Informing action for United Nations SDG target 8.7 and interdependent SDGs: Examining modern slavery from space

The United Nations Sustainable Development Goals (SDGs) form an interdependent framework for global development, with Target 8.7 calling for effective measures to end modern slavery and human trafficking by 2030. Measuring progress toward 8.7 is challenging due to the hidden nature of affected populations. Conventional approaches—event, standards, and survey-based data—have provided valuable global prevalence estimates (e.g., the Global Slavery Index), but suffer from limitations related to country coverage, extrapolation, and temporal updating. The article argues that satellite Earth observation (EO) can complement these approaches by identifying physical sites where exploitative labour is likely, enabling spatial and temporal analysis of risk and impacts. The study focuses on Bull’s Trench brick kilns across the 1.5 million km² South Asian ‘Brick Belt’ (Pakistan, northern India, Nepal, Bangladesh), where reports document widespread bonded and child labour and trafficking into forced labour. Beyond social harms, kilns impact environmental quality and health through resource extraction and emissions, linking SDG 8.7 with SDGs on health (3), urbanisation and consumption (11, 12), life below water and on land (14, 15), and climate action (13). By mapping the geography and construction dates of kilns, the study establishes a robust, updateable baseline for the industry’s extent, examines drivers of kiln installation (e.g., urban demand), and quantifies environmental impacts, thereby informing policy and targeted anti-slavery interventions.

Prior work on modern slavery measurement highlights reliance on surveys and statistical models (e.g., Global Slavery Index; Landman 2018, 2020; Landman and Silverman 2019; Silverman 2020) and their limitations for hidden populations and temporal monitoring. Reports document extensive exploitation in South Asian brick kilns, including bonded labour and child labour (Bales 2012; Kara 2014; ILO 2005; ASI 2017; Save the Children 2007), with national kiln counts previously estimated but inconsistent across time and methods. Emerging evidence links modern slavery to environmental degradation and climate change (Brown et al., 2019; Decker Sparks et al., 2021), suggesting SDG interdependencies (Nilsson et al., 2016; Pradhan et al., 2017). Remote sensing has been demonstrated for detecting brick kilns and related environmental analyses (Boyd et al., 2018; Foody et al., 2019; Nazir et al., 2020; Misra et al., 2020; Lee et al., 2021), but a comprehensive, contiguous spatiotemporal mapping across the entire Brick Belt had not been achieved. This study builds on those insights to provide the first rigorous mapping and dating of kilns region-wide.

Study area: The ‘Brick Belt’ spans ~1,551,997 km² across Pakistan, northern India, Nepal, and Bangladesh along the Indo-Gangetic Plain—an agriculturally vital and rapidly urbanising region where Bull’s Trench and Fixed Chimney Bull’s Trench kilns are common and exploitative labour practices have been reported.

Spatiotemporal mapping of kilns: The team used contiguous very high-resolution (VHR) Airbus Pléiades optical imagery (50 cm, blue/green/red/NIR, 20 km swath) covering the entire Brick Belt and a YOLOv3 convolutional neural network (CNN) for object detection of kilns, leveraging distinctive features (ring shape, size/shape, central chimney/shadow) to distinguish from similar objects. Training followed an iterative three-stage process: (1) initial model trained on ~1000 crowdsourced, expert-validated kiln annotations from Rajasthan (Zooniverse; Boyd et al., 2018) and applied across the Belt; (2) refinement by randomly selecting 3000 detections, expert visually reviewing 2×2 km windows to curate a diverse training set of 2709 kilns representing spatial variability; (3) final training on 2209 kilns (500 held out for validation/early stopping), network size 608×608, trained >24 h over >20,000 epochs. Operational inference split the region into 50 km grid cells (893 cells), further tiled into 1 km² patches with overlaps to avoid edge loss; duplicate/overlapping predictions were merged via geospatial filtering. Results were stored in a PostGIS database. Accuracy assessment via expert visual interpretation yielded overall accuracy 92.83%, precision 98%, recall 87%.

Dating kiln construction: For each detected kiln, the construction year was estimated using Landsat 5 TM, 7 ETM+, and 8 OLI surface reflectance time series (five bands: green, red, NIR, SWIR1, SWIR2), with cloud/shadow masking (CFmask), spanning 1988–2018. Visual inspection of earliest available VHR imagery in Google Earth Engine (GEE) since 2000 established kiln presence/absence windows (t_f first observed; t_o′ closest image before t_f), constraining the likely construction interval. Breakpoint detection identified significant changes in mean, slope, and variance in the Landsat time series. A Random Forest (RF) classifier, trained on kilns with visually determined construction windows (training n=15,408; validation n=15,075 distributed across India, Bangladesh, Pakistan, Nepal), distinguished ‘true’ construction-related breaks from other land cover transitions using 30 features per break. An empirically determined probability threshold selected candidate breaks; if none were detected, construction was inferred pre-1988. Validation assumed correctness if predicted date fell within ±182.5 days of the target year; mean differences were 1.83 years and −1.55 years relative to reference bounds.

Supply–demand analysis: As a proxy for annual population/urban demand, the study used NOAA DMSP-OLS Night-Time Lights (NTL) avg_lights_x_pct composites (30 arc-second) for 1996–2013, computing annual totals across the Brick Belt. Annual cumulative kiln counts were correlated with NTL totals using Pearson’s r to assess association between kiln growth (supply) and urban/economic activity (demand).

Environmental impact assessments: Land cover impacts were quantified by intersecting annual kiln construction locations with MODIS MCD12Q1 IGBP land cover (2001–2017) to determine the distribution of kiln siting by class and its evolution. Resource extraction was estimated using conservative production of 2 million bricks per kiln per year and literature-based extraction factors: topsoil usage benchmarked to Indian estimates (approx. 400 million tonnes per year nationally, scaled by kiln counts) and groundwater use at ~0.75 L per brick (Shrestha et al., 2013), yielding ~150,000 L per kiln-year. Proximity to hydrological features was assessed by buffering rivers/canals/lakes/reservoirs at 1–5 km and intersecting kiln locations; kilns within channels/floodplains were recorded. Air quality impacts were examined using Global Annual PM2.5 (MODIS/MISR/SeaWiFS AOD with GWR, 0.01°, 1998–2016) to characterise temporal and spatial trends over the Belt and, at city scale (GADM v3.6 boundaries), relate annual PM2.5 concentrations to local kiln counts in major urban centres.

• Comprehensive mapping identified 66,455 brick kilns across the Brick Belt as of 2018: 69.7% (46,319) in India, 17.3% (11,497) in Pakistan, 11.3% (7,509) in Bangladesh, and 1.7% (1,130) in Nepal.
• Detection performance: overall accuracy 92.83%, precision 98%, recall 87%.
• Temporal dynamics: Year-by-year maps from 1988–2018 show sustained growth in kiln installations with spatially heterogeneous patterns.
• Construction dating: RF-based breakpoint classification on Landsat time series achieved mean difference of ~1.83 years between predicted and reference construction timing windows.
• Supply–demand relationship: Strong positive correlation between cumulative kiln counts and NTL-based demand proxy over 1996–2013 (Pearson r = 0.84, p < 0.001), indicating kiln growth tracks urban/economic activity.
• Land cover siting: Approximately 80% of kilns are located on cropland, indicating a pronounced conversion/occupation of agricultural land.
• Resource extraction (2018): Estimated topsoil extraction totaled ~266 million metric tonnes; groundwater use across all kilns totaled ~101.64 billion litres.
• Hydrological proximity (2018): The majority of kilns are within 5 km of water bodies; 1.27% located within channels (<1 km). By buffer across all waters: 29.35% within 1 km, 23.86% within 2 km, 16.88% within 3 km, 11.28% within 4 km, 6.55% within 5 km, 9.54% beyond 5 km (Table 3).
• Air quality: PM2.5 concentrations increased across 1998–2016, with spatial peaks near major urban centres (e.g., Delhi). City-level analyses (e.g., Delhi, Dhaka, Lahore, Multan, Gujranwala, Kathmandu, Patna, Varanasi, Kanpur/Barne, Assam) show co-increasing trends in kiln numbers and PM2.5.
• Policy-relevant insight: Kilns cluster around large urban centres and infrastructure, suggesting localised supply chains and labour demand pressures.

Mapping brick kilns as “objects of SDG intersectionality” provides a spatial proxy for potential prevalence and vulnerability to modern slavery (SDG 8.7), while simultaneously enabling assessment of co-impacts on health (SDG 3), sustainable cities and consumption/production (SDGs 11, 12), climate (SDG 13), and terrestrial/aquatic ecosystems (SDGs 14, 15). Although kiln presence does not directly measure slavery, high documented likelihood of exploitative labour in this sector means the dataset offers crucial, actionable intelligence for policy makers and NGOs to target prevention, inspections, regulation, and liberation efforts. The strong association between kiln growth and NTL indicates urban demand for bricks as a key driver, implying that continued urbanisation without safeguards risks perpetuating exploitative labour practices. Environmental analyses reveal significant pressures on croplands (kiln siting and topsoil extraction), water resources (groundwater use and proximity to rivers/canals), and air quality (rising PM2.5), underscoring social-ecological risks including reduced agricultural productivity, health burdens, and increased vulnerability to climate-related flooding. Interventions such as converting traditional kilns to more efficient Zig-Zag technologies can substantially reduce emissions and fuel use, improve brick quality, and potentially enhance governance and visibility of labour conditions. However, short-term cost increases may heighten vulnerability to exploitation in some contexts, necessitating complementary social safeguards and monitoring. The EO-based approach is readily triangulated with event-, standards-, and survey-based slavery metrics, focusing scarce ground resources where risks are highest. The dataset already supports operational tools (e.g., UNDP India Accelerator Lab mobile app) for coordinated action. Partnerships and systematic EO monitoring can strengthen integrated SDG measurement and accelerate progress toward 8.7.

This study delivers the first contiguous, high-resolution map and construction chronology of all Bull’s Trench brick kilns across South Asia’s Brick Belt, linking kiln proliferation to urban demand and quantifying environmental externalities (soil, water, air). The updateable EO- and ML-based framework provides a robust baseline for targeting anti-slavery interventions, informing environmental governance, and monitoring SDG interdependencies centered on SDG 8.7. The authors advocate establishing an Anti-Slavery Digital Observatory to integrate satellite data with novel and traditional data streams for global, repeatable analyses of slavery-linked sites and activities (e.g., kilns, fishing fleets, logging, mining). Scaling this work will benefit from cross-sector partnerships to access VHR imagery, reduce costs, and integrate ground data for validation and richer attribution. Future research should refine kiln typology (e.g., active vs. abandoned; technology class), improve temporal attribution of construction and operations, attribute pollutant sources more precisely, and deepen analyses of social-ecological feedbacks driving vulnerability, thereby supporting evidence-based pathways to eradicate modern slavery and mitigate environmental harms.

The approach does not directly observe or quantify slavery; kiln presence is a proxy for elevated risk. VHR imagery availability is uneven across time and space, constraining precise construction dating and activity status; Landsat-based breakpoints can reflect other land cover changes, and the RF model, while validated, has dating uncertainty (~1.8 years). Detection errors remain (recall 87%), potentially missing or misclassifying kilns or similar structures. The analysis does not fully disaggregate kiln technologies (e.g., Zig-Zag vs. traditional) or distinguish active vs. abandoned kilns, limiting environmental attribution. PM2.5 trends cannot be attributed solely to kilns given multiple emission sources. Resource extraction estimates use literature-based coefficients and conservative production assumptions, introducing uncertainty. NTL as a proxy for demand/population has known saturation and calibration issues and covers 1996–2013 only. Access to VHR data can be cost-prohibitive for routine monitoring, and limited ground truth data reduces the ability to validate at scale.