PaperlayerView exact revision on arXiv

Source-linked AI summary

A large-scale climate-aware satellite image dataset for domain adaptive land-cover semantic segmentation

Songlin Liu, Linwei Chen, Li Zhang, Jun Hu, Ying Fu

arXiv:2610.21238v1cs.CV

TL;DR

Existing land-cover datasets overlook climate diversity, even though climate-linked land-cover differences create domain gaps. The paper introduces CASID, a large-scale dataset spanning four climate domains, and analyzes climate-related image differences. DeepLabV2’s cross-climate score illustrates the resulting challenge, reaching 12.2 mIoU when trained on TemMs and evaluated on TroRf.

  • Problem

    Existing land-cover datasets overlook climate-driven geographic diversity, while land-cover differences across climate zones create domain gaps that can impair model effectiveness on new zones.

    Source: §1 · pp. 1–2

  • Method

    The authors introduce CASID, a 980-image dataset spanning four climate domains, analyze climate-related image characteristics, and evaluate segmentation and unsupervised domain adaptation methods.

    Source: §1 · p. 2

  • Results

    12.2 mIoU is DeepLabV2’s score when trained on TemMs and evaluated on TroRf, illustrating cross-climate performance degradation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    CASID offers a challenging multi-climate benchmark, and its evaluations provide a baseline for future land-cover segmentation and domain-adaptation research.

    Source: §1 · p. 2

Abstract

from arXiv · show

A few well-annotated datasets for land-cover semantic segmentation have recently been introduced to advance the field of earth observation technologies. However, these datasets overlook the significant diversity among geographic areas with different climates, which can greatly impact and diversify land cover. Consequently, this leads to a domain gap in remote sensing images and severe performance degradation of the segmentation models. To enhance land-cover semantic segmentation with improved generalization ability, we conducted the first investigation into the impact of climate on this task. In this paper, we present a unique large-scale Climate-Aware Satellite Images Dataset (CASID) specifically designed for domain adaptive land-cover semantic segmentation. It consists of 980 satellite images with a size of 5000 x 5000 pixels, collected from 30 different regions around Asia, covering over 24,500 square kilometers. These images are gathered from four distinct climate zones, namely temperate monsoon, subtropical monsoon, tropical monsoon, and tropical rainforest. It includes four sub-datasets/domains, each representing one of the aforementioned climate zones. This characteristic makes CASID the first climate-aware land-cover semantic segmentation dataset with multiple domains. Additionally, we provide a comprehensive analysis of the samples from the four climate zones, emphasizing differences in global image features, image texture, category distribution, spectral value, and object shape. These analyses offer valuable insights for subsequent research in this field. Moreover, we conduct extensive experiments to evaluate the latest semantic segmentation and unsupervised domain adaptation methods on the CASID dataset. These results serve as a robust baseline for future research endeavors. Our dataset will be made publicly available at https://github.com/Linwei-Chen/CASID.

1. Introduction

Existing annotated land-cover datasets omit climate diversity, although climate-linked differences in land-cover characteristics create cross-zone domain gaps and can undermine model transfer. The paper introduces CASID, a four-domain dataset, analyzes climate-related image differences, and evaluates segmentation and adaptation methods.

1. Introduction

  • Climate-related differences in land-cover characteristics, including vegetation distributions and forest spectra, contribute to domain gaps that can reduce cross-climate model effectiveness.The introduction notes that temperate monsoon and tropical rainforest zones can differ substantially in forest distribution and pixel spectra.Source: §1 · pp. 1–2
  • CASID contains 980 5,000 × 5,000-pixel images from 30 Asian regions across four climate-domain subdatasets, covering over 24,500 km² and five labeled classes.The climate domains are temperate monsoon, subtropical monsoon, tropical monsoon, and tropical rainforest; labels cover background, building, forest, road, and water.Source: §1 · p. 2
  • The paper analyzes four climate zones for differences in global image features, texture, category distribution, spectral values, and object shape.These analyses are presented as insights for subsequent research.Source: §1 · pp. 2–3
  • 12.2 mIoU is DeepLabV2’s score when trained on TemMs and evaluated on TroRf, exemplifying cross-climate degradation.The cross-domain visualizations also show inaccurate predictions on TroRf from models trained on TemMs, SubMs, and TroMs.Source: Table 1 · p. 2, §1 · p. 2, Figure 1 · p. 2
  • The authors evaluate state-of-the-art semantic segmentation and unsupervised domain adaptation methods on CASID as a baseline for future research.The introduction describes domain adaptation experiments as targeting performance degradation associated with the climate domain gap.Source: §1 · pp. 2–3

2. Related works

The related work reviews land-cover datasets and segmentation methods before motivating climate-aware data and domain adaptation. Existing datasets and methods leave gaps in geographic diversity and climate-focused adaptation.

2.1. Land-Cover Datasets for Semantic Segmentation

  • Earlier land-cover datasets provide pixel-wise labels, but their limited area and sample counts constrain algorithm generalization; later large-scale datasets still underrepresent geographic domain gaps.The review names OpenEarthMap as a global benchmark, while noting climate’s impact has received limited investigation.Source: §2.1 · p. 3

2.2. Semantic Segmentation

  • Land-cover segmentation assigns a label to each pixel; prior methods address foreground-background imbalance and small objects amid complex backgrounds.FarSeg uses relation-based and optimization-based foreground modeling, while FactSeg uses a dual-branch decoder and collaborative probability loss.Source: §2.2 · p. 3

2.3. Unsupervised Domain Adaptation

  • Unsupervised domain adaptation transfers a source-trained model to a target domain, with adversarial training and self-training as two main branches.Adversarial approaches align domain features or structured outputs, while self-training methods generate target pseudo-labels for alternating fine-tuning.Source: §2.3 · pp. 3–4
  • Remote-sensing domain adaptation has been explored, but limited datasets with consistent annotation granularity and high diversity have constrained its advancement.CASID is proposed as a more challenging benchmark for future remote-sensing unsupervised domain adaptation research.Source: §2.3 · p. 5

3. Dataset Description

CASID provides large, labeled satellite images from four climate domains for land-cover segmentation and unsupervised domain adaptation. Its analyses find that image features, texture, category distributions, spectral values, and object shapes vary across climates.

3.1. Overview and Dataset Properties

  • CASID collects satellite imagery from 30 Asian cities/regions across more than 24,500 km2, with four climate-specific domains and pixel-wise labels for five categories: background, building, forest, road, and water.Approximately 70% of images are assigned to training and the remaining ∼30% to testing.Source: §3.1 · pp. 6–7

3.2. Statistics Analysis for Climate Domain Gap

  • t-SNE features separate the four climate domains, with TemMs and TroRf most distinct and SubMs and TroMs mostly between them.The visualization’s domain pattern resembles the geographic distribution of climate zones.Source: §3.2 · p. 9
  • GLCM texture properties—contrast, dissimilarity, homogeneity, and energy—change gradually from TemMs to TroRf.The analysis reports that image texture is associated with the domain gap and correlated with climate zone.Source: §3.2 · p. 9
  • Forest pixels account for 76.2% of the tropical rainforest area versus 45.4% of the temperate monsoon area.The paper reports significant class-distribution variation among climate zones.Source: Figure 7 · p. 8, §3.2 · p. 9
  • The same land-cover category has different spectral values across climates; tropical rainforest forest has lower mean red and blue values.The paper also notes distinct RGB histograms by domain and higher mean red values for temperate-monsoon background.Source: Figure 9 · p. 8, Figure 8 · p. 8, §3.2 · p. 9
  • Object sizes and polygon vertex counts vary across climates; roads are thin, long, and curved, while forests cover larger areas.The paper interprets average size as object scale and vertex count as shape complexity.Source: Figure 10 · p. 9, Figure 11 · p. 9, §3.2 · p. 10

4. Experiments for Land-Cover Semantic Segmentation

The experiments benchmark semantic segmentation methods on CASID using pixel-level accuracy metrics and shared training protocols. SegNeXt achieves the strongest reported mIoU, while qualitative results highlight persistent difficulty with roads.

4. Experiments for Land-Cover Semantic Segmentation

  • The benchmark includes representative remote-sensing segmentation networks, from FCN8S and DeepLabV3+ to transformer-based and CNN-based methods.The selected methods include FCN8S, DeepLabV3+, PAN, U-Net, U-Net++, PSPNet, LinkNet, FarSeg, FactSet, HRNet, SegFormer, PoolFormer, and SegNeXt.Source: §4 · p. 10

4.1. Evaluation Metric

  • IoU measures segmentation accuracy by class, mIoU averages IoU across categories, and the experiments also report Dice.The training protocol uses 40k iterations, batch size 8, and randomly cropped 512×512 patches; experiments are repeated five times with mean and variance reported.Source: §4.1 · pp. 10–11, §4.2 · p. 12

4.2. Implementation details

  • Training uses a common data split and optimization setup, with original training settings retained for SegFormer, PoolFormer, and SegNeXt.The common setup includes SGD with momentum 0.9, weight decay 1e-4, and a poly learning-rate schedule; SegFormer uses AdamW with initial learning rate 6e-5.Source: §4.2 · pp. 11–12

4.3. Quantitative Results

  • 63.4 mIoU is the highest reported result, achieved by SegNeXt, ahead of SegFormer at 63.0 and PoolFormer at 62.7.LinkNet reaches 62.3 mIoU; road and water are more challenging than building and forest.Source: §4.3 · p. 12

4.4. Qualitative Results

  • Roads are difficult to segment because their thin, elongated shapes can be obscured, while forest predictions benefit from spectral homogeneity and distinct spectral features.HRNet provides the best road visualization, attributed to high-resolution retention and multi-scale fusion.Source: §4.4 · pp. 12–13

5. Experiments for Unsupervised Domain Adaptation

The UDA experiments compare cross-domain performance on CASID with existing datasets and test several adaptation methods across climate domains. CASID shows larger cross-domain gaps, while adaptation improves results but does not eliminate difficult domain shifts.

5. Experiments for Unsupervised Domain Adaptation

  • The evaluation covers adversarial-training and self-training UDA approaches, plus recent ASA and DAFormer methods, using mIoU and five-run mean and variance results.Methods include AdaptSeg, CLAN, FADA, CBST, and PyCDA; all UDA methods share a feature extractor and discriminator by default.Source: §5 · p. 13, §5.1 · p. 13

5.1. Implementation details

  • UDA implementation uses DeepLabV2 with ResNet-50 as the default feature extractor and a fully convolutional discriminator, with method-specific training settings.Adversarial training runs for 10k iterations with batch size 8; source and target images are input alternately.Source: §5.1 · p. 14

5.2. Quantitative Results

  • 31.4∼51.5 mIoU is CASID’s cross-domain performance gap, versus 14.4∼15.1 on LoveDA and 17.3∼25.6 on OpenEarthMap.The paper presents the larger gap as evidence that CASID is a more challenging benchmark for semantic segmentation and multi-domain UDA.Source: §5.2 · p. 14
  • UDA improves cross-domain results, but sizable climate shifts can leave performance below target-only results; TroRf→SubMs with FADA reaches 54.4 mIoU versus 60.9 mIoU for SubMs→SubMs.For SubMs→TemMs, AdaptSeg reaches 35.7 mIoU, comparable to the TemMs→TemMs result; the best TroMs→TemMs result with DeepLabV2 is 31.0 mIoU using CALN.Source: §5.2 · p. 15
  • Reducing sampled regions lowers intra-domain variance and UDA performance, while uniform sampling at the same sample counts preserves higher variance and better performance.TemMs-100% has σ = 0.64 and TemMs-25% has σ = 0.32; the uniformly sampled variants have variance closer to full TemMs, and TemMs-50%-U performs comparably to TemMs-100%.Source: §5.2 · p. 16, Table 11 · p. 16

5.3. Qualitative Results

  • Visualizations show source-only models failing under climate shifts; UDA improves segmentation, but road and water accuracy and segmentation consistency remain unsatisfactory.The paper notes that false segmentations persist even with DAFormer; Figures 15 and 16 show domain-transfer examples, and Figure 17 presents failure cases.Source: Figure 15 · p. 14, Figure 16 · p. 15, §5.3 · p. 16

6. Conclusion

The conclusion presents CASID as a climate-aware, multi-domain dataset and a benchmark for land-cover segmentation and UDA. It identifies future work in developing climate-focused adaptation methods and expanding the dataset globally.

6. Conclusion

  • CASID comprises 980 5,000 × 5,000-pixel images from 30 Asian regions across four climate zones, establishing a multi-climate-domain dataset for land-cover segmentation.The paper also reports statistical differences in image features, texture, category distribution, spectral values, and object shapes across zones.Source: §6 · p. 17
  • The authors conclude that CASID can benchmark land-cover semantic segmentation and UDA, with experiments providing a baseline for future research.The paper frames the dataset as supporting research on domain-adaptive land-cover segmentation and potentially robust vision algorithms beyond remote sensing.Source: §6 · p. 17
  • Future work will develop an unsupervised domain-adaptation method for climate-caused domain gaps and expand CASID to more climate zones and global samples.The proposed method development is intended to use CASID as its foundation.Source: §6 · p. 17

Related papers

Loading 2610.21238v1…