PAK-001

High-Resolution Rural Poverty Mapping for Sindh – Research-to-Policy Pilot

Download PDF
Pakistan Middle East, North Africa, Afghanistan & Pakistan Lower middle income Pilot / Controlled Trial Phase Likely

Strategic Social Protection Unit (SPSU), Government of Sindh; University of Bristol (School of Geographical Science); University of Manchester (Department of Planning and Environmental Management); University of Qatar (Social and Economic Survey Research Institute)

At a Glance

What it does Prediction (including forecasting) — Vulnerability, needs and risk assessment, including predictive analytics
Who runs it Strategic Social Protection Unit (SPSU), Government of Sindh; University of Bristol (School of Geographical Science); University of Manchester (Department of Planning and Environmental Management); University of Qatar (Social and Economic Survey Research Institute)
Programme Sindh Social Protection Strategic Unit – High-Resolution Rural Poverty Mapping Pilot
Confidence Likely
Deployment Status Pilot / Controlled Trial Phase
Key Risks Data-related risks
Key Outcomes Ensemble model achieved 71% recall and 67% precision on hold-out test; 59–81% recall across 6-fold spatial cross-validation; 59% recall in 2022 ground-truth validation (vs 50% random baseline).
Source Quality 4 sources — Report (multilateral / development partner), Academic journal article

The High-Resolution Rural Poverty Mapping initiative in Sindh Province, Pakistan, is a research-to-policy pilot that uses ensemble deep learning applied to satellite imagery to generate 1 km² resolution poverty predictions for rural areas. The project was developed by an academic research team comprising Felix Agyemang (University of Manchester), Rashid Memon (University of Qatar), Levi John Wolf and Sean Fox (University of Bristol), in collaboration with the Strategic Social Protection Unit (SPSU) of Pakistan's Sindh Province. Funded by the Center for Effective Global Action's (CEGA) Targeting Aid Better Initiative at UC Berkeley, the initiative was motivated by the twin shocks of the Covid-19 pandemic (March 2020) and one of the wettest monsoon seasons since 1961, which devastated rural households in Sindh and created an urgent need to identify communities eligible for emergency cash relief without relying on costly and time-consuming household surveys.

The system employs a transfer learning approach using three convolutional neural network (CNN) architectures — ResNet-50, ResNet-50V2, and ResNet-101 — initialised with ImageNet weights and trained end-to-end on three concatenated input streams. The primary data inputs are Sentinel-2 daytime satellite imagery (10 m resolution, captured January–April 2016), VIIRS nighttime lights data (2016 median annual product, resampled from 500 m to 10 m), and a global accessibility layer measuring travel time to settlements of 5,000–10,000 population. The three CNN models are combined into an ensemble using a modal (majority vote) classification approach, producing binary predictions at the 1 km² grid cell level: cells with median Simple Poverty Scorecard (SPS) scores below 19 are classified as chronically poor, while those at or above 19 are classified as not chronically poor.

The training data comprised approximately 1.67 million anonymised household-level poverty score records (from an original collection of approximately 1.9 million), gathered through computer-assisted personal interviews (CAPI) on Android tablets between 2016 and 2019. These records were collected under two existing Sindh government programmes: the Sindh Union Council Economic Strengthening Support (SUCCESS) programme, covering 8 districts, and the People's Poverty Reduction Program (PPRP), covering 6 additional districts — together spanning 14 of Sindh's 24 districts. The SPS methodology, originally adopted by the Government of Pakistan as the targeting tool for the Benazir Income Support Program (BISP), its flagship cash transfer programme since January 2009, assigns households a score from 0 to 100, where higher scores indicate less deprivation.

The research team implemented a rigorous three-stage validation framework. First, a hold-out test on randomly sampled data showed the ensemble model achieving 71% recall, 67% precision, 69% overall accuracy, and an AUC of 0.69, outperforming a comparable Uganda study (66% recall, 44% precision). Second, six-fold spatial cross-validation — where entire districts were held out in turn — demonstrated ensemble recall of 59–81% and median AUC of 0.66, with the ensemble consistently performing as the most stable predictor across iterations. Third, a novel ground-truth validation exercise was conducted in Ghotki district in 2022, surveying approximately 7,000 households across 174 grid cells, where the ensemble model achieved 59% recall versus 50% for random classification. The models were trained using Keras and TensorFlow in Python with a batch size of 16, learning rate of 0.00005, 30 epochs, and early stopping after 10 consecutive epochs without validation improvement.

The resulting poverty maps cover 32,281 grid cells across rural Sindh at 1 km² resolution, representing a tenfold improvement in spatial granularity over prior geographic targeting models. The SPSU intended to use these maps to develop a targeting strategy for social welfare interventions, particularly emergency cash relief. However, the SPSU subsequently launched a USD 230 million initiative with the World Bank — the Strengthening Social Protection Delivery System in Sindh (SSPDSS) project — that relies on a Multidimensional Poverty Index (MPI) for geographic targeting instead of the ML-generated poverty prevalence estimates (CEGA, 2023b; World Bank, 2022). No primary evidence has been identified confirming that the poverty maps were operationally adopted into government targeting workflows or that they directly informed benefit allocation decisions. The project remains characterised as a research-to-policy initiative with demonstrated technical feasibility but no documented policy uptake.

The study received ethical approval from the Institutional Review Board at Lahore University of Management Sciences (LUMS-IRB/05/30/2022/RM-FWA-00019408). Verbal informed consent was obtained from all households using a standardised consent script, and all data was aggregated to 1 km² cells to prevent individual household identification. All data files and scripts were made publicly available on Figshare. Key limitations acknowledged include substantial spatial noise in the original GPS-referenced survey data, the asset-based SPS measure differing from the official consumption-based poverty line, geographic coverage limited to 14 of 24 districts, and temporal separation between 2016 training imagery and 2022 validation data.

Classifications follow the DCI AI Hub Taxonomy. Hover over field labels for definitions.

Social Protection Functions

Implementation/delivery chain
Assessment of needs/conditions + enrolment primary
SP Pillar (Primary) The social protection branch: social assistance, social insurance, or labour market programmes. Social assistance
Programme Name Sindh Social Protection Strategic Unit – High-Resolution Rural Poverty Mapping Pilot
Programme Type The type of social protection programme, classified under social assistance, social insurance, or labour market programmes. View in glossary Poverty targeted Cash Transfers (conditional or unconditional)
System Level Where in the social protection system the AI is applied: policy level, programme design, or implementation/delivery chain. View in glossary Implementation/delivery chain
Programme Description Research-to-policy pilot using ensemble deep learning on satellite imagery to produce 1 km² resolution rural poverty maps for Sindh Province. Developed to support the Strategic Social Protection Unit (SPSU) in geographic targeting for emergency cash relief during Covid-19 and monsoon flood response, building on existing household poverty score data from the SUCCESS and PPRP programmes.
Implementation Type How the AI output is produced: Classical ML, Deep learning, Foundation model, or Hybrid. Affects validation, compute requirements, and governance profile. View in glossary Deep learning
Lifecycle Stage Current stage in the AI lifecycle, from problem identification through to monitoring, maintenance and decommissioning. View in glossary Model Selection and Training
Model Provenance Origin of the AI model: developed in-house, adapted from open-source, commercial/proprietary, or accessed via third-party API. View in glossary Developed in-house
Compute Environment Where the AI system runs: on-premise, government cloud, commercial cloud, or edge/device. View in glossary Not documented
Sovereignty Quadrant Classification of data and compute sovereignty: I (Sovereign), II (Federated/Hybrid), III (Cloud with safeguards), or IV (Shared Innovation Zone). View in glossary Not assessed
Data Residency Where the data used by the AI system is stored: domestic, regional, or international. View in glossary Not documented
Cross-Border Transfer Whether data crosses national borders, and if so, whether documented safeguards are in place. View in glossary Not documented
Decision Criticality The rights impact of the decision the AI supports. High criticality requires HITL oversight; moderate requires HOTL; low may operate HOOTL. View in glossary Moderate
Human Oversight Type Level of human involvement: Human-in-the-Loop (active review), Human-on-the-Loop (monitoring), or Human-out-of-the-Loop (periodic audit). View in glossary HITL
Development Process Whether the AI system was developed fully in-house, through a mix of in-house and third-party, or fully by an external provider. View in glossary Fully in-house
Highest Risk Category The most significant structural risk source identified: data, model, operational, governance, or market/sovereignty risks. View in glossary Data-related risks
Risk Assessment Status Whether a formal risk assessment, informal assessment, or independent audit has been conducted for this system. Informal assessment
  • Bias audit
  • Data minimisation controls
  • Independent evaluation
CategorySensitivityCross-System LinkageAvailabilityKey Constraints
Administrative data from other sectorsNon-personalLinks data across multiple systemsCurrently available and usedGlobal accessibility layer (travel time to settlements of 5,000–10,000 population); publicly available
Geospatial and remote sensing dataNon-personalLinks data across multiple systemsCurrently available and usedSentinel-2 imagery (10 m resolution, Jan–Apr 2016) and VIIRS nighttime lights (2016 median annual); temporal mismatch between imagery date and poverty score collection period (2016–2019); satellite data is publicly available
Survey and census dataPersonalLinks data across multiple systemsCurrently available and usedSimple Poverty Scorecard data from SUCCESS and PPRP programmes covering 1.67 million households across 14 of 24 Sindh districts; substantial spatial noise in GPS coordinates; asset-based scoring differs from consumption-based official poverty line; collected 2016–2019 via CAPI on Android tablets

Center for Effective Global Action (CEGA) (2023) 'Combining satellite imagery and machine learning to target social protection in Pakistan'. Berkeley: UC Berkeley. Available at: https://cega.berkeley.edu/collection/ml-targeting-sindhpakistan/ (Accessed: 23 March 2026).

View source Report (multilateral / development partner)

Center for Effective Global Action (CEGA) (2023b) 'Using Ensemble Deep Learning to Deliver Aid Better in Post-Flood Pakistan', CEGA Blog, 15 May. Available at: https://cega.berkeley.edu/article/using-ensemble-deep-learning-to-deliver-aid-better-in-post-flood-pakistan-2/ (Accessed: 24 March 2026).

View source Report (multilateral / development partner)

Agyemang, F. S. K., Memon, R., Wolf, L. J. and Fox, S. (2023) 'High-resolution rural poverty mapping in Pakistan with ensemble deep learning', PLOS ONE, 18(4), e0283938. Available at: https://doi.org/10.1371/journal.pone.0283938 (Accessed: 23 March 2026).

View source Academic journal article

World Bank (2022) 'Factsheet: Strengthening Social Protection Delivery System in Sindh', 19 December. Available at: https://www.worldbank.org/en/news/factsheet/2022/12/19/factsheet-strengthening-social-protection-delivery-system-in-sindh (Accessed: 24 March 2026).

View source Report (multilateral / development partner)
Deployment Status How far the system has progressed into real-world operational use, from concept/exploration through to scaled and institutionalised. View in glossary Pilot / Controlled Trial Phase
Year Initiated The year the AI system was first initiated or development began. 2020
Scale / Coverage The scale and geographic or population coverage of the deployment. 14 of 24 Sindh districts; 32,281 grid cells at 1 km² resolution; trained on 1.67 million household records; ground-truth validation with ~7,000 households in Ghotki district
Funding Source The source(s) of funding for the AI system development and deployment. Center for Effective Global Action (CEGA), UC Berkeley – Targeting Aid Better Initiative
Technical Partners External technology vendors, academic partners, or development partners involved. University of Bristol; University of Manchester; University of Qatar; Sindh Rural Support Organization (SRSO)
Outcomes / Results Ensemble model achieved 71% recall and 67% precision on hold-out test; 59–81% recall across 6-fold spatial cross-validation; 59% recall in 2022 ground-truth validation (vs 50% random baseline). Produced 1 km² poverty maps for 32,281 rural grid cells across Sindh. Tenfold improvement in spatial granularity over prior geographic targeting models. SPSU subsequently adopted World Bank MPI-based targeting for the USD 230M SSPDSS project rather than the ML poverty maps (CEGA, 2023b; World Bank, 2022).
Challenges Substantial spatial noise in GPS-referenced household survey data; asset-based Simple Poverty Scorecard differs from official consumption-based poverty line; geographic coverage limited to 14 of 24 Sindh districts; temporal gap between 2016 training imagery and 2022 ground-truth validation; mediocre precision attributed to inclusion errors in underlying SPS data; SPSU chose World Bank MPI-based targeting over ML poverty maps for operational use (CEGA, 2023b).

How to Cite

DCI AI Hub (2026). 'High-Resolution Rural Poverty Mapping for Sindh – Research-to-Policy Pilot', AI Hub AI Tracker, case PAK-001. Digital Convergence Initiative. Available at: https://socialprotectionai.org/use-case/PAK-001 [Accessed: 1 April 2026].

Change History

Created 30 Mar 2026, 08:41
by v2-import (import)