Tempov: Satellite Wealth Mapping

• Tempov is a satellite foundation model that leverages bi-temporal Landsat imagery and self-supervised pretraining to create detailed wealth maps.
• Its architecture features an enhanced Vision Transformer with bi-DINO and bi-iBOT losses to maintain temporal consistency despite seasonal variations.
• Low-Rank Adaptation and an uncertainty-aware regression head enable robust wealth estimation from sparse survey labels and mitigate challenges of temporal shift.

Tempov is a satellite foundation model for wealth monitoring that is pretrained by self-supervision on three million bi-temporal Landsat pairs and then adapted with parameter-efficient fine-tuning to sparse survey labels. It is designed to generate large-scale, high-resolution maps of asset-based wealth indices, support zero-shot nowcasting up to a decade after observed labels, retrospective hindcasting, and decadal change tracking, and mitigate the temporal shift that has limited earlier satellite-based poverty models. The model is presented as an open-source pathway to timely, scalable, low-cost monitoring of wealth and poverty from routinely collected satellite data, with its most extensive demonstration centered on Africa (Zheng et al., 25 Apr 2026).

1. Definition and problem setting

Tempov addresses the problem of measuring poverty and wealth accurately, frequently, and at fine spatial scales in settings where censuses and household surveys are expensive, infrequent, temporally lagged, spatially sparse, and sometimes noisy. The model is motivated by the observation that daytime satellite imagery encodes proxies for development such as urban form, roads, building density, and agricultural patterns, while previous imagery-based poverty models often required country-specific training, were built on relatively small task-specific datasets, struggled to capture local variation, and degraded under temporal shift (Zheng et al., 25 Apr 2026).

The target variable is the asset-based wealth index (AWI), constructed via PCA from household assets and housing characteristics. In the reported experiments, labels come from full-count Population and Housing Censuses in Malawi and Mozambique, Malawi IHS 2016, and DHS country-year entries across Africa, with harmonization across DHS waves following the recoding of housing materials, water, sanitation, and assets. Tempov is therefore not framed as a direct estimator of consumption or income; it is calibrated to AWI.

A recurrent misconception is to read “bi-temporal” as long-horizon sequence modeling. In Tempov, “bi-temporal” refers specifically to two seasonal views of the same location and year. The point of this design is not to supervise interannual change directly during pretraining, but to make the representation invariant to short-term phenological and illumination variation so that later supervised adaptation can focus on multi-year socioeconomic variation.

2. Architecture and self-supervised pretraining

Tempov uses an enhanced Vision Transformer with a ViT-Large backbone adapted to multispectral Landsat imagery. The input is a 6-channel Landsat surface reflectance tensor with Blue, Green, Red, NIR, SWIR-1, and SWIR-2 bands. Patch embedding is implemented with a convolutional patch embed of kernel size and stride $16$, producing $1024$-dimensional patch tokens. The token set includes one learnable class token, four storage tokens, and a learnable mask token. The encoder has 24 Transformer blocks with 16 attention heads per block, pre-norm LayerNorm, multi-head self-attention plus a 2-layer MLP with intermediate dimension $4096$ and GELU, residual connections with LayerScale initialized at $10^{-5}$, 2D rotary positional encoding, and stochastic depth with drop-path rate $0.3$ during pretraining. The normalized class token is used as the global embedding for wealth prediction, while normalized patch tokens are used in self-supervised objectives (Zheng et al., 25 Apr 2026).

To initialize multispectral inputs, Tempov reuses RGB-pretrained patch projection weights for Blue, Green, and Red, while initializing the NIR, SWIR-1, and SWIR-2 channels to zero. This preserves the original feature distribution at initialization and then allows the network to learn the additional bands during pretraining.

The pretraining corpus contains 3 million bi-temporal Landsat pairs from the SSL4EO-L archive, spanning Landsat 7, 8, and 9 over 2001–2022 and about 250,000 human settlement locations globally. For each location, the model sees two images from different seasons within the same year, chosen to maximize phenological and illumination variance. Each training example includes one global crop $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$ and four local crops $$x_{t,j}^{\ell} \in \mathbb{R}^{96 \times 96 \times C}$$.

Self-supervision is built on a teacher-student joint-embedding architecture extending DINO and iBOT to the bi-temporal setting. With teacher $g_{\theta_t}$ defined as an EMA of student $g_{\theta_s}$ with momentum $0.992$, the bi-DINO image-level loss enforces temporal consistency across seasons,

$1024$0

while the bi-iBOT patch-level loss aligns masked patch predictions across the two seasonal views,

$1024$1

Uniformity regularization is added to avoid feature collapse. Pretraining runs for 800,000 steps on 4 NVIDIA H100 GPUs with batch size 64 per GPU, AdamW, weight decay $1024$2, constant learning rate $1024$3 after a 10,000-step warm-up, bf16 mixed precision, FlashAttention-2, and FSDP.

3. Fine-tuning, labels, and uncertainty-aware prediction

For downstream wealth estimation, Tempov is adapted with Low-Rank Adaptation rather than full fine-tuning. For each selected transformer weight matrix $1024$4, specifically the query and value projections, the model learns

$1024$5

where $1024$6, $1024$7, rank $1024$8, and $1024$9 remains frozen. This sharply reduces the number of trainable parameters and is reported to improve robustness in low-label regimes (Zheng et al., 25 Apr 2026).

The regression head is heteroscedastic Gaussian. Given the global embedding $4096$0, the model predicts

$4096$1

with mean $4096$2 and variance $4096$3. Training uses Gaussian negative log likelihood,

$4096$4

This downweights samples with large predicted uncertainty, which is relevant because survey-based labels can be noisy due to survey error, GPS jitter, and sampling noise.

Spatial alignment is handled by pairing each census enumeration area or DHS cluster with a Landsat patch built from median composites over multi-year windows. Patch footprints are chosen to accommodate DHS GPS jitter of 2 km in rural areas and 5 km in urban areas. For mapping, the reported grid resolution is $4096$5, and for continent-wide products grid cells with population density $4096$6 person/$4096$7 are masked.

4. Tasks, metrics, and temporal shift

The model is evaluated on five task families: static high-resolution wealth mapping, dynamic measurement of decadal wealth change, zero-shot nowcasting, few-shot nowcasting, and retrospective hindcasting. A sixth regime, described as a no-census extreme data-scarce scenario, uses only small survey fractions in historical and current periods. For Africa-wide evaluation, the protocol distinguishes out-of-country, in-country out-of-year, in-country in-year, all-countries out-of-year, and all-countries in-year scenarios, with 1,715 Tempov models trained under a unified spatiotemporal cross-validation design (Zheng et al., 25 Apr 2026).

Performance is reported with both coefficient of determination and squared Pearson correlation: $4096$8 and

$4096$9

The distinction matters because $10^{-5}$0 is sensitive to level bias, whereas $10^{-5}$1 measures correct ordering.

Temporal shift is central to the article’s framing. It denotes the change in the statistical relationship between imagery and wealth across years because of infrastructure change, urban expansion, agricultural change, climate, or sensor variation. The paper quantifies the underlying instability by noting that the $10^{-5}$2 between earlier and later census wealth labels is $10^{-5}$3 in Malawi and $10^{-5}$4 in Mozambique. This means that even the ground-truth target is only modestly persistent in one case and negatively correlated in the other, so models that mainly memorize static spatial structure can fail badly in change estimation.

5. Empirical performance and continent-scale outputs

Across most reported scenarios, Tempov outperforms XGBoost, a task-specific CNN, SwinV2-T, and the geospatial foundation models DOFA, DINOv3 RGB, Prithvi-v2, and CLAY-v1.5. The strongest contrast appears in decadal change tracking and temporal transfer, where several baselines are described as falling to near-random predictions or near zero $10^{-5}$5, while Tempov retains substantial predictive power (Zheng et al., 25 Apr 2026).

Setting	Reported metric	Value
Africa-wide unified model	$10^{-5}$6	0.63
Africa-wide unified model	$10^{-5}$7	0.68
Country-specific Africa models	$10^{-5}$8	0.64
Country-specific Africa models	$10^{-5}$9	0.68
Malawi decadal change	$0.3$0	$0.3$1
Mozambique decadal change	$0.3$2	$0.3$3
Malawi hindcasting, 2008 levels	$0.3$4	0.62
Mozambique hindcasting, 2007 levels	$0.3$5	0.51
Malawi hindcasting, changes	$0.3$6	0.31
Mozambique hindcasting, changes	$0.3$7	0.29

Static wealth prediction is also strong: Malawi 2008 and 2018 are reported at average $0.3$8 and $0.3$9, while Mozambique 2007 and 2017 are reported at average $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$0 and $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$1. In zero-shot nowcasting, Tempov is reported to retain substantial predictive power and often match or exceed geospatial foundation models that are allowed a few-shot adaptation. In few-shot nowcasting with 5% of target-year labels, performance “rapidly recovers” and clearly exceeds all baselines. In low-label regimes, Tempov is reported to reach similar accuracy with 10% of survey labels to what other models require with 100%.

The unified Africa-wide model is trained on DHS data from 34 African countries and 42,993 clusters, and a 5-fold ensemble is used to generate continent-wide predictions. The resulting maps cover Africa on a $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$2 grid for 2015, 2025, and 2015–2025 wealth change. The reported spatial patterns include strong within-country heterogeneity, with decomposition of the Theil index indicating that approximately 80% of inequality is within countries rather than across them. Wealth gains are concentrated in parts of West and East Africa, while declines appear in parts of Southern and Central Africa. At 50-km resolution, the paper reports statistically significant $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$3-convergence with $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$4, 95% CI $x_t^{g} \in \mathbb{R}^{224 \times 224 \times C}$5, and notes that country effects explain only about one-third of the variance in wealth change.

6. Interpretation, applications, and limitations

Tempov is positioned at the intersection of satellite-based poverty mapping, geospatial foundation models, and research on temporal domain shift in remote sensing. Its novelty lies in the scale of pretraining, the explicitly bi-temporal self-supervised design of bi-DINO and bi-iBOT, the use of LoRA for sparse-label adaptation, the uncertainty-aware regression head, and the production of an open-source, continent-wide African wealth map at 6 km resolution with a single unified model (Zheng et al., 25 Apr 2026).

The model’s practical appeal is tied to routine Landsat acquisition and relatively cheap inference after pretraining. The reported inference pipeline parallelizes data acquisition with one CPU node per country and GPU inference with one A100 GPU node per country, mapping all of Africa in approximately 3 hours with 36.5 GPU-hours. The intended users include national statistical offices seeking inter-census monitoring and small-area estimates, governments and NGOs targeting poverty alleviation or evaluating interventions, and researchers studying growth, inequality, conflict, and climate impacts with a consistent fine-scale proxy.

Several limitations are explicit. AWI is an asset proxy rather than a direct measure of consumption or income. DHS cluster locations are jittered, introducing nontrivial spatial noise. Landsat’s 30 m resolution limits the ability to resolve within-neighborhood inequality. The model uses two-time-point reasoning rather than full temporal sequences, and it relies only on satellite imagery rather than multimodal fusion. The growth-determinant analysis is described as associational rather than causal. Ethical concerns are also implied: coarse imagery and aggregate labels reduce individual privacy risks relative to high-resolution imagery or mobile phone data, but wealth maps could still be misused for surveillance or discriminatory targeting.

These limitations help delimit what Tempov is and is not. It is not a substitute for household microdata, not a direct welfare accounting system, and not a full temporal sequence model. It is a geospatial representation-and-estimation framework that converts medium-resolution optical imagery and sparse survey labels into temporally robust, high-resolution AWI surfaces, with unusually strong empirical performance under temporal shift and label scarcity.