Models that map floods, crops or poverty from satellite images lose accuracy when they are moved to a region they were not trained on, sometimes by up to 30% on unseen disaster sites (Position: Mission Critical). Today nobody can say why in terms an analyst can act on. One possible cause is that the water itself looks different. Another is that the landscape around the water is different, for example steeper terrain, more cropland or a different built environment, and the model had learned to lean on that surrounding context. We propose a model that makes this question answerable. The encoder first predicts a set of physically meaningful maps, which we call "concepts": land-cover fractions, slope, elevation, permanent water and similar quantities. A small head then makes the final prediction from those maps alone. The head has two parts. One looks only at each location's own neighbourhood. The other uses attention, a mechanism that lets any location read from any other location in one step, to pull in distant context.
Earth observation has a property that ordinary concept-based models lack. Most of these concepts do not need human annotators, because co-registered measured products already exist for every pixel on Earth. We can therefore take the model's own predicted concepts and swap them for the measured values, one concept at a time, on data from the new region. The accuracy recovered by each swap tells us how much of the out-of-region error flows through that concept. Because the head separates local from distant evidence, the same swap also tells us whether the error comes from the site itself or from its surroundings. The deliverable is a per-region failure report grounded in interventions, together with a test of whether that report actually helps fix the model.
Core Claim
We claim that intervention-based concept attribution in such a model will identify the causes of geographic failure better than three simpler routes. The first is comparing the regions' measured products directly with statistical tests. The second is probing a black-box foundation model after training. The third is black-box accuracy estimation. We define "better" operationally: the diagnosis should pick out a known planted cause, and when a small budget of new labels is spent according to the diagnosis, those labels should repair more of the lost accuracy. A second, riskier hypothesis is that a substantial share of out-of-region error is carried by the long-range, distant-context part of the head rather than by on-site appearance. If that holds, it would explain why methods that align image style do little on satellite shifts. That weakness is documented in Extending the WILDS Benchmark for Unsupervised Adaptation, which found that domain-adversarial training, feature alignment and self-training give marginal or inconsistent gains on PovertyMap-WILDS.
The closest prior work is concept-based shift detection. Failing Conceptually runs per-concept two-sample tests on the outputs of a concept bottleneck model to rank shifted concepts, but only on synthetic shape datasets (dSprites, 3dshapes) with hand-labeled concepts. SGShift fits a sparse correction to explain changes in the label-given-input relationship, and its authors state that extending it beyond tabular data is open. We differ from both in three ways. We use real geographic holdouts. We take concept supervision from measured products rather than annotators. Most importantly, we rank concepts by their effect on the model's error under intervention, not by how much their distributions differ. A concept can shift a great deal and not matter, and that distinction is exactly what the two-sample approach cannot make.
Research Program
The backbone is a geospatial foundation model, Prithvi 2.0 300M, adapted with low-rank adaptation (LoRA), which trains small added weight matrices instead of the whole network. We choose this setup because Fine-tune Smarter, Not Harder found it generalizes to geographic hold-out sets better than full fine-tuning. The encoder outputs a patch-level stack of $K$ concept maps.
The concept set comes from the kinds of layers Galileo already ingests as input modalities: Dynamic World land-cover fractions, elevation and slope, ERA5 weather and nighttime lights. We add permanent-water extent where a task needs it. Galileo feeds these layers in as inputs to a black-box model. We instead make them the only channel through which information reaches the task head, with no bypass connection. This follows the warning in Concept Bottleneck Models that such "side channels" let the network route around the concepts. That paper also showed the benefit of forcing predictions through concepts. On the synthetic CUB+Places background shift, the concept model held error near 48–49% while a standard end-to-end model rose to 62.7%. It also showed that correcting two concepts at test time cut knee-arthritis grading error from over 0.4 to about 0.3.
The head has a local branch, a small convolutional network over the concept maps, and a non-local branch, attention over concept tokens. The two are summed before the output layer. This design also addresses a problem raised by ConEx. ConEx explains vision transformers after training by averaging patch embeddings inside concept masks, and it names faithful per-location concept attribution in transformers as an open problem, because attention blends patch information across the whole image ("smearing"). In our head, attention operates over tokens that are already named and located concepts. What gets blended is interpretable by construction, and the unconstrained backbone attention sits upstream of the bottleneck.
We also borrow an idea from Graph Concept Bottleneck Models. That work learns dependencies among concepts in an abstract graph and gains 1–2% accuracy over independent-concept models on CUB and medical data. Our dependencies differ in kind: they run between concepts at different places in the scene. Transformer-based Dual Relation Graph couples spatial attention with a graph over labels, but it has no inspectable bottleneck and was evaluated only on in-distribution benchmarks. We keep its spatial-plus-semantic intuition and add the bottleneck and the shift setting.
Diagnosis works as follows. Let $L_T(\cdot)$ be task loss on a held-out target region, $\hat c$ the predicted concept maps, and $\hat c^{(k)}$ the same maps with concept $k$ replaced by its measured value. The attributed effect of concept $k$ is
$$E_k = L_T\big(h(\hat c)\big) - L_T\big(h(\hat c^{(k)})\big),$$
where $h$ is the frozen head. $E_k$ measures perception failure: the error removed by giving the head the correct value of concept $k$.
A second quantity measures context reliance. We replace concept $k$ only outside a radius around each evaluated pixel, and replace it with source-typical values rather than measured ones. The accuracy change under this substitution is the part of the head's error that enters through distant context about concept $k$. A per-region report then lists the few concepts with the largest effects, splits each effect into perception and context parts, and splits the context part between the local and non-local branches.
The measured products are imperfect; Dynamic World is itself a model's output. We therefore treat them as noisy references. We report how sensitive the conclusions are to that noise by rerunning the analysis on concepts where independent references exist, such as elevation against a second terrain product.
Baselines
The naive reference points are simple. On Sen1Floods11 they are Otsu thresholding, which reports mIoU 0.4571 on permanent water, and a ResNet50 fully convolutional network trained on Sentinel-2 weak labels, which reports flood-water mIoU 0.3389 (Sen1Floods11). On PovertyMap-WILDS the reference is empirical risk minimization, i.e. ordinary training, which falls from Pearson 0.60 in-distribution to 0.45 out-of-distribution (WILDS).
The strongest accuracy competitor is LoRA-adapted Prithvi 2.0 300M with a UNet decoder. It reports 87.57% mIoU on the Sen1Floods11 geographic hold-out set, against 82.07% for full fine-tuning (Fine-tune Smarter, Not Harder). It is the strongest competitor because it is the best reported holdout number in the file and uses the same backbone we use. We rerun it ourselves on identical splits, with the same LoRA rank, training steps and hyperparameter search budget, and give our model the same backbone and budget. Our model is not expected to beat it on accuracy. The bar is to stay within three mIoU points without measured-concept substitution. If deployment-available static concepts (elevation, slope, pre-event land cover) are swapped in at test time, the bar is to match it or exceed it. A larger accuracy loss would undercut practical use, and we would report that as a result.
For diagnosis, the naive baseline ranks concepts by Kolmogorov–Smirnov distance between source and target measured products, with no model involved. The strongest existing method is concept-shift scoring in the style of Failing Conceptually, applied to our predicted concepts. A post-hoc baseline trains linear probes for the same concepts on the LoRA black box and computes TCAV-style sensitivities. TCAV measures how much a prediction changes when internal activations move along a concept direction. As an aggregate-only reference we include ATC, which estimates target accuracy from unlabeled data.
Two ablations isolate the key ingredients. The first removes the non-local branch. The second restores a bypass connection from backbone to head, with concepts used only as an auxiliary loss. The first tests whether distant context carries the error. The second tests whether diagnosis still works once information can leak around the concepts.
Why not just…?
Why not just feed the measured products as extra input channels, as Galileo does? For accuracy alone, this may match or beat our model, and we treat that as a real threat. We will train that variant under the same budget. However, a model that simply receives the layers as inputs gives no single place to intervene. Its reliance on each layer is mixed into the backbone features, so its diagnosis has to fall back on post-hoc probing. The planted-shift and repair experiments below compare the two directly. If probing the input-channel model diagnoses as well as our interventions, the bottleneck is not worth its cost, and we will say so.
Why not just compare the regions' products with statistical tests and skip the model? This is the cheapest diagnosis available and it will often flag real differences. What it cannot tell us is whether the model's errors depend on those differences. Our decisive test targets this gap. We plant shifts in which a concept changes strongly but was made irrelevant during training, alongside shifts in a concept the model does rely on. Statistical-test ranking should flag both. Intervention-based ranking should flag only the second. If the two rankings agree, the model-free method wins.
Why not just probe LoRA-Prithvi after training? Probes show what information is present in the features, not what the prediction depends on. Directional sensitivities of the kind used in TCAV may also be affected by the attention smearing that ConEx reports. We test this on the same planted shifts and repair budget. We do not assume probes lose.
Why not just use LoRA and label some target data? That is the repair step, not an alternative to diagnosis. The question is which target data to label, and our main experiment compares diagnosis-guided selection against random selection at equal label counts.
Proposed Experiments
The primary testbed is Sen1Floods11, using its geographic hold-out set and its events spread across six continents and all 14 biomes (Sen1Floods11). We also run leave-one-event-out splits so that each event serves as a target region. The second testbed is PovertyMap-WILDS with its country-level splits. We expect distant context, such as surrounding settlements, roads and nightlights, to matter most here.
The first decisive check uses planted shifts with a known cause. We filter training chips so that one concept is skewed. For example, training floods occur only on low-slope terrain or only near low cropland fractions, and testing is on the complementary slice. We also build a control in which a concept shifts strongly but the model was trained to ignore it. The metric is how often each method ranks the planted concept first, and how often it wrongly ranks the irrelevant concept highly.
The second decisive check is repair on real holdouts. Each method gives a top-ranked concept. We spend a fixed budget of target labeled chips, stratified to cover that concept's range in the target region, and fine-tune. We compare the recovered mIoU (Sen1Floods11) or Pearson correlation (PovertyMap) against random selection and against the other methods' rankings at equal budget. We also check whether the attributed effects $E_k$ predict, in rank order, which concept-targeted repairs recover the most.
The third check tests the context hypothesis. On real holdouts we measure how much out-of-region error disappears when distant concepts are swapped for measured values versus when on-site concepts are, and we repeat this with the non-local branch ablated.
Throughout, we report accuracy against the LoRA-Prithvi baseline, with and without swapping in measured static concepts at deployment. We use several seeds and bootstrap intervals over regions, following the resampling practice of GEO-Bench.
The project succeeds if interventions beat both statistical-test ranking and post-hoc probing on planted-cause recovery and on repair gain, at an accuracy cost of a few points or less. It fails informatively if the model-free statistical tests do just as well. That outcome would show that geographic failures in these tasks are mostly visible in the raw covariates, without needing a model at all.
References
- Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning
- Extending the WILDS Benchmark for Unsupervised Adaptation
- Failing Conceptually: Concept-Based Explanations of Dataset Shift
- Explaining Concept Shift with Interpretable Feature Attribution
- Fine-Tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities
- Concept Bottleneck Models
- ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
- Graph Concept Bottleneck Models
- Transformer-based Dual Relation Graph for Multi-label Image Recognition
- Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1
- WILDS: A Benchmark of in-the-Wild Distribution Shifts
- Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)
- Leveraging unlabeled data to predict out-of-distribution performance
- GEO-Bench: Toward Foundation Models for Earth Monitoring