PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning for Socioeconomic Estimation in Forced-Displacement Settings

Steven Ndung'u, Adel Daoud, Ismael Yacoubou Djima, Hai-Anh H. Dang, Patrick Michael Brock

arXiv:2609.15773v1cs.LGcs.AI

TL;DR

Periodic surveys may miss changes between rounds, while tools developed for general populations remain underexplored in forced-displacement settings. This study adapts a DHS-pretrained spatiotemporal vision transformer using FDS and RMS indices; after fine-tuning, it explains up to 66% of variation in camp-intersecting grids.

  • Problem

    Tools developed for general populations have seen limited adaptation and evaluation in forced-displacement settings, where socioeconomic conditions and settlement patterns differ.

    Source: §1.1 · p. 3, §5 · p. 10

  • Method

    The study partially fine-tunes a DHS-pretrained spatiotemporal vision transformer using FDS/RMS socioeconomic indices and Landsat, VIIRS, and building-derived inputs.

    Source: §4.1 · p. 9, §4.3 · p. 10, §5 · p. 11

  • Results

    Fine-tuning yields up to 66% explained variation in camp-intersecting grids, while the zero-shot model has near-zero or negative R2.

    Source: §2 · p. 6

  • Takeaways & Limitations

    Spatially granular, regularly updated estimates can extend periodic surveys between rounds and help screen for change requiring field verification.

    Source: §5 · p. 11

    Without later survey waves as independent test data, the hindcasting and nowcasting applications lack temporal validation.

    Source: §2 · p. 8

Abstract

from arXiv · show

Progress in inclusive household surveys has strengthened socioeconomic evidence for forcibly displaced populations, providing indispensable benchmarks on living conditions and welfare. However, these surveys remain resource-intensive and periodic, while conditions can change between rounds, particularly in settings affected by fragility, conflict, and violence. More frequently updated, spatially granular complementary evidence is therefore needed to identify where socioeconomic conditions may be changing between survey rounds and to inform operational prioritization. Earth observation and machine learning offer a scalable source of spatially explicit socioeconomic information. However, tools developed for general populations have not been systematically adapted and evaluated in forced displacement settings, where living conditions, settlement patterns, and displacement impacts may differ substantially. We address this gap by adapting a multimodal spatiotemporal vision transformer, pretrained on Demographic and Health Survey data from approximately 1.2 million households across 36 African countries, to forced displacement and host community settings in South Sudan, Cameroon, and Zambia. We develop and evaluate the updated, adapted model using socioeconomic indices derived from UNHCR FDS and RMS data. Our results show that satellite-derived geospatial covariates explain up to 66% of the variation in socioeconomic outcomes in camp-intersecting grids, with a mean absolute error (MAE) of 4.37 index points, and 41% in non-camp-intersecting areas, with an MAE of 5.41. The framework complements and adds value to periodic household surveys by filling critical spatial and temporal data gaps with regularly updated, model-based socioeconomic estimates. These estimates sustain insight between survey rounds and support timely humanitarian prioritization and field verification.

1. Results

The results characterize socioeconomic-index differences across survey sources and regions, evaluate model performance across settlement contexts, and demonstrate hindcasting and nowcasting in South Sudan.

1.1. Socioeconomic index.

  • RMS adds fine-tuning coverage despite lacking asset variables; displacement indices cluster lower than DHS, a broad distributional difference rather than a cardinal welfare comparison.The RMS index uses variables aligned with FDS, while the FDS index includes asset ownership.Source: §1.1 · p. 3
  • Regions around Yaoundé, Douala, and Lusaka have higher and wider index distributions than other areas.The study suggests greater access to labor markets, services, and infrastructure may contribute to this urban pattern.Source: §1.1 · p. 3
  • Mayukwayukwa, Meheba, and Mantapala have similar mean index values of 17.3, 17.6, and 15.9, with considerably overlapping distributions.Compared with broader urban and rural variation, the three Zambian camp settings occupy a comparatively low, compressed portion of the index distribution.Source: §1.1 · p. 3

1.2. Model performance.

  • Fine-tuning achieves pooled R2 0.40 and MAE 5.49 index points, whereas zero-shot predictions have near-zero or negative R2.The zero-shot model's narrow predictions indicate that its DHS-calibrated mapping requires domain-specific fine-tuning.Source: §1.2 · p. 3
  • Across countries, test R2 is 0.62 in South Sudan, 0.44 in Zambia, and 0.28 in Cameroon.The corresponding MAEs are 3.55, 4.53, and 5.48 index points, respectively.Source: Table 1 · p. 5, §1.2 · p. 5
  • The model performs better in camp-intersecting grids, attaining R2 0.66 and mean absolute error 4.37 index points, versus R2 0.41 and 5.41, respectively, in host community grids.Differences in outcome variability, sample composition, and household density may partly explain this pattern.Source: §1.2 · p. 5

1.3. Using the model: hindcasting and nowcasting.

  • With the fine-tuned model fixed, South Sudan estimates show a 5.9-point mean-index decline from 2023, including a 12 percent decline by 2024 and a cumulative 27 percent decline by 2025.The direction and geography of the shift are described as consistent with consequences of the war in Sudan and its cross-border spillover.Source: §1.3 · p. 6
  • These projections identify areas of possible socioeconomic change for field verification before follow-up survey rounds.The model is applied to later imagery to produce updated estimates where survey data are unavailable.Source: §1.3 · p. 6

2. Discussion

Displacement-specific adaptation makes national-survey representations useful for socioeconomic estimation in forced-displacement settings, while temporal projections and their validation remain bounded. The estimates are intended to complement surveys by helping identify areas for further assessment and field verification.

2. Discussion

  • After FDS and RMS fine-tuning, geospatial covariates explain up to 66% of socioeconomic variation in camp-intersecting grids, while zero-shot predictions have near-zero or negative R2.The result is the paper's main evidence that representations learned from national populations retain useful information after displacement-specific adaptation.Source: §2 · p. 6
  • South Sudan's model-predicted mean index declines by 5.9 points from 2023 to 2025, but these unvalidated projections are not observed welfare changes.The model is held fixed while Earth observation inputs vary.Source: Figure 4 · p. 7
  • Camp-intersecting performance is broadly comparable to related EO–ML wealth-estimation studies, although their larger, nationally representative samples and different evaluation designs limit direct comparison.The cited benchmarks report R2 = 0.72 across held-out countries and R2 = 0.75 across held-out time spans; the study reports R2 = 0.66 for spatial variation in camp-intersecting grids.Source: §2 · pp. 6–7
  • The estimates are best used to flag possible deterioration and prioritize assessment and field verification, rather than as official poverty or welfare estimates.The model is intended to complement household surveys, and future survey data are needed to test whether predicted changes match measured changes.Source: §2 · p. 7
  • The evaluation is spatially blocked but only partially blocked in time, because the surveys lack repeated measurements at the same locations for testing predicted temporal changes.Later FDS and RMS rounds could support fully blocked spatiotemporal evaluation using independent test data.Source: §2 · pp. 7–8

3. Data

The study links harmonized survey-derived socioeconomic outcomes in South Sudan, Cameroon, and Zambia with satellite and building features on a common spatial grid. Its data combine displacement-focused survey sources with imagery and spatially augmented training examples.

3. Data

  • The dataset links socioeconomic outcomes from three forced-displacement contexts in sub-Saharan Africa to remotely sensed settlement and environmental characteristics.The data section describes the survey reference data, satellite imagery, and building-derived features used in the study.Source: §3 · p. 8

3.1. Survey reference data.

  • UNHCR's standardized FDS provides comparable household data on living standards, well-being, services, and vulnerability; South Sudan RMS adds coverage focused primarily on internally displaced people.Combining FDS and RMS increases the analysis's spatial, temporal, and demographic coverage.Source: §3.1 · p. 8

3.2. Satellite imagery.

  • Landsat daytime imagery and VIIRS night-time lights provide complementary environmental and activity-related features aligned to survey-year observations on a 6.72 km × 6.72 km grid.Landsat bands capture vegetation, water, soil, and built-environment characteristics; VIIRS radiance is used as a proxy for artificial lighting and economic activity, not as a direct welfare measure.Source: §3.2 · p. 8

3.3. Building footprints and settlement morphology.

  • Building presence, height, and fractional count add settlement-morphology features absent from the pretrained model's original inputs.These features complement Landsat land-surface information and VIIRS activity signals.Source: §3.3 · pp. 8–9

3.4. Integrating satellite data with survey data.

  • Survey locations are aggregated to fixed grid cells, with a five-household minimum, yielding 211 South Sudan, 238 Cameroon, and 159 Zambia observations.The common grid preserves household privacy and matches the model input structure.Source: §3.4 · p. 9
  • Sliding spatial windows increase the dataset to 1,123 grids while keeping windows from each original grid in the same partition to prevent spatial leakage.The shifted windows provide spatially perturbed examples around densely populated settlements.Source: §3.4 · p. 9

4. Methodology

The methodology constructs survey-based socioeconomic indices, uses spatially disjoint validation regions, and adapts a DHS-pretrained multimodal spatiotemporal transformer to estimate welfare from satellite time series and auxiliary geospatial features.

4.1. Construction of the socioeconomic index.

  • The 0–100 socioeconomic index combines asset, housing, services, food-security, and land indicators; the South Sudan RMS version uses harmonized housing, services, and food-consumption variables.Preprocessing reduces score heaping before principal component analysis; the first component is rescaled so lower scores indicate poorer households.Source: §4.1 · p. 9

4.2. Experimental setup.

  • Fourfold validation holds out geographically distinct regions, but distributes observations across months and survey years randomly rather than testing a fully independent future survey round.Training, validation, and test regions are spatially disjoint, with the test set reserved for final evaluation.Source: §4.2 · p. 10

4.3. Model architecture.

  • Landsat time series are spatially encoded and temporally aggregated, while VIIRS radiance and building features supplement the regression head’s grid-level index prediction.Each Landsat observation is a 6.72 km × 6.72 km tile.Source: §4.3 · p. 10

4.4. Transfer learning and fine-tuning.

  • The model retains pretrained patch embeddings and early spatial layers while fine-tuning remaining spatial layers, the temporal transformer, auxiliary components, and regression head on FDS and RMS data.Pretraining used approximately 1.2 million households from 140 DHS surveys across 36 African countries; the number of frozen early layers was selected as a hyperparameter.Source: §4.4 · p. 10

4.5. Performance metrics for evaluation.

  • Evaluation uses MAE, mean squared error, root mean squared error, R2, and Pearson correlation to assess prediction error and goodness of fit.The metrics compare predicted socioeconomic indicators against ground-truth measurements.Source: §4.5 · p. 10

5. Conclusion

The study adapts DHS-pretrained representations to estimate socioeconomic welfare in displacement settings, addressing the limited adaptation of EO-ML methods to refugee and host-community settlements. It presents the framework as a screening aid between survey rounds, not a replacement for household surveys.

5. Conclusion

  • DHS-pretrained representations retain predictive value after partial fine-tuning for forcibly displaced populations in South Sudan, Cameroon, and Zambia.The conclusion frames this as proof-of-concept evidence.Source: §5 · p. 11
  • New satellite imagery supports spatially granular, regularly updated estimates of how selected indicators vary across locations and change over time.These estimates extend periodic surveys between rounds or when follow-up surveys are not feasible.Source: §5 · p. 11
  • The framework is most useful for screening where medium-term conditions may be changing and where further assessment or field verification is warranted.Household surveys remain the welfare benchmark and are essential for periodic model recalibration.Source: §5 · p. 11

Data, Materials, and Software Availability.

The paper reports that its FDS and RMS microdata are available through the UNHCR Microdata Library, subject to access requirements.

Data, Materials, and Software Availability.

  • FDS and RMS microdata are available through the UNHCR Microdata Library, subject to application, approval, and applicable data-protection requirements.The paper also states that its findings and conclusions are the authors’ own and do not necessarily represent the views of the World Bank Group or UNHCR.Source: Data, Materials, and Software… · p. 11

5 CONCLUSION

The listed references cover socioeconomic-index construction, earth-observation data and features, and methods for spatially structured validation.

5 CONCLUSION

  • References 30–32 cover deriving an asset-based wealth index from refugee and host-community survey data, DHS wealth-index approaches for rural and urban areas, and the International Wealth Index.Source: §5 · p. 12
  • References 33–36 address Landsat data, VIIRS night-time lights, and satellite-based building detection.These entries concern open Landsat data, global land observations, night-time lights, and building detection.Source: §5 · p. 12
  • The bibliography also lists work on spatially structured cross-validation and regression dilution bias.These are references 43 and 37, respectively.Source: §5 · p. 12

6. Appendix

The appendix labels a zero-shot performance table, presents socioeconomic-index materials for three refugee camps in Zambia, and illustrates sliding-window augmentation.

6. Appendix

  • Table 4 is identified as reporting zero-shot model performance.Source: Table 4 · p. 13
  • Table 5 presents socioeconomic-index information across Mayukwayukwa, Meheba, and Mantapala refugee camps in Zambia.Source: Table 5 · p. 13
  • Figure 6 presents the socioeconomic-index distribution across Mayukwayukwa, Meheba, and Mantapala refugee camps in Zambia.Source: Figure 6 · p. 13
  • Figure 7 illustrates sliding-window augmentation, with shifted extraction windows shown relative to the original window.The red dot marks the center of the original grid.Source: Figure 7 · p. 13

Related papers

Loading 2609.15773v1…