that these gaps are not random but are often systematically associated with specific socio-demographic groups, such as low-income populations, informal settlement residents, or digitally disconnected users. Despite discussions across various sectors, such as digital media and data analysis, there remains a lack of systematic understanding of the issue of digitally invisible groups. By reviewing the existing literature, we categorise the concept of digital invisibility broadly into three dimensions of invisibility, namely spatially invisible, temporally invisible and demographically invisible, as elaborated below.

It is important to underscore that these dimensions are interrelated. For instance, one’s residential location often correlates with their socio-economic status (Luo et al., 2021). Understanding these categories helps in tailoring interventions that address unique challenges faced by each group. These dimensions of digital invisibility therefore provide a framework for linking data coverage gaps to specific population groups, which is essential for interpreting patterns of underrepresentation in data-driven applications.

Lack of standardised methods to infer digitally invisible groups

Existing research on digital data bias has largely been developed in the context of the Global North, where data availability is relatively high and digital infrastructure is more evenly distributed (Özkula and Reilly, 2024). In these settings, studies often focus on identifying and correcting biases within datasets, such as sampling bias or measurement error, assuming that most population groups are at least partially represented. In contrast, in the Global South, the challenge is more fundamental: large segments of the population may be entirely absent from digital datasets due to limited access to technology and infrastructure (Raihan et al., 2024). As a result, digital invisibility is not only a matter of bias within data, but also of complete exclusion from data generation processes (Arora et al., 2023). Despite this critical difference, relatively few studies explicitly compare how these challenges manifest across contexts or develop methods to systematically identify who remains digitally invisible.

Previous relevant work presents a critical gap in methodology design. We found no standardised methods to reveal the presence of digitally invisible groups or to assess how data biases affect indices derived from significant data sources. Although some studies address data biases and representativeness issues within big data, they often lack systematic ways to infer who is excluded and why. For example, Poledna et al. (2018) analysed biases in two mobile-phone-derived datasets, finding that they largely mirror telecom market share and subscription patterns rather than population coverage (i.e., who has a chance to be observed by the data). Moreover, for privacy reasons, this method was designed not to map exact locations of individuals but to make general estimates about movement patterns and locations visited. Similarly, Jürgens et al. (2019) looked at biases in data collected through self-reporting by comparing it to data gathered automatically from mobile devices. They found that self-reported information and automatically collected data produced different results, with mobile device tracking reducing some measurement errors. This study underscores that different methods capture different types of information. Neffendorf et al. (2018) applied a similar comparison in the UK, examining location data from 254 mobile phone users against their self-reported travel records. They found that while work locations were accurately identified, home locations were somewhat reliable, though some devices were inactive during travel times, leading to gaps. Conversely, Embury (2023) used mobile app data and travel surveys to detect biases in synthetic population models, which have the potential to harm marginalised individuals and groups. This work further highlights the variability in methods and the need for standardised approaches to address who remains ‘digitally invisible’.

To address the knowledge gap on digital invisibility in urban settings, we propose a scalable and transferable methodology that identifies digitally invisible groups through stepwise integration of multi-source data. We apply this workflow to land use classification in Nairobi, demonstrating its effectiveness in inferring digitally invisible groups. This approach enhances the inclusiveness of urban mapping by integrating diverse datasets that capture unique aspects of human activity and the physical environment.

Study area setting

Kenya, located in East Africa, is a hub of African economic, political and digital development (Ngui Muchai and Kimuyu, 2017; Ren et al., 2020) (Figure 1). From 2021 to 2022, the number of internet users in Kenya grew by 1.6 million, a 7.4% increase (Kemp, 2022). The country has also launched the Kenya National Spatial Data Infrastructure to promote spatial data usage (Mwungu et al., 2017) and support land use planning (Ministry of Lands and Physical Planning, 2017). Nairobi, Kenya’s capital and economic hub, is recognised globally for its innovation, ranking among the top 100 cities in the 2019 Global Innovation Index (Ren et al., 2020; Ministry of ICT Innovation and Youth Affairs, 2022). The city increasingly leverages data sciences and technology for development monitoring (Nairobi City County, 2023).

OPEN IN VIEWER

Figure 1.

Location of Nairobi and population count distribution across 100 m × 100 m grid cells. Data sources: Administrative boundaries from GADM; population count data per 100 m grid from WorldPop (2020).

Nairobi offers significant potential for data-driven applications, supported by robust infrastructure and a population willing to engage in digital initiatives. According to Kiburi et al. (2022), phone access among individuals was estimated to be 77.2%, and among them, 59.7% used smartphones. Nearly all households had access to some form of phone communication. Safaricom, Kenya’s leading mobile network, provides the most comprehensive mobile data coverage, making it a valuable resource for spatially referenced analyses of human behaviour. However, accessing mobile phone data poses challenges due to privacy concerns. Social media data, while more readily available due to voluntary participation, remain limited in representativeness in Kenya. Social media platforms collectively reach approximately 23.5% of the total population, whereas Twitter reaches only 3.4% (Kemp, 2024). This disparity underscores the restricted penetration of platform-based data, with many excluded users residing in informal settlements due to persistent technical and infrastructural barriers. With strong government support for digital initiatives and relatively high levels of data availability, Nairobi stands out as one of the most digitally advanced cities in the Global South. It is positioned at the forefront of urban digital transformation, facing challenges that many other cities in the region are likely to encounter as they progress through similar data-driven development pathways.

Nairobi faces rapid population growth, especially in informal settlements. According to Georganos et al. (2021), informal settlements occupy less than 5% of the city’s residential land but home more than 60% of the population, as shown in Figure 1. Although there has been some progress in expanding digital access in larger informal settlements, such as Kibera, such effort is limited to a local scale, leaving many communities disconnected (Williams et al., 2020).

Workflow: Reveal digitally invisible groups via land use classification

There is no standard method or reliable ground truth for identifying digitally invisible groups. We address this by comparing modelled outcomes from a single data source with outcomes produced after stepwise integration of additional sources. Digitally invisible groups are conceptualised as systematic misclassification across different combinations of data inputs. We demonstrate this idea with land-use classification, using a simple diagnostic: hold the modelling framework constant and compare outputs from a single data source with outputs after adding other sources step by step.

To implement this, we adopt a stepwise data integration strategy, in which models are trained repeatedly using incremental combinations of input data sources (e.g., R, RN, RNO, RNT, RNS, . . ., RNOTS). This design allows us to isolate the contribution of each dataset and examine how classification performance changes when specific data sources are included or excluded.

We choose land-use classification because it is a well-established task with benchmarks (Pan et al., 2013; Soliman et al., 2017; Talukdar et al., 2020) and because land-use classes correspond to human activities (Toole et al., 2012), allowing us to see where some groups are likely to be missed by particular sources (e.g., building footprints, Twitter, street view imagery). More specifically, when a data source is introduced or excluded, certain land-use types, and the associated activity patterns, may be systematically under- or over-classified because the data disproportionately omit or capture these activities.

Because there is no reliable ground truth for digitally invisible groups, these output comparisons are more informative than raw coverage checks and represent our attempt to address this gap. The workflow has four steps as shown below and in Figure 2. A full explanation of the workflow can be found in Appendix 8.1

OPEN IN VIEWER

Figure 2.

Workflow via land use classification.

In this study, digital invisibility is operationalised as systematic differences in classification performance across data source combinations. Rather than directly observing socio-demographic groups, we infer potential underrepresentation by analysing class-specific changes (e.g., producer’s accuracy and misclassification patterns) when particular data sources are included or excluded. Population groups are therefore represented indirectly through land-use categories, which act as proxies for activity-based population segments. We do not directly measure socio-demographic groups; instead, we infer potential underrepresentation through model behaviour across data combinations.

Data

Multi-source data

In typical land use classification, remote sensing data, including satellite images (R) and nighttime light imagery (N), usually serve as the base layer for mapping land use. Although nighttime light imagery is derived from satellite sensors, it is treated as a separate data source due to its role as a proxy for human activity. To demonstrate the utility of our workflow in inferring digitally invisible groups, we started by using satellite imagery for land use classification, and then following a stepwise approach to incorporate additional data sources including building footprint (O), social media post (T) and street view imagery (S), as illustrated in Figure 2. Understanding the strengths and limitations of each data source is crucial for interpreting findings on digital invisibility. Table 1 summarises the input data sources and indicates which groups are likely to be over- or underestimated.

OPEN IN VIEWER

Table 1.

Summary of multi-source big data.

Data (abbr.)	Satellite imagery(R)	Nighttime light imagery(N)	Building footprint (O)	Social media post (T)	Street view imagery(S)
Source	Sentinel-2 MSI (Copernicus/ESA)	VIIRS-DNB (NOAA)	Google Open Building V3	Twitter posts (via API)	Mapillary
Method of data generation	Remote sensing	Remote sensing	Remote sensing	Smart phone	Smart phone
Sample time	2022.01-2022.12	2022.01-2022.12	2023.05	2019-2022	2017-2022
Data type	Raster data	Raster data	Vector data	Vector data	Vector data
Data format	Pixel size 10 m × 10 m	Pixel size 760 m × 760 m	Polygon	Geo-referenced text	Geo-located image
Feature selection	Cloud masked basic bands	Night Band	Building density	Post language, Post time	Elements identified from image
Feature interpretation	Land cover characteristics	Urban extentEconomic activity via light intensity	Structural density and morphology	Human activity, language, mobility patterns	Ground-level features: roads, vendors, walls
Likely overrepresented groups	N/A All regions equally (no cloud coverage)	Electrified urban extent	All built environments covered; more low-confidence buildings appear near urban poor areas	Young, educated, urban, smartphone users	Accessible/formal neighbourhoods, roadside vendors
Likely underrepresented/invisible groups	No human activity	Off electricity grid; No social activity	No human activity	Older adults, low-income groups, informal workers, rural users	Deep informal settlements, gated areas, footpaths
Representation bias	Limited to physical features, no insight into population	Underrepresents low-income settlements	Capture no human activity or socio-economic function; purely physical	Biased by tech access, literacy, platform usage	Skewed toward roads and public-facing infrastructure

Data processing and random forest classifier

Building footprint, Twitter posts and street view imagery were pre-processed before being used in the classification model, summarised in Table 2. We used building data from Google Research open data source to calculate building density, providing insights into urban morphology (Sirko et al., 2021).

OPEN IN VIEWER

Table 2.

Features selected from multiple data sources.

Feature	Data type	Feature Name	Processing	Feature Description
1–13	Satellite Image (R)	b1–b13	Raw	Original satellite bands (e.g., blue, green, red) representing spectral information
14	Nighttime Light Imagery (N)	NTL	Raw	Nighttime light day-night band
15	Building Footprint (O)	OB	Density calculation	Building density derived from open building data
16–22	Twitter Post (T)	TTime_[Location]D TLang_[Language]D	NLP with SpaCy	Locations identified as home, work and leisure based on Twitter posts, with language classification into English, Swahili, or other languages
23–51	Street View Imagery (S)	SVI_[SingleObject] SVI_[Combine_Object]	Panoramic segmentation with Detectron2	Object detection and segmentation in street view imagery (e.g., roads, buildings, trees, cars, wallbrick)

Georeferenced Twitter posts were pre-processed using cleaning procedure proposed by Lansley and Longley (2016). Following this method, we removed tweets that meet the following criteria: a) Repeated tweets from the same user exceeding three instances, and b) Tweets from users who posted more than 3,000 times within a year. The cleaned georeferenced Twitter posts were then categorised into three groups based on the timing of posts: work/school locations, leisure activity locations and home locations. Language detection techniques from spaCy (Montani et al., 2020) were applied to analyse the content of these posts, classifying them into three language groups: English, Swahili and other non-official languages.

Street view imagery was processed through panoramic segmentation using Detectron2 (Wu et al., 2019), a pre-trained image segmentation and object detection algorithm developed by Meta. Detected objects from these images were transformed into features for the classification model. The resulting point dataset was then aggregated into a grid to align with the spatial resolution of the remote sensing data. A summary of the selected features used for inferring and estimating land use types is presented in Table 2.

Input data were resampled to a uniform 10 m resolution to ensure spatial alignment across datasets, although this may introduce smoothing effects and reduce fine-scale variability. Classifications were also tested at coarser scales (30–100 m in 10 m increments) to assess the sensitivity of model performance to spatial resolution.

We randomly sampled 1,000 pixels per land-use class and split them into training and validation sets (70:30). A random forest classifier with 200 trees and a fixed random seed (42) was used to ensure stable and reproducible performance. Missing values resulting from incomplete data coverage (e.g., absence of Twitter or street view observations) were retained and encoded as zero, thereby implicitly representing data absence within the feature space. All models were trained using identical hyperparameter settings across data combinations to ensure comparability.

Overall model performance

Figure 3 illustrates the output land use classification for Nairobi using a combination of all data sources, and Figure 4 illustrates changes in overall accuracy across spatial resolutions as they are introduced step by step into the model. A consistent improvement in overall accuracy is observed across all spatial resolution groups, with the highest OA of 73.57% achieved at a 30-meter resolution when all input datasets were integrated. This improvement reflects the complementary information provided by different datasets, including variations in spatial coverage and the ability to capture distinct aspects of the urban environment. However, these gains should not be interpreted as a direct indication of improved representation of underrepresented population groups. Instead, they reflect aggregate model performance, which may arise from multiple factors unrelated to population visibility. Therefore, in the following sections, we focus on class-specific performance changes to examine how representation varies across data sources.

OPEN IN VIEWER

Figure3.

Land use classification of Nairobi at 30-meter spatial resolution with all data.

OPEN IN VIEWER

Figure 4.

Change of overall accuracy by data type combinations across spatial resolutions and standard deviation: The x-axis highlights different data type combinations (outlined in Table 2), while each colour corresponds to a spatial resolution group. The left y-axis represents overall accuracy (depicted as lines), while the right y-axis corresponds to the standard deviation for mean overall accuracy (represented as bars). This depiction illustrates both the variations in overall accuracy across spatial resolutions and the associated variability.

Error analysis and feature interpretation

This section provides an overview of how classification errors are redistributed across land-use types under different data combinations, based on confusion matrix analysis. Detailed class-specific evaluation is presented under ‘Class-specific evaluation by land use type’. The confusion matrix analysis focuses on false-negative misclassifications resulting from horizontal omission errors (Barsi et al., 2018). Figure 5 shows the baseline datasets: the basic combination of remote sensing and nighttime light imagery (RN, in Figure 5a) and the fully integrated dataset of RNOST (All, in Figure 5h). To assess the individual contribution of building footprint (O), Twitter (T) and street view imagery (S) data sources, we compare RNO, RNT and RNS with RN. We then compare the All with RNOT, RNTS and RNOS to evaluate the additional value of each dataset when introduced at a later stage of integration.

OPEN IN VIEWER

Figure 5.

Confusion matrix across different data combinations: The x-axis denotes the predicted classification of pixels, while the y-axis represents the actual reference pixels. Each cell contains the count of pixels, colour coded by value.

By comparing RNO with RN and ALL with RNTS, we observe that incorporating the building footprint led to a consistent improvement in accuracy for residential areas. This change significantly reduced misclassifications of vegetation as residential areas. In addition, in the absence of Twitter and street view imagery, the inclusion of building density also enhanced the identification of industrial areas. However, this improvement overlapped with the contributions from Twitter as shown by comparing RNTS and RNOS with RNS, as well as RNOT with RNT and RNO. Despite a similar level of improvement in revealing industrial areas, building footprint led to an overestimation of industrial areas at the expense of commercial areas, as both share similar morphological features, including larger and more sparsely distributed buildings. These misclassifications suggest that building data alone may not fully capture activity-based distinctions. Thus, although building footprints improve spatial recognition of physical land use, they may fall short in representing the social or economic activities of certain population groups.

Twitter data highlights patterns of underrepresentation of commercial activities at any stage of integration, as seen in the comparison from RN to RNT, RNS to RNTS, RNO to RNOT and RNOS to ALL. However, the trade-off is an overestimation of commercial activities at the expense of residential and institutional activities. This overestimation may arise because many commercial activity patterns overlap with other land uses: for example, individuals working in office buildings may follow similar schedules to those in government institutions, and retail and catering activities often occur outside standard office hours and can overlap with times when people are likely to be at home. As Twitter captures patterns of digitally engaged users during activity-rich periods, these overlaps introduce classification ambiguity.

SVI improved the identification of commercial activities at both early and later stages of data integration (RN to RNS and RNOT to All), indicating that the commercial activities identified from street view imagery represent a different group of commercial activities from those identified through Twitter data. In addition, it also contributed significantly to the accurate classification of informal settlements. This is noteworthy given that many informal settlements are not directly accessible by vehicle and therefore have limited street view imagery coverage. The strong performance of SVI in this context suggests that even partial visibility, particularly along settlement boundaries, can capture distinctive environmental features that differentiate informal settlements from formal areas. These features may include differences in the presence of pavement, fencing, or other physical infrastructure. This result highlights how comparing outputs from different data sources helps reveal how even incomplete street view coverage can capture previously underrepresented activity patterns associated with informal settlements.

While the confusion matrix highlights how classification errors are distributed across land-use types, it does not explain the contribution of individual variables to these outcomes. To complement this analysis, we apply SHapley Additive exPlanations (SHAP) to interpret feature contributions (Figure 6). The results show that OB is the most influential feature across multiple land-use classes, particularly for distinguishing residential and informal settlement areas. NTL and spectral features provide more general contributions, while social media features (e.g., TTime and TLang) have comparatively lower influence, especially for informal settlements. These patterns indicate that different data sources contribute unevenly to the classification process, helping to explain the class-specific patterns observed in the confusion matrix.

OPEN IN VIEWER

Figure 6.

Global SHAP feature importance showing mean absolute SHAP values across all land use classes, with contributions stacked by class. Higher values indicate greater influence on model predictions.

Class-specific evaluation by land use type

We focus on commercial, industrial and informal settlement land use types, as these were most impacted by multi-source data integration. Their differing responses to different data inputs indicate that activity patterns associated with these areas are captured unevenly across datasets, making them particularly relevant for demonstrating our workflow and identifying patterns of underrepresentation across data sources. Within this framework, ‘digitally invisible groups’ refer to population segments approximated through land-use types, where persistent underrepresentation or misclassification is observed when relying on specific data sources. We interpret these patterns as indirect evidence of uneven representation of the activities, and by extension, the populations, associated with these land-use categories. For example, informal settlement areas may be underrepresented in social media data, while certain commercially active areas may be overrepresented.

For each selected land use type, we provide the producer accuracy and feature importance plot to further understand the role of different data sources in revealing and mitigating digital invisibility. The producer importance plots depict the accuracy of a land use type under various data combinations, represented by bars and colours. RN is used as the baseline model, and the effects of additional data sources are evaluated relative to this combined reference rather than individual datasets. The subplots should be interpreted in an anticlockwise direction. The feature importance plots list features derived from five input datasets and are used to interpret how specific data sources may influence the representation of activity patterns. The central x-axis in each sub-figure represents the mean feature importance for each type. A lower value below the mean does not indicate an adverse impact; rather, it suggests the feature has less influence. However, multiple low-impact features can collectively affect the results. Full producer accuracy and feature importance results are provided in Figures S2 and S3 in Appendix.

Commercial land use area

Incorporating Twitter and street view imagery significantly improves the classification accuracy of commercial activities, with RNT and RNS improving accuracy by 16.89% and 12.84%, respectively, compared to RN, as shown in Figure 7a. This enhancement demonstrates that both Twitter and SVI effectively capture human commercial activities that traditional RS and building data fail to detect.

OPEN IN VIEWER

Figure 7.

Producer’s accuracy (a) and variable importance (b) of the commercial land use type. (a) Producer’s accuracy: the x-axis corresponds to different data combinations, and the bars indicate the estimated accuracy values. (b) Variable importance: the x-axis is the feature importance value, and the y-axis is the selected feature by input data, from left to right: satellite imagery, nighttime light imagery, building footprint, street view imagery and Twitter.

Figure 7b further explores the groups of people identified and missed by Twitter and SVI. Twitter-derived location features such as work, home and recreational location effectively distinguish between daily leisure activities and home-based activities, potentially reflecting patterns like dining out or entertainment. The notable contribution of TLang_OthD, which is the ‘other’ language, supports the identification of international commercial activities, such as international business and tourism. This finding is intuitive: people often share their travel and dining experiences on social media, and consequently, the tourism and food industries are more inclined to advertise online (Azhar et al., 2023; Rini et al., 2024). Incorporating Twitter data into ALL results in a 6.75% improvement over using RNOS alone, highlighting the irreplaceable ability to infer formal commercial activities.

Among SVI features, SVI4_InformalBUD, which aggregates objects commonly associated with informal settlements (e.g., unpaved roads and umbrellas), exceeds the average feature importance for commercial classification. In particular, the presence of ‘umbrella’ often indicates local commercial activities like street vendors, kiosks and food carts, as shown in Figure 8b. This aligns with findings from the World Bank Group (2016), which reported that 95% of all businesses in Kenya operate within the informal economy.

OPEN IN VIEWER

Figure 8.

Panoramic segmentation results for street views in commercial and industrial areas: umbrellas are visible in the commercial area (7a and 7b), while pavement and gravel are visible in the industrial area (7c and 7d).

Importantly, the improvements brought by Twitter and SVI are complementary rather than overlapping. This is evidenced by the performance differences between RNT, RNS and RNTS, as well as among RNOS, RNOT and All. These differences indicate that while there is some overlap – such as in locations like the Maasai Market, where local markets attract tourists – each data source also captures distinct groups and spatial patterns of commercial activity. This complementary nature underscores the value of integrating both Twitter and SVI data for a more comprehensive understanding of urban commercial dynamics.

The producer accuracy for commercial land use decreases with the introduction of building footprint data, as seen in the RNO combination. This suggests that building footprints alone, which reflect only physical structures, are not effective in identifying commercial activities. However, when combined with SVI features (RNOS), the accuracy improves by 3.39% over RNS, indicating that building footprints can still contribute to infer a certain group of people when integrated with other data sources. For example, while building data alone may cause confusion between commercial and industrial areas, SVI features like SVI_Pavement help distinguish them: local markets often have goods displayed directly on the ground, while industrial areas typically maintain visible pavement for workers commuting. This demonstrates the synergistic effect of combining datasets to reveal the underrepresentation of certain digitally invisible groups.

Industrial land use area

Figure 9a shows that the inclusion of building footprint data, Twitter and SVI improves the identification of industrial activities, achieving an average improvement of 7.07% over RN. Further gains are observed when multiple datasets are combined in pairs, reaching an average producer accuracy of 81.67%, and achieving the highest producer accuracy of 82.86% when all datasets are integrated. This synergy arises because each dataset captures distinct characteristics of industrial activities.

OPEN IN VIEWER

Figure 9.

Producer’s accuracy (a) and variable importance (b) of the industrial land use type.

In Figure 9b, nighttime light imagery is widely recognised as a reliable indicator of industrial and economic activity (Cui et al., 2023; Qi et al., 2017). Building footprints highlight areas with larger building sizes and lower density, typical of industrial zones designed for vehicle loading for heavy manufacturing. However, similar patterns are also found in some commercial areas, such as shopping malls, which can introduce confusion between industrial and commercial classifications. The inclusion of SVI features helps mitigate this confusion. For example, SVI4_InformalBUD contributes significantly, possibly due to elements like gravel, as industrial sites often have debris piles. The presence of SVI_Pavement further supports the hypothesis that industrial areas maintain footpaths for daily worker commuting, as shown in Figure 8d. A similar effect is observed for SVI_Wall, which reflects security infrastructure typically associated with more managed environments such as industrial activities.

From Figure 9b, TTime_relaxD and TTime_workD are the main contributors to the identification of the industrial area. This is likely due to the long working hours in Nairobi’s manufacturing plants, factories and warehouses, where the distinction between work and recreation is often blurred (Ongeri Ondimu, 2024). Additionally, Nairobi’s industrial sector employs a diverse workforce, including individuals from various regions of Kenya, other African countries and expatriates from around the world, as indicated by TLang_othD (Castañeda-Ramírez and Martínez-Mares, 2025).

Informal settlement land use area

From Figure 10a, both O and S significantly enhance the classification of informal settlements, improving producer accuracy to 83.95% and 88.27%, respectively, as shown in Figure 10b. O is expectedly the dominant contributor for identifying the high building density characteristic. This is further supported by SHAP analysis (Figure 11), which shows that OB has the strongest contribution to the predicted probability of informal settlement areas, with higher values consistently increasing the likelihood of this class. Partial dependence analysis (Figure. S4 in the Appendix) further confirms the non-linear relationship between OB and the predicted probability of informal settlement areas. This feature is widely used for urban slum mapping in the Global South (Bettencourt and Marchio, 2023; Boo et al., 2022).

OPEN IN VIEWER

Figure 10.

Producer’s accuracy (a) and variable importance (b) of the informal settlement land use type.

OPEN IN VIEWER

Figure11.

SHAP beeswarm summary for informal settlements. Points represent individual samples, coloured by feature value.

SVI also contributes significantly to the identification of informal settlements. This is counterintuitive, as informal settlements are generally considered inaccessible for SVI collection. However, even without access to the core areas of informal settlements, SVI helps delineate their boundaries, as these settlements are mainly located within urban areas. As shown in Figure 10b, the SVI_Grass and SVI_Wall features stand out despite the overall higher contribution of building footprint. SHAP analysis further supports the contribution of SVI, although the specific features highlighted differ from those identified through impurity-based importance. This reflects that SVI-derived variables capture a range of contextual signals associated with informal settlements, even if their individual contributions vary across methods.

The inclusion of Twitter data from RN does not significantly influence the classification of informal settlement residents. This is also reflected in the SHAP analysis, where Twitter-derived variables exhibit minimal contributions. This pattern is intuitive, given that smartphone ownership and internet penetration within highly informal settlements are still relatively low (Williams et al., 2020; Wyche, 2015).

Identification of digitally invisible groups from multi-source data

Table 3 summarises which groups are well represented or underrepresented across different data sources, revealing patterns of digital invisibility in Nairobi. These patterns are not inferred from data availability alone, but from systematic differences in model performance when specific datasets are included or excluded.

OPEN IN VIEWER

Table 3.

Summary of digitally invisible groups from multi-source data.

Dataset	Groups/Factors well captured	Digitally invisible groups	Notes
Satellite imagery + nighttime light	Land cover, physical structures, general urban form and industrial zones	Commercial activities, informal vendors, some industrial workers	Good for mapping structures, limited in distinguishing between activity types*
Building footprint	Urban morphology, Informal settlements	People engaged in activities	Good for mapping informal settlements
Twitter post	Commercial actors, tourists, middle- and high-income users	Informal settlement residents, low-income workers	Reflects active social media users and public spaces
Street view imagery	Informal settlement boundaries, street-level commerce (e.g., vendors)	Core areas of informal settlements, inaccessible neighbourhoods	Captures visible street objects, limited coverage inside slums
Combined sources	Broader spectrum across groups	Highly marginalised, digitally disconnected populations still underrepresented	Integration improves but gaps remain

Activity types*: refers to the differentiation of land-use-related activities (e.g., distinguishing between commercial, residential and informal uses), rather than the presence of human activity in general.

Twitter data significantly improves commercial and industrial classification but contributes little to informal settlements. This selective improvement suggests a digital silence from informal settlement residents, likely due to limited internet access or device ownership, making them invisible in social media-derived data. Similarly, while SVI improves classification at settlement boundaries, core areas remain unmapped due to limited vehicle access, reinforcing spatial exclusion. Building footprint data, though valuable in highlighting dense, informal urban fabric, lacks behavioural information and cannot differentiate between land uses shaped by human activity, leaving economically marginalised groups partially obscured in activity-based models.

These disparities reflect not only technical limitations but also social and spatial inequalities in data coverage. Persistent misclassification or only small gains for certain land use types, even after adding new sources, indicate that some populations remain digitally underrepresented. Adding a single dataset does not reliably improve performance and can reinforce errors when that source underrepresents particular groups. Combining complementary sources can reduce blind spots because each captures different people and places, but improvements are class-dependent and should be verified through systematic, per-class evaluation rather than by aggregating everything. In short, digital invisibility is exposed, and partly mitigated, by how performance changes with each data source and urban context.