A big data approach to assess progress towards Sustainable Development Goals for cities of varying sizes

The paper addresses how to quantitatively assess city-level progress toward all 17 SDGs across cities of different sizes, especially where statistical data are scarce. While many SDG assessments exist at global, regional, national and subnational scales, comprehensive city-level evaluations covering all SDGs are limited and often focus on large cities because small and medium cities lack comparable statistics. The authors argue that cities are central to achieving the SDGs, yet face pressing challenges (inequalities, pollution, infrastructure deficits) and need data-driven tools to set baselines and track progress. The study aims to develop and test a systematic approach that leverages open-source big data and machine learning to estimate city-level SDG Index scores, validate these against scores computed using statistical indicators for 254 Chinese cities, and then extend the method to 83 additional cities lacking relevant statistical data. The goal is to provide an efficient, scalable, and low-cost means to monitor SDG progress and to illuminate disparities by city size, geography, and income level, thereby informing evidence-based policy.

Prior work highlights the SDG Index as a useful composite for comparing overall SDG performance, but city-level voluntary local reviews have often lacked quantitative baselines and progress metrics. Big data has emerged as a promising resource for monitoring specific SDG indicators and targets, with applications including nighttime lights for economic growth and poverty mapping, POI and OpenStreetMap data for infrastructure and urban form, and machine learning models (random forest, boosted trees, ANNs) to improve indicator estimation. However, most studies monitor a subset of indicators for single SDGs rather than providing comprehensive, multi-SDG city-level assessments. There remains a need to integrate multisource big data and ML to evaluate overall SDG performance across cities, particularly where official statistics are missing or delayed.

Study design: The authors conducted a two-stage assessment for Chinese prefecture-level and above cities in 2017. Stage 1 computed individual SDG scores and an SDG Index for 254 cities using statistical indicators; Stage 2 built a big-data-based indicator system and trained an artificial neural network (ANN) model to predict SDG Index scores from big data, then applied it to 83 additional cities lacking statistics, yielding coverage of 337 cities. Stage 1 — Statistical indicator framework and SDG scoring: A localized framework of 54 statistical indicators (average ~3 per SDG) was developed by combining the UN global SDG indicator list with prior SDG Index reports and China-focused studies, emphasizing relevance to urban challenges and data availability at city level. For each indicator, upper and lower bounds were set to control extremes: fixed technical/ideal values for some indicators (e.g., gender equality, hazardous waste), the average of the top 5 performers for others as the upper bound, and the 2.5th percentile as the lower bound. Indicators were normalized to a 0–100 scale where higher is better; values beyond bounds were capped at 0 or 100. Within each SDG, indicators were equally weighted, and the 17 SDGs were equally weighted to compute each city’s SDG Index as the arithmetic mean of SDG scores. Uncertainty and sensitivity analyses were performed (details in Supplementary Methods). Stage 2 — Big data indicators and ANN model: Multisource open big data were compiled, including remote sensing and geospatial datasets: nighttime lights (EOG VIIRS), land use (Geographical Information Monitoring Cloud), road networks (OpenStreetMap), POIs (Amap), company information (Tianyancha), gridded population (WorldPop), and others. Candidate big data indicators were derived for each SDG and correlated (Pearson) with the statistically-derived individual SDG scores across 254 cities. Eighteen monitoring indicators with significant correlation (p < 0.05) to their target SDGs were selected to form a generic big data indicator system (examples: NIC—nighttime light intensity on construction land for SDG 8; DME—density of manufacturing enterprises on construction land and DRE—density of research enterprises for SDG 9; SDG 15 used prior indicators). These indicators were also partially correlated with other SDGs, reflecting cross-linkages. ANN training and configuration: Using the 254-city dataset, the 18 big data indicators served as inputs and the statistical SDG Index scores as the expected outputs. Data were randomly split into training (70%), validation (15%), and test (15%) sets. A backpropagation neural network with the Levenberg–Marquardt optimization was implemented in MATLAB (R2021a). After tuning, the final architecture comprised an 18-neuron input layer, two hidden layers with 9 neurons each, and a 1-neuron output layer. Performance was evaluated using R, RMSE, and R-squared on train/validation/test subsets. The trained model was then applied to estimate SDG Index scores for 83 cities with insufficient statistical data. Data sources: Statistical data came from national population surveys, city statistical communiques (2017), China City Statistical Yearbooks (2016–2018), and China Urban Construction Statistical Yearbooks (2016–2018). Carbon emissions: CEADs. Government performance: financial transparency and political–business relations reports. Marine/coastal water quality bulletins. Big data sources: EOG nighttime lights, OpenStreetMap roads, Amap POIs, Tianyancha company data, WorldPop population, land use from DSAC platform. Supplementary materials and data are available via figshare (DOI provided).

- Individual SDG performance (254 cities with statistics): Large cities generally scored higher across most SDGs but lagged on SDG 10 (reduced inequalities) and SDG 17 (partnerships). Small and medium cities performed relatively better on SDG 15 (life on land) but still with low scores. Cities performed well in the planet dimension, notably SDG 6 (clean water) and SDG 12 (responsible consumption/production), with averages exceeding 77. People/prosperity dimensions showed strengths in SDG 5 (gender equality) and SDG 7 (energy), but weaknesses in SDG 3 (health), SDG 9 (infrastructure/innovation), and SDG 10 (inequalities). - Regional patterns: Eastern cities outperformed most SDGs but scored lowest on SDG 14 (life below water) due to poor coastal water quality; Eastern cities also scored high on SDG 17. Central cities performed better on energy-related SDGs (7, 12). Northeastern cities led SDG 2 (zero hunger) due to fertile agricultural lands. Western cities scored lower due to limited ecological assets (SDG 15), educational facilities (SDG 4), and external connectivity (SDG 17). - Big data indicators validity: Eighteen big data indicators were significantly correlated with their target SDGs (p < 0.05). Examples: NIC (nighttime light intensity on construction land) correlated with SDG 8 score at r = 0.621; for SDG 9, DME and DRE correlated at r = 0.462 and r = 0.574, respectively. SDG 15 used existing indicators. - ANN performance: Using 18 big data indicators, the ANN achieved test RMSE = 3.13 and R^2 = 0.7625, indicating strong predictive ability for SDG Index scores. Overall correlation (all sets) R ≈ 0.90. Compared to collecting 54 statistical indicators, the big-data-plus-ANN approach provided comparable accuracy with greater efficiency and availability. - Group-wise ANN accuracy (Table 1): Large cities MAE 1.676, RMSE 2.270, R 0.899; medium MAE 1.809, RMSE 2.256, R 0.822; small MAE 2.000, RMSE 2.562, R 0.852. Eastern MAE 1.935, RMSE 2.446, R 0.845; central MAE 1.765, RMSE 2.371, R 0.851; western MAE 1.651, RMSE 2.133, R 0.889; northeastern MAE 2.116, RMSE 2.615, R 0.737. By income: high-income MAE 1.370, RMSE 1.988, R 0.939; upper-middle MAE 1.919, RMSE 2.418, R 0.815; lower-middle MAE 2.002, RMSE 2.562, R 0.890. - City-level SDG Index (337 cities via ANN): Higher scores concentrated in Eastern China, especially coastal Guangdong, Zhejiang, and Jiangsu; lower scores in Western China, notably Tibet and Qinghai. Central provincial capitals (Changsha, Wuhan, Zhengzhou) ranked in the top 15. Liaoning outperformed Heilongjiang in the Northeast. Provincial capitals generally exceeded other prefecture-level cities in their provinces. - Distribution by group (337 cities): Median SDG Index by region—East 50.13, Central 46.90, Northeast 43.77, West 41.31. By size—average SDG Index: large 49.66, medium 46.50, small 42.10; medium cities’ distribution more balanced, small cities skewed lower. By income—high-income cities’ median > 50; lower-middle-income median < 40; lower tiers clustered at and below Q3. Overall, large, high-income, eastern cities had the highest average and median (53.92 and 54.42, respectively).

The study demonstrates that open-source big data, coupled with an ANN model, can effectively approximate city-level SDG Index scores, offering a practical alternative where official statistics are limited or delayed. This addresses the core challenge of monitoring SDG progress across cities of varying sizes. Findings reveal systemic disparities: smaller and western cities lag behind larger and eastern cities, with pronounced gaps in SDG 3 (health), SDG 9 (infrastructure/innovation), and SDG 17 (partnerships). The results suggest targeted policy interventions, including enhancing infrastructure connectivity, fostering innovation ecosystems, and strengthening inter-city and international partnerships. The approach enables timely, cost-effective monitoring and can guide allocation of resources and prioritization of lagging SDGs at subnational scales. The authors advocate developing a standardized big data monitoring platform for consistent, automated assessments, and combining big-data-based overall assessments with detailed statistical analyses where available.

The paper contributes a scalable big-data-plus-ANN methodology to assess overall SDG performance across cities, validated against statistical measures and applied nationally to 337 Chinese cities. It evidences systematic differences in SDG Index scores by city size, region, and income—large, eastern, high-income cities lead; small, western, lower-income cities lag. To promote balanced sustainability, integrated socioeconomic communities (metropolitan areas, urban agglomerations, economic belts) can leverage large-city advantages to uplift surrounding small and medium cities through industrial collaboration and improved connectivity. Future directions include: (1) increasing investments in SDG data and evaluation systems to expand and harmonize city-level indicators and reduce reporting lags; and (2) modeling SDG interactions (trade-offs and synergies) to move beyond simple averaging, integrating inter-goal dynamics into assessment frameworks.

- Data coverage: Not all 17 SDGs could be measured with suitable big data indicators; the city-level statistical indicator framework, though broad (54 indicators), still cannot fully capture all aspects of SDG progress. - Comparability and resolution: Big data sources vary in spatiotemporal resolution and localization, which can reduce comparability and certainty across cities and over time. - Statistical data gaps: Many health-related outcome indicators (e.g., disease prevalence for SDG 3) are unavailable at city level, restricting validation and detail; other indicators are inconsistently reported across city yearbooks. - Temporal lag vs. timeliness: Official statistics lag by at least a year, while big data are more current; discrepancies in timing may affect alignment between training targets and big data inputs. - Generalizability: The ANN was trained on Chinese cities (2017); application to other countries requires adaptation of indicators and retraining with local statistical baselines. - Model transparency: While ANN performance is strong, interpretability is limited relative to transparent statistical aggregation methods.