A satellite foundation model for improved wealth monitoring

Abstract: Poverty statistics guide social policy, but in many low- and middle-income countries, censuses and household surveys that collect these data are costly, infrequent, quickly outdated, and sometimes error-prone. Satellite imagery offers global coverage and the possibility of predicting economic livelihoods at scale, yet existing approaches to predicting livelihoods with imagery or other non-traditional data often fail to reliably identify local-level variation and, as we show, degrade under temporal shift. Here we introduce Tempov, a satellite foundation model pretrained by self-supervision on three million bi-temporal Landsat pairs and adapted with parameter-efficient fine-tuning to sparse survey labels. The model enables large-scale, high-resolution wealth mapping and dynamic measurement, including zero-shot nowcasting up to a decade after observed labels, retrospective hindcasting, and decadal change tracking, while outperforming existing neural network and geospatial foundation-model baselines. In low-label regimes, Tempov achieves competitive accuracy with only 10% of survey samples, indicating substantially reduced dependence on expensive label collection. The model further generalizes across populous countries within and outside Africa, and scales to a unified Africa-wide model with strong continent-level performance ($R^2=0.63$, $r^2=0.68$), from which we generate high-resolution decadal maps of wealth and wealth changes for the African continent. Analysis of these maps shows large variation in recent economic performance both within and across countries. Our open-source approach provides a pathway to timely, scalable, low-cost monitoring of wealth and poverty from routinely collected satellite data.

• The paper presents Tempov, a satellite foundation model that uses bi-temporal self-supervised pretraining to robustly map wealth indices in low-data settings.
• It employs a custom Vision Transformer with multispectral embedding and probabilistic regression, outperforming traditional methods on both static and dynamic predictions.
• Results show high R² scores across African nations, demonstrating scalable, cost-efficient wealth monitoring and effective handling of temporal shifts.

Satellite Foundation Modeling for Dynamic Wealth Monitoring: The Tempov Framework

Introduction and Context

The measurement and monitoring of economic well-being in low- and middle-income countries are persistently impeded by the scarcity, expense, and temporal lag of traditional ground-based surveys and censuses. Satellite imagery has emerged as a scalable alternative, but existing approaches exhibit diminished performance under temporal shifts and limited capacity to model dynamic, fine-grained socioeconomic variation. The paper "A satellite foundation model for improved wealth monitoring" (2604.23166) introduces Tempov, a multimodal satellite foundation model, to address these gaps. Tempov leverages large-scale bi-temporal self-supervised pretraining, robust multispectral embedding, and parameter-efficient fine-tuning to enable high-resolution, continent-scale mapping of both static and dynamic wealth indices.

Figure 1: Tempov framework showcasing global pretraining, diverse application regimes (nowcasting, hindcasting, change-tracking), dataset distribution, pretraining paradigm, and downstream pipeline.

Model Architecture and Training Paradigm

Tempov employs an enhanced Vision Transformer (ViT-Large) architecture customized for six-channel Landsat surface reflectance inputs. Patch embedding weights for RGB bands are initialized from conventional pre-trained models, whereas extraspectral bands (NIR, SWIR-1, SWIR-2) begin as zero to preserve the feature distribution. The model is pretrained using bitemporal self-supervised objectives on a curated dataset of three million image pairs spanning multiple seasons and years, constructed to maximize radiometric (phenological) variance and enforce temporal invariance in learned representations.

The self-supervised framework utilizes teacher-student DINO and iBOT losses, adapted to focus on seasonal, rather than geometric, variation, facilitating semantic disentanglement of structural socioeconomic changes from transient environmental noise. Performance is further enhanced by a probabilistic regression head during fine-tuning, optimizing Gaussian negative log-likelihood to dynamically attenuate noise from label uncertainty inherent to household survey and census data. Fine-tuning employs low-rank adaptation (LoRA) to mitigate overfitting and accommodate low-label regimes.

Benchmarking Static and Dynamic Wealth Estimation

Tempov is evaluated against a suite of supervised and foundation model baselines, including DOFA, Prithvi-v2, DINOv3, CLAY-v1.5, CNN, XGBoost, and SwinV2-T. In static wealth estimation conducted on 2008/2018 Malawi and 2007/2017 Mozambique census datasets, Tempov consistently achieves the highest $R^2$ (85–87% for Malawi; 73–74% for Mozambique), outperforming both geospatial and traditional baselines.

Critically, Tempov exhibits strong resilience in longitudinal tracking tasks where conventional models sharply degrade, often attaining near-random predictive power. Tempov maintains robust dynamic prediction with decadal-change $R^2$ of 69% for Malawi and 46% for Mozambique (Figure 2). Embedding visualizations reveal that learned representations correspond with salient urban and agricultural features and maintain superior temporal consistency, underpinning dynamic wealth change estimation.

Figure 2: Decadal wealth change prediction benchmarking in Malawi/Mozambique; PCA visualization highlights semantic embedding structure.

Performance in Data-Scarce Regimes

A key contribution is Tempov’s adaptability in scenarios with severely limited ground truth. In zero-shot settings (training on historical labels, direct prediction on novel imagery), Tempov retains substantial predictive power and outperforms alternatives. Few-shot adaptation (incorporating only 5% of contemporary census labels) rapidly restores accuracy beyond competing models, many of which suffer from negative transfer across temporal shifts (Figure 3).

Retrospective hindcasting (training on current year, sparse historical adaptation) further demonstrates Tempov’s temporal bidirectionality, achieving $R^2 = 0.62$ (Malawi) and $0.51$ (Mozambique) for reconstructing past wealth, and $R^2 = 0.31$/$0.29$ for decadal change. Even under the most resource-constrained regime (5% historical plus 5% current labels, no census), Tempov maintains reasonable accuracy where alternatives collapse ($R^2\sim0$). Joint scaling of historical census priors and contemporary survey fractions reveals synergistic cost-efficiency, with Tempov matching full-sample baseline models using only 10% of survey data.

Figure 3: Wealth prediction and change benchmarking under extreme data scarcity, showing zero-shot, few-shot, hindcasting, and sample size scaling robustness.

Scalability and Global Generalization

Tempov’s capacity to generalize spatially and temporally is validated in Kenya, Nigeria, and Bangladesh, using sparse DHS survey clusters. Fine-tuned models achieve the highest $R^2$ across states and demonstrate flexible spatial resolution deployment. Predicted maps delineate known intranational socioeconomic gradients, readily transforming sparse ground points into interpretable, country-scale asset wealth distributions (Figure 4).

Figure 4: Ground truth and predicted Asset Wealth Index (AWI) maps for Kenya, Nigeria, and Bangladesh, with performance benchmarking against geospatial foundation models.

Continental-Scale Wealth Mapping for Africa

Leveraging harmonized, multi-country DHS datasets and efficient inference pipelines, Tempov is scaled to continent-wide, high-resolution asset wealth mapping across Africa. A unified model achieves cross-country $R^2 = 0.63$, matching specialized single-country deployments. Wealth maps for 2015 and 2025 reveal substantial within- and cross-country heterogeneity; Theil decomposition shows 80% inequality is within-country. Decadal wealth change maps indicate spatially diverse gains and losses aligned with macroeconomic trends (GDP per capita), but also expose finer spatial heterogeneity not captured by national aggregates (Figure 5; Supplementary Figure 1).

Figure 5: Predicted static wealth in 2015/2025 and decadal change across Africa with population density masking.

Analysis of Growth Determinants

Tempov’s decadal wealth changes display modest evidence for $β$-convergence (poorer areas marginally outpace richer), consistent with recent national account analyses. Decomposition analyses indicate local-level determinants (temperature trends, conflict intensity) explain more variance in wealth change than country-level institutional factors or remoteness (Supplementary Figure exfig:6). This points to the critical importance of integrating environmental and conflict granularity into future socioeconomic modeling.

Methodological and Practical Implications

Tempov demonstrates that satellite foundation modeling, anchored by robust seasonal and spatial invariance in representation learning, provides an efficient framework for high-frequency, high-resolution wealth monitoring. Its resilience to label scarcity, temporal drift, and spatial transfer positions it as a practical backbone for evidence-driven development targeting, disaster response, and evaluation, especially amid declining coverage of ground surveys.

Future work may focus on integrating multimodal auxiliary signals (e.g., mobile-phone metadata, object-level commercial imagery), explicit temporal modeling for richer sequence-based inference, and generalizing beyond asset wealth to more direct poverty statistics (income/consumption). Design-consistent calibration and uncertainty quantification remain central requirements for deployment in policy settings.

Conclusion

Tempov advances the state-of-the-art in dynamic wealth monitoring, unifying static and decadal measurement in resource-constrained environments, and scaling robustly to global deployments. Its methodological innovations facilitate accurate, timely, and scalable asset wealth mapping, enabling policy-relevant socioeconomic surveillance and targeted interventions using accessible satellite-based data. The paradigm represents a significant step toward democratized, low-cost, spatially-detailed economic evidence infrastructure for the global South.