Fine-resolution land-use segmentation and synthetic image-label augmentation in the Kathmandu Valley

Pranil Parajuli, Sugam Dahal, Kiran Chandra DahalarXiv 2026 · arXiv

High-resolution land-use data is a luxury often missing in the rapidly urbanizing cities of the Global South. While global satellite products offer 10-meter resolution, a single pixel in a city like Kathmandu can swallow an entire house and its adjacent street. A new paper by Pranil Parajuli and colleagues tackles this "data deficit" head-on, introducing a challenging new dataset and revealing some uncomfortable truths about how we evaluate and augment remote sensing models.

TL;DR

The authors introduce a fine-resolution (0.3m) land-use dataset for the Kathmandu Valley and demonstrate that standard random data splitting severely inflates performance due to spatial leakage. They propose an enhanced transformer architecture that improves thin-structure segmentation, but find that blindly augmenting training with diffusion-generated synthetic data actually degrades accuracy for the rarest classes.

What Problem Does This Paper Address?

Most semantic segmentation benchmarks are built on the "planned" geographies of the Global North or China, featuring regular grids and distinct spectral bands (like near-infrared). The Kathmandu Valley is different: it’s dominated by small, irregular, densely packed parcels where different land uses look nearly identical in standard RGB imagery.

The authors identify three primary hurdles:

  1. Spectral Inseparability: Classes like "Unused Land" and "Agricultural Land" are virtually indistinguishable by color alone (see Figure 3c).
  2. Geometric Under-resolution: Key features like rivers and roads are often only 2-3 pixels wide at the model's prediction stride, making them easy to lose during downsampling.
  3. Extreme Class Imbalance: The ratio between the most and least common classes is a staggering 94:1.
Figure 3: Properties of the corpus that make it hard. (a) Share of annotated pixels by class. (c) Jeffries–Matusita separability for every class pair.

Methods

The study follows a rigorous two-part pipeline (Figure 5). First, they benchmark five architectures under two partitioning regimes. Second, they test whether a two-stage diffusion generator can mitigate class imbalance.

The TC-SegFormer

The authors modified a SegFormer-B0 backbone with three "thin-structure-aware" components:

  • Residual Detail Path: A shallow convolutional branch that processes the image at higher resolution to recover fine details lost by the transformer's hierarchical encoder.
  • Reliability Reweighting: A novel weighting scheme that discounts pixels near hand-drawn boundaries, where annotations are most likely to be noisy.
  • Topology-Aware Loss: Integration of soft centreline Dice (clDice) to ensure roads and rivers maintain connectivity rather than just pixel accuracy.

Spatial Splitting Protocol

To combat spatial leakage—where a model "memorizes" a neighborhood because adjacent tiles appear in both training and validation—the authors compared standard random splits against a strict sequence-block split. This proxy for geographical blocking ensures that validation tiles are spatially distant from training tiles (Figure 4).

Figure 4: Two partitioning regimes. (a) Sequence-block fold showing dropped buffers. (b) Random split of the same scenes.

Datasets

  • Kathmandu Valley Dataset (New): 408 hand-annotated RGB scenes covering 202.4 km². It features six classes: Residential, Agricultural, Unused Land, Forest, Road, and River.
  • LoveDA: An external benchmark used to validate that the proposed model improvements weren't just over-fitting the new Kathmandu data.

Key Insights and Ideas

  • Random Splits are "Cheating": Switching from random to block splitting dropped performance by up to 0.095 mIoU. This "spatial penalty" is often larger than the performance gap between different model architectures, suggesting that many published results in remote sensing may be significantly over-optimistic.
  • The Synthetic Data Trap: While diffusion models can generate visually stunning satellite imagery, they often fail at the topology of rare classes. The generator produced fragmented, "blobby" rivers instead of continuous ones. Training on these flawed samples taught the segmentation models incorrect structural priors, harming the very classes the authors hoped to help.
  • Denoising Loss is a Poor Proxy: When training generative models for downstream tasks, the standard "denoising error" didn't correlate with how well the generated labels actually helped the segmentation model. Layout fidelity (how well a frozen scorer can read back the intended labels from a generated image) is a much better metric for checkpoint selection.

Key Results

  • Architecture Gains: The proposed TC-SegFormer outperformed its backbone in all six fold-runs, adding only 1.9% more parameters.
  • Segmentation Performance: On the Kathmandu dataset, models achieved mIoU scores between 0.39 and 0.47 under strict block splitting.
  • Augmentation Failure: Adding 804 synthetic image-label pairs reduced mIoU across every tested architecture by an average of 0.014. The "River" class saw a relative decline of 18.2% in IoU when synthetic data was included.

Limitations and Future Work

The biggest structural limitation is the lack of a near-infrared (NIR) band, which the authors believe is the "spectral ceiling" preventing better separation of unused and agricultural land.

Future work should move away from "blind" synthetic augmentation. Instead of focusing on making synthetic images look more realistic, researchers should focus on layout models that explicitly preserve the connectivity and width of thin structures like roads and rivers. Finally, the authors urge the community to always report results under spatial/block partitioning to provide a more honest assessment of model generalization.

Summary updated

Abstract

Parcel-scale land-use data is missing for most rapidly urbanizing Global South cities, as current segmentation architectures rely on multispectral imagery and planned layouts. We introduce a hand-annotated six-class land-use corpus covering 202 km$^2$ of Kathmandu Valley across 408 fine-resolution RGB scenes. The dataset is uniquely challenging: RGB spectral similarity is high, class imbalance approaches 100:1, and key structural features like rivers measure just two to three pixels wide at a standard hierarchical decoder's prediction stride. Evaluating five architectures reveals that standard random data splitting conceals massive spatial leakage. Switching to a strict sequence-block split drops performance by up to 0.095 mIoU. This penalty exceeds the performance gap between the architectures themselves. We modified the SegFormer baseline by integrating a half-resolution residual detail path, boundary-reliability reweighting, a topology-aware loss, and a presence-balanced sampler. Adding only 1.9% more parameters, the enhanced architecture improves mean IoU across every test fold, with gains concentrated in rare, thin-structure classes. However, both models achieve identical results on the external LoveDA benchmark. Finally, augmenting with diffusion-generated pairs at matched compute degraded accuracy across all architectures, especially harming rare classes it aimed to improve. This demonstrates that blindly incorporating synthetic data can severely damage performance, highlighting a critical need to rigorously evaluate and filter generated samples before training.
BibTeX
@misc{parajuli2026fine,
  author = {Pranil Parajuli and Sugam Dahal and Kiran Chandra Dahal},
  title = {{Fine-resolution land-use segmentation and synthetic image-label augmentation in the Kathmandu Valley}},
  year = {2026},
  eprint = {2611.10847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10847},
}
Figure