Stain Jitter for the Night Sky: Sensor-Targeted Augmentation for Satellite Poverty Regression

Pathology models break when they move between hospitals because they learn each lab's dye color instead of tissue shape. The fix that works best does not try to make features identical across hospitals. It randomizes stain color in a physically meaningful way while leaving the tissue untouched, so the color cue stops being useful. Satellite poverty mapping has the same failure, and its version of stain color is the radiometric fingerprint of the satellite sensors. Nighttime-light images, which record how much artificial light leaves the ground, look systematically different across the older military sensor (DMSP/OLS) and the newer calibrated radiometer (VIIRS). Their gain, saturation and "blooming" around bright cities all differ by sensor. Daytime multispectral bands differ by atmosphere and acquisition conditions. The field has no label-preserving augmentations for this data, and that gap is the stated reason unsupervised adaptation failed on satellite poverty regression in Extending the WILDS Benchmark for Unsupervised Adaptation.

We propose to build the satellite analogue of H&E stain jitter. We would estimate, from the data itself, how sensor and country change nighttime-light gain, offset, saturation ceiling and spatial blur, and how acquisition conditions change each daytime band. We then randomize exactly those factors. Because the transforms leave a village's wealth label unchanged, they give regression something it has lacked: two views of the same unlabeled image whose predictions should agree. That agreement can drive self-training on unlabeled target-country imagery. The core claim is that sensor-targeted augmentation, plus a simple regression consistency loss built on it, raises out-of-country correlation on PovertyMap-WILDS where every existing adaptation method gave no significant gain.

What the two directions supply

Histopathology supplies the method and the theory. Out-of-Distribution Robustness via Targeted Augmentations splits features into object features, noise, spurious domain features and predictive domain features. It proves that with fewer training domains than domain-feature dimensions, models memorize the spurious ones. It shows that randomizing only stain color lifts Camelyon17 out-of-distribution accuracy from 75.3% to 90.5%, and it argues that full invariance methods discard predictive domain context. That last point matters here. Country and urbanization carry real economic signal, a point supported by the shift from population-tracking to GDP-tracking light described in A Prolonged Artificial Nighttime-light Dataset of China. So we randomize the sensor fingerprint only, not "domain" as a whole.

Structure-Preserving Multi-Domain Stain Color Augmentation (HistAuGAN) and Model-Based Domain Generalization achieve the same goal with learned style translators. We deliberately do not port them. They need discrete domains and costly generators, and MBDG assumes the label relationship is fixed across domains, which the nighttime-light literature contradicts.

The nighttime-light direction supplies the application and its constraints. Using publicly available satellite imagery and deep learning to understand economic well-being in Africa fuses Landsat and nighttime lights in a ResNet-18. It shows the two channels are complementary, with lights driving spatial differences and daytime bands driving temporal change, so we augment both rather than drop either.

Two papers motivate the proposal by documenting the gap. WILDS reports an ERM drop on PovertyMap from Pearson 0.60 in-distribution to 0.45 out-of-distribution. Make the U in UDA Matter (ICON) gains 2.4% on Camelyon17 but underperforms on PovertyMap. Its authors attribute this to no pretrained backbone and to clustering that needs a known class count. Our consistency loss needs neither.

A separate paper constrains our claims. Fairness and Representation in Satellite-Based Poverty Maps shows that models use urbanization as a shortcut, so rank correlation collapses within urban-only or rural-only subsets. Urbanization is not a sensor artifact. We therefore do not claim our augmentation fixes that shortcut, but we measure whether it helps, since saturated, bloomed city lights are one channel through which the shortcut is expressed.

Core mechanism

For a nighttime-light image $n$, the augmentation draws sensor parameters $\theta = (g, b, s, \sigma)$ and produces

$$n' = \min\big(s,\; g\,(K_\sigma * n) + b\big),$$

where $g$ is gain, $b$ is offset, $s$ is a saturation ceiling, and $K_\sigma$ is a Gaussian blur of width $\sigma$ that mimics DMSP-style blooming. The ranges of $\theta$ are fit per sensor-by-country group from the training imagery, much as stain matrices are estimated per hospital. We match histograms of lit pixels and compare the edge profiles of bright clusters. Augmentations sample from the convex hull of the fitted groups, slightly extended.

Daytime bands receive per-band affine perturbations whose covariance is estimated from between-country differences in the reflectance of dark and vegetated pixels. Those pixel classes should not depend on wealth, so the covariance captures acquisition conditions rather than economics. A guard rail keeps the transform label-preserving: within any single augmented image, the brightness ordering of pixels is preserved. Labels in this setting depend on relative brightness, not absolute calibration.

For adaptation, a teacher network averaged over training steps predicts on a weakly augmented unlabeled target image $x$. The student predicts on a strongly sensor-augmented view $T_\theta(x)$, and we minimize

$$\mathcal{L} = \mathcal{L}_{\text{sup}} + \lambda\, \mathbb{E}_{x,\theta}\big[\, w(x)\,(f(T_\theta(x)) - \bar f(x))^2 \,\big],$$

where $f$ is the student, $\bar f$ is the teacher, $\lambda$ weights the unlabeled term, and $w(x)$ down-weights examples whose teacher predictions vary strongly across augmentations. This replaces the confidence threshold that pseudo-labeling uses for classification and that WILDS 2.0 found does not carry over to regression.

Research Program

The work proceeds in three stages. The first characterizes the sensor fingerprint. We quantify how much PovertyMap's nighttime-light statistics differ across sensor eras and countries after conditioning on survey wealth. If the fingerprint is small relative to the economic signal, the premise fails and we stop. The second trains supervised models with targeted augmentation alone. The third adds the consistency term using the unlabeled PovertyMap imagery released with WILDS 2.0.

As a sanity check on implementation, the identical consistency code is run on Camelyon17 with H&E jitter, where it should approach the published targeted-augmentation numbers. We tune $\lambda$ and the range extension only on the benchmark's out-of-distribution validation countries, never on test countries, so that the unsupervised model-selection problem is not quietly assumed away.

Baselines

The naive reference point is ERM with the standard WILDS PovertyMap recipe, a ResNet over multispectral bands plus nighttime lights. WILDS reports Pearson 0.45 out-of-distribution against 0.60 in-distribution. We rerun it on the same splits and seeds.

The strongest existing competitors are the WILDS 2.0 adaptation methods: Noisy Student, FixMatch-style pseudo-labeling, DANN, CORAL, and self-supervised pretraining. WILDS 2.0 states that none gave a significant boost on PovertyMap. No reported result for a method that significantly beats ERM on PovertyMap was found on Lacuna, so the bar is ERM's 0.45 out-of-distribution Pearson plus whatever these methods reach in our reruns. ICON is included because it is the strongest recent adaptation method, though its own authors report weaker PovertyMap performance. All methods get the same backbone, the same unlabeled pool, and the same number of tuning trials selected on validation countries.

Beating them means a mean out-of-distribution Pearson gain over ERM that exceeds the across-seed spread on the benchmark's country-held-out splits. It also means no loss on worst-group (urban versus rural) correlation.

Three ablations isolate the key ingredient, the fact that the augmentation is targeted. The first is untargeted jitter: random gain, noise and blur with matched total perturbation magnitude but ranges not fit to sensors. The second applies targeted augmentation to nighttime lights only and to daytime bands only. The third uses targeted augmentation without the consistency term, which separates the supervised robustness gain from the adaptation gain.

Why not just…?

Why not just normalize each country's or sensor's nighttime lights (histogram matching, the analogue of stain normalization)? This is the most serious threat, because sensor offsets are partly a calibration problem. Normalization fixes a single canonical appearance and needs the target domain's statistics. It also removes real between-country brightness differences that carry wealth signal, the predictive domain context the targeted-augmentation paper warns against discarding. We have no reported comparison of the two on Lacuna, so we run per-country and per-sensor histogram matching as a full baseline. If it matches our method, the honest conclusion is that calibration, not augmentation, is what PovertyMap needs, and we will report that.

Why not just use generic augmentation such as MixUp, noise, or random color jitter? The targeted-augmentation analysis shows that generic augmentations do not scale with the number of spurious domain dimensions when training domains are few. MixUp-MIL found that mixing across slides with conflicting stains hurt accuracy, and that only within-slide mixing helped. The untargeted-jitter ablation tests this directly on PovertyMap.

Why not just drop nighttime lights and use daytime imagery, avoiding sensor inconsistency altogether? Yeh et al. show the lights carry most of the spatial wealth differentiation, so dropping them should cost accuracy. We include a daytime-only model to measure that cost against the fingerprint we remove.

Why not just apply post-hoc urban/rural mean calibration? Steele et al. show it needs urban/rural labels that are often unavailable and does not repair within-group ranking. It also targets a different failure than ours. We report it as a complementary layer, not a competitor.

Proposed Experiments

The primary benchmark is PovertyMap-WILDS with its country-held-out splits and unlabeled pool. We report out-of-distribution Pearson, worst urban/rural group Pearson, and within-urban and within-rural Spearman correlation, following Steele et al.'s diagnostic.

Three decisive checks decide the outcome. First, a sensor-era transfer test trains on one nighttime-light sensor era and tests on the other. Targeted augmentation should shrink this gap far more than untargeted jitter. If it does not, the fingerprint hypothesis is wrong. Second, the consistency term must improve over augmentation-only training on held-out countries. If it does, regression adaptation becomes workable once label-preserving views exist, which answers WILDS 2.0's negative result. Third, the histogram-normalization baseline must be beaten, or the contribution reduces to calibration.

Camelyon17 serves only as an implementation check that the shared consistency code reproduces stain-jitter gains of the published magnitude (90.5% in the targeted-augmentation paper).