Many high-resolution land-cover maps are trained from labels that do not say exactly what each pixel is. A coarse 10 m map, two weak labeling sources that disagree, or a hand-drawn polygon that spills across a road edge each give a pixel a small set of possible classes, only one of them correct. Partial label learning (PLL) is the field that trains models from such candidate sets. It works by recovering the true label from within each set. Its main tool rests on one assumption: if two examples look alike, they share the true label, while the wrong "distractor" labels differ at random between them. That assumption breaks in satellite imagery. Distractors there are not random. They come from the labeling process, which fails in the same way across a whole region. An optical water index can misfire across an entire biome, and a coarse map can blur one district's farmland into bare soil everywhere. Pixels that look alike and sit near each other therefore tend to share their errors. A model that looks to them for evidence will confirm the mistake.
We propose a fix for dense segmentation in this setting. When the model gathers evidence from similar-looking pixels to decide a label, it should count only pixels that are geographically far enough away that their label errors are unlikely to be shared. We call this geographically decorrelated disambiguation. We expect it to matter most at object boundaries and thin structures, because that is where candidate sets come from in practice. So the method is built on a detail-preserving decoder with multi-scale attention, and it is judged on held-out regions using boundary-sensitive metrics as well as overall accuracy.
Core Claim
PLL methods that propagate labels over a feature-similarity graph assume that distractors are independent across neighbors. This includes instance-based propagation in Solving the Partial Label Learning Problem: An Instance-Based Approach, credible-label elicitation in Partial Multi-Label Learning via Credible Label Elicitation, and the bag-plus-instance graphs of Dual Graph Disambiguation for Multi-Instance Partial-Label Learning. These papers supply the propagation technique we reuse. None of them models where their examples come from, and that independence assumption is exactly what spatial autocorrelation violates. Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning argues that this autocorrelation is a defining property of Earth observation data and reports drops of up to 30% when models move out of region.
Our claim makes testable predictions. First, if label errors are spatially correlated, the precision of nearest-neighbor label votes should fall as neighbors get closer in space. Second, excluding nearby neighbors should help in proportion to how far that correlation extends. Third, the gain should be largest on geographically held-out test regions, where memorizing regional label quirks is most costly. When distractors are independent, the exclusion should do nothing, and that null case is part of the test.
Method
Each training pixel or superpixel $i$ carries a candidate set $S_i$, a location, and a feature vector $z_i$ from the decoder. We keep a running posterior $q_i$ over classes that is restricted to $S_i$. At each disambiguation step it is updated as
$$ q_i(y) \;\propto\; \mathbf{1}[y \in S_i]\; p_\theta(y \mid x_i)^{\alpha}\Big(\sum_{j} w_{ij}\, q_j(y)\Big)^{1-\alpha}, \qquad w_{ij} = \mathrm{sim}(z_i, z_j)\,\big(1 - e^{-d_{ij}/\rho}\big). $$
The terms are defined as follows. $p_\theta(y\mid x_i)$ is the network's current class probability. $\alpha$ balances the model's own belief against evidence from neighbors. $\mathrm{sim}$ is cosine similarity restricted to the $k$ nearest feature neighbors drawn from a memory bank. $d_{ij}$ is the geographic distance between $i$ and $j$. $\rho$ is the correlation range of label disagreement.
We do not tune $\rho$ freely. We estimate it once from the training labels with an empirical variogram, a standard geostatistics plot of how disagreement between label sources grows with distance. Setting $\rho \to 0$ recovers ordinary graph propagation, which gives a built-in ablation. The network is trained with cross-entropy against $q$. For warm-up we use the class-activation selection rule of Exploiting Class Activation Value for Partial-Label Learning. That paper notes a weakness of the rule: a poor early warm-up spreads wrong labels. The distance-weighted graph is meant to stop that spread from staying inside one region.
For the network, we use a UNet-style decoder with skip connections. Fine-tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models found this decoder important for dense prediction from geospatial foundation models, and it also found that LoRA preserves pretrained features out of region. We add a per-pixel scale gate, a small attention module that weights decoder levels in the spirit of A Novel Attention Fully Convolutional Network Method for Synthetic Aperture Radar Image Segmentation. The gate's output weights determine which decoder level supplies $z_i$ for the neighbor search. The intent is that interior pixels match on coarse context and boundary pixels match on fine detail.
This component is secondary, and we treat it as a hypothesis rather than a requirement. If fixing the neighbor search at a single scale does as well, we will drop the gate. One concern from the detail-preservation literature applies directly: raw skip features can carry shadow and clutter into the decoder. We will check whether fine-scale neighbor search is fooled by those textures.
Research Program
The program has three parts. The first is diagnostic. On real weak-label data, we measure how distractor agreement between two pixels changes with their spatial separation. This decides whether the premise holds before any method is built. If the variogram is flat, the proposal reduces to ordinary pixel-wise PLL and we say so.
The second part builds candidate-set benchmarks from existing datasets. On Sen1Floods11 (Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1), each pixel's candidate set over {permanent water, flood water, land} is the union of the labels assigned by three sources: the dataset's Sentinel-2 index weak labels, Otsu thresholding of the SAR image, and the permanent-water layer. This is a naturally occurring set, and its errors follow event and biome. We evaluate against the hand-labeled chips and the Geographic Hold-Out Set.
On LoveDA (LoVEDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation), which has 7 classes and 18 districts, we coarsen the clean masks into cells. Each fine pixel's candidate set is the set of classes present in its cell. We then add distractor fields whose spatial correlation length we control, ranging from independent per pixel to district-wide. This controlled sweep is what tests the core claim.
The third part trains, compares, and analyzes the methods, with all splits spatially blocked by district or flood event, following the evaluation warnings in Deploying an Artificial Intelligence Application to Detect Flood from Sentinel 1 Data.
Baselines
The naive reference treats each pixel's candidate set as a single noisy hard label and trains with cross-entropy. For LoveDA we take the majority class of the cell. For Sen1Floods11 we take the Sentinel-2 weak label, which is what Sen1Floods11 itself did and reported a flood-water mIoU of 0.3389. The most obvious PLL baseline applies candidate-masked PLL losses independently to each pixel, using the CAVL selection rule and a PRODEN-style reweighting. A "drop ambiguity" baseline trains only on pixels whose candidate set has one class and ignores the rest.
The strongest competitor is the same pipeline with ordinary feature-graph propagation ($\rho = 0$), which is a pixel-level version of the graph propagation in DualG. It has every advantage of our method except the geographic weighting. It therefore isolates the key ingredient. To control for the fact that excluding neighbors also makes the graph sparser, a further ablation removes a random set of edges matched in number to the edges our weighting suppresses.
All methods share the same backbone, decoder, data, and memory-bank size. On Sen1Floods11 this is Prithvi 2.0 with LoRA. On LoveDA it is HRNet, to stay comparable with that paper. Each method gets an equal hyperparameter search budget, tuned on a spatially separate validation block.
Clean-label training is an upper bound. The reported reference points for that bound are Prithvi 2.0 with LoRA on the Sen1Floods11 Geographic Hold-Out Set at 87.57% mIoU, from Fine-tune Smarter, and HRNet on the LoveDA test set at 49.79% mIoU, from LoveDA. LoveDA's cross-domain numbers give the out-of-domain range: rural-to-urban 41.32% for CBST against an oracle of 45.67%.
We found no reported result on Lacuna for pixel-level PLL on either dataset. "Beating" therefore means two things. Our method must exceed the $\rho = 0$ graph by a margin larger than the seed-to-seed spread on held-out regions. It must also close a clearly larger share of the gap between the naive baseline and the clean-label bound. Both should hold for overall mIoU and for boundary F-score.
Why not just…?
Why not just ignore ambiguous pixels and train on the confident ones? This is the most serious threat. If ambiguity sits at boundaries, discarding it costs little interior accuracy. But it also removes almost all boundary supervision. The approach also fails when distractors are spread across a region, because whole regions then lose their labels. The correlation-length sweep and the boundary F-score separate the two cases. If dropping ambiguous pixels matches us on both measures, we will report that pixel-level disambiguation is unnecessary in that regime.
Why not just apply an existing PLL loss to each pixel? Pixel-wise CAVL or PRODEN ignores neighborhood evidence entirely. CAVL's own analysis shows that it locks in early mistakes. These baselines are in the comparison. We expect them to be competitive only when distractors are independent per pixel.
Why not just use a geospatial foundation model with LoRA and a noise-robust loss? A strong backbone helps every method equally. The LoRA backbone is shared by all methods here, so this alternative becomes the naive and per-pixel baselines on top of it. Fine-tune Smarter shows that the backbone resists forgetting under geographic shift. It says nothing about region-correlated label errors, which a better feature extractor can fit just as easily.
Why not use unsupervised domain adaptation? It addresses a different gap: unlabeled target regions rather than ambiguous source labels. LoveDA reports that self-training methods degrade when class distributions differ between domains. The two approaches can be combined, but adaptation would not remove the systematic errors inside the source labels.
Decisive Checks
The central experiment is the LoveDA correlation-length sweep. Our method's advantage over the $\rho = 0$ graph should be near zero when distractors are independent per pixel. It should grow steadily as the correlation length grows toward district scale. It should be larger on held-out districts than on held-out tiles from the training districts.
A second check measures how often neighbor votes are correct, binned by spatial distance, on Sen1Floods11's naturally ambiguous labels. Vote precision should rise with distance if the premise holds.
Two further results would sink the specific mechanism even if the overall method does well. One is the random edge-removal control matching our method. The other is a variogram-estimated $\rho$ doing no better than an arbitrary fixed exclusion radius. Finally, comparing fixed-scale and gated neighbor search, scored separately on boundary pixels and interior pixels, decides whether multi-scale attention earns its place in the method.
References
- Solving the Partial Label Learning Problem: An Instance-Based Approach
- Partial Multi-Label Learning via Credible Label Elicitation
- Dual Graph Disambiguation for Multi-Instance Partial-Label Learning
- Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning
- Exploiting Class Activation Value for Partial-Label Learning
- Fine-Tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models
- A Novel Attention Fully Convolutional Network Method for Synthetic Aperture Radar Image Segmentation
- Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1
- LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation
- Deploying an Artificial Intelligence Application to Detect Flood from Sentinel 1 Data