cs.CVOct 15, 2026

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

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

Organizations: School of Remote Sensing Information Engineering, Wuhan University, 430079, Wuhan, Hubei, China · State Key Laboratory of Geo-information Engineering, 710054, Xi’an, Shaanxi, China · Advanced Research Institute of Multidisciplinary Science, Beijing Institute of Technology, 100081, Beijing, China · School of Computer Science and Technology, Beijing Institute of Technology, 100081, Beijing, China · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Table 1: Results of semantic segmentation model ( Chen et al., 2017 ) trained and evaluated on images from different climate zone. The model learned from a certain climate zone exhibits severe performance degradation when evaluated on images from different climate zone.
Figure 1: Visualized results of semantic segmentation model ( Chen et al., 2017 ) . The “A →\rightarrow B" notation indicates that the model is trained on domain A and evaluated on domain B. Due to the climate domain gap, models trained on TemMs, SubMs, and TroMs fail to predict accurate results on TroRf climate zone.
DatasetYearResolution (m)Area (km 2 )Image sizeImagesCategoriesDomainsSS TaskUDA Task
ISPRS Vaihingen20130.091.41887∼38161887\sim 38163361✓\checkmark 
ISPRS Postdam20130.053.46000×60006000\times 60003861✓\checkmark 
Zurich Summer ( Volpi and Ferrari, 2015 )20150.69.41000×11501000\times 11502081✓\checkmark 
USGS SIRI-WHU ( Zhong et al., 2015 )20150.633.509000×100009000\times 10000141✓\checkmark 
EvLab-SS ( Zhang et al., 2017 )20170.1∼20.1\sim 2722.04500×45004500\times 450060111✓\checkmark 
SpaceNet ( Van Etten et al., 2018 )20170.52544.0406∼439406\sim 439600021✓\checkmark 
Table 2: Comparison between our CASID and the main land-cover datasets for semantic segmentation (SS) and unsupervised domain adaptation (UDA).
Figure 2: The distribution of geographical locations of images in CASID. The “ □\Box ” indicates the sampling areas. The blue, green, yellow, and red in the climate map indicate four typical climates, i.e ., temperate monsoon, subtropical monsoon, tropical monsoon, and tropical rainforest ( Shuzhen, 1980 ) . We provide visual examples of representative images from the four climate zones along with their corresponding pixel-wise land-cover labels.
ClimateRegionImagesTrain/Test splitSubtotal
Temperate monsoonHokkaido, Japan21Train: 201284
(TemMs set)Seoul, South Korea55
Zhangjiakou, Hebei, China35
Weifang, Shandong, China48
Daqing, Heilongjiang, China42
Belkhabarovsk, Russia35Test: 83
Table 3: The division of the CASID dataset. Images are collected from 30 regions, covering 4 typical climate zones. For each climate zone, we choose two regions as the test set and leave the rest as the train set. The train set and test set account for around 70% and 30% of the whole dataset, respectively.
Figure 3: The visualization of global image features analysis on CASID. We employ two-dimensional t-SNE ( Van der Maaten and Hinton, 2008 ) for visualization. Colored squares indicate global image features of 1024×10241024\times 1024 image patches sampled from four climate domains in CASID. To extract the global image features from the image patches, we utilize InceptionV3 ( Szegedy et al., 2016 ) . The colors blue, green, yellow, and red represent image patches from the following climate domains: temperate monsoon, subtropical monsoon, tropical monsoon, and tropical rainforest, respectively. After applying t-SNE, we observe that the distribution of image features from the four climate domains closely resembles the distribution of geographical locations shown in Figure 2 . This demonstrates the correlation between land cover and climate zone.
Figure 4: The visualization of global image features analysis on CASID. We use two-dimensional t-SNE ( Van der Maaten and Hinton, 2008 ) for visualization. The symbol “ ∘\circ " indicates the marginal data, which are predicted by the k-Nearest Neighbors (k-NN) algorithm. If a data point’s nearest neighbor belongs to a different category, then that data point can be considered one of the marginal points, i.e ., they are at the boundary of different domains. Zoom for better details.
Figure 5: The visualization of grey level co-occurrence matrix (GLCM) ( Haralick et al., 1973 ) for each domain in the CASID.
Figure 6: The visualization of four GLCM properties on the CASID, i.e ., contrast, dissimilarity, homogeneity, and energy. Blue: TemMs, Green: SubMs, Yellow: TroMs, Red: TroRf.
Figure 7: Statistics for category distributions. We present the category distribution and corresponding pixel sample numbers for each climate zone. These results demonstrate that the distribution of categories varies across climate zones.
Figure 8: Image coloration analysis. We show the histograms of red, green, and blue color for each domain in the CASID. The histogram of each domain is very different from the other domains. This difference may contribute to the climate domain gap.
Figure 9: Statistics for spectral values. We present the mean and standard deviation (sigma) for the red, green, and blue spectral values of each category in the four climate zones. Our findings indicate that the spectral values of the same category vary significantly across different climate zones, resulting in large intra-class variance.
Figure 10: Statistics of object shape. The average size depicts the scale of objects, and the number of vertices in an object polygon represents the complexity of its shape.
Figure 11: Histogram of object shape distribution. The size depicts the scale of objects, and the vertice represents the complexity of its shape.
MethodsBackboneIoU (%)mIoUDice
BkgdBldgFrstRdWtr
FCN8S ( Long et al., 2015 )VGG-1651.975.574.732.251.657.2 ( ±\pm 0.3)71.3 ( ±\pm 0.3)
U-Net ( Ronneberger et al., 2015 )ResNet-5053.377.576.341.060.961.8 ( ±\pm 1.2)75.6 ( ±\pm 1.0)
PSPNet ( Zhao et al., 2017 )ResNet-5053.777.175.334.045.357.1 ( ±\pm 0.4)71.2 ( ±\pm 0.3)
LinkNet ( Chaurasia and Culurciello, 2017 )ResNet-5052.878.076.840.863.462.3 ( ±\pm 0.2)75.8 ( ±\pm 0.2)
DeepLabV3+ ( Chen et al., 2018 )ResNet-5053.476.274.538.942.556.9 ( ±\pm 0.5)71.4 ( ±\pm 0.6)
Table 4: Semantic segmentation results on the test set of CASID. We conduct all experiments five times and present the mean and variance results below. Bkgd: Background, Bldg: Building, Frst: Forest, Rd: Road, Wtr: Water.
Figure 12: Representative visualized segmentation results. The top two rows are the samples from TemMs set, and the bottom rows are the samples from SubMs set.
Figure 13: Representative visualized segmentation results. The top two rows are the samples from TroMs set, and the bottom rows are the samples from TroRf set.
Figure 14: Representative visualized segmentation results. From top to bottom are the samples from TemMs, SubMs, TroMs, and TroRf set.
Table 5: Results reported on LoveDA ( Wang et al., 2021a ) with semantic segmentation model ( Chen et al., 2017 ) . It is trained and evaluated on images from different domains.
Table 6: Results reported on OpenEarthMap ( Xia et al., 2023 ) with semantic segmentation model ( Chen et al., 2017 ) . It is trained and evaluated on images from different domains. Asia: AS, Europe: EU, Africa: AF, North America: NA, South America: SA, and Oceania: OC.
Table 7: Results of semantic segmentation model ( Chen et al., 2017 ) trained and evaluated on images from different climate zone on the proposed CASID. TemMs: temperate monsoon, SubMs: subtropical monsoon, TroMs: tropical monsoon, TroRf: tropical rainforest.
Method Domain IoU (%) mIoU (%) ±\pm (%) Source →\rightarrow Target Bkgd Bldg Frst Rd Wtr Target only TemMs →\rightarrow TemMs 49.4 60.9 62.6 24.7 1.9 39.9 0.4 Source only SubMs →\rightarrow TemMs 0.3 8.3 41.2 0.1 0.4 10.0 0.5 AdaptSeg ( Tsai et al., 2018 ) 44.4 60.2 45.3 21.9 6.4 35.7 1.4 CBST ( Zou et al., 2018 ) 37.7 54.0 23.8 27.7 4.3 29.5 1.7 CLAN ( Luo et al., 2019 ) 41.1 59.7 36.0 22.6 7.2 33.3 0.6 PyCDA ( Lian et al., 2019 ) 37.3 32.3 22.8 7.2 0.3 20.0 0.3 FADA ( Wang et al., 2020b ) 48.4 60.1 59.8 19.5 5.0 38.6 0.4 ASA ( Zhou et al., 2020 ) 34.2 28.2 60.0 9.9 4.9 27.4 1.9 DAFormer* ( Hoyer et al., 2022 ) 40.1 64.4 30.0 45.1 18.2 39.6 1.1 Source only TroMs →\rightarrow TemMs 0.5 3.1 56.6 0.0 0.6 12.2 0.2 AdaptSeg ( Tsai et al., 2018 ) 42.3 44.1 29.7 25.1 0.7 28.4 3.1 CBST ( Zou et al., 2018 ) 45.2 39.1 27.2 28.0 0.3 28.0 2.0 CLAN ( Luo et al., 2019 ) 41.8 50.1 34.9 26.6 1.7 31.0 2.0 PyCDA ( Lian et al., 2019 ) 37.7 32.0 19.6 1.9 0.2 18.3 0.2 FADA ( Wang et al., 2020b ) 45.7 47.1 41.3 19.3 0.6 30.8 0.3 ASA ( Zhou et al., 2020 ) 40.0 31.7 36.9 1.8 0.2 22.1 1.9 DAFormer* ( Hoyer et al., 2022 ) 42.2 68.6 38.4 43.3 0.7 40.0 1.2 Source only TroRf →\rightarrow TemMs 3.1 3.0 35.8 0.1 0.5 8.5 0.3 AdaptSeg ( Tsai et al., 2018 ) 38.1 59.2 65.4 24.2 4.2 38.2 0.9 CBST ( Zou et al., 2018 ) 36.1 57.5 59.2 27.9 1.3 36.4 2.1 CLAN ( Luo et al., 2019 ) 37.6 60.3 51.9 25.5 3.5 35.8 0.8 PyCDA ( Lian et al., 2019 ) 30.6 20.3 25.7 1.2 0.2 15.6 0.2 FADA ( Wang et al., 2020b ) 40.1 57.1 64.9 19.3 1.4 36.6 0.1 ASA ( Zhou et al., 2020 ) 37.4 18.7 53.0 46.0 4.0 21.9 2.2 DAFormer* ( Hoyer et al., 2022 ) 45.2 66.1 56.5 39.3 16.8 44.8 1.0 Target only SubMs →\rightarrow SubMs 56.0 78.2 83.8 24.4 62.3 60.9 0.1 Source only TemMs →\rightarrow SubMs 0.7 2.8 41.4 0.1 1.9 9.4 0.9 AdaptSeg ( Tsai et al., 2018 ) 40.3 76.6 74.8 22.5 51.7 53.2 0.9 CBST ( Zou et al., 2018 ) 40.9 75.7 66.3 25.8 29.7 47.7 4.5 CLAN ( Luo et al., 2019 ) 45.5 76.9 77.4 23.5 39.7 52.6 0.6 PyCDA ( Lian et al., 2019 ) 13.3 57.5 44.0 4.4 17.9 27.4 0.8 FADA ( Wang et al., 2020b ) 40.0 75.6 70.9 21.9 49.1 51.5 0.5 ASA ( Zhou et al., 2020 ) 24.2 53.0 62.6 66.0 13.7 30.8 3.6 DAFormer* ( Hoyer et al., 2022 ) 61.5 80.1 84.9 39.3 62.4 65.6 1.0 Source only TroMs →\rightarrow SubMs 1.5 17.2 43.2 0.0 2.2 12.8 0.5 AdaptSeg ( Tsai et al., 2018 ) 50.5 75.7 73.2 27.7 37.0 52.8 1.7 CBST ( Zou et al., 2018 ) 46.6 70.7 64.7 32.5 35.6 50.0 4.1 CLAN ( Luo et al., 2019 ) 50.2 77.9 73.8 29.8 39.1 54.1 0.8 PyCDA ( Lian et al., 2019 ) 37.1 73.0 44.9 4.5 9.3 33.8 0.2 FADA ( Wang et al., 2020b ) 54.1 73.8 80.6 16.6 46.7 54.4 0.2 ASA ( Zhou et al., 2020 ) 41.2 65.6 69.4 1.1 16.7 38.8 2.1 DAFormer* ( Hoyer et al., 2022 ) 54.8 81.6 73.5 44.8 49.5 60.9 0.9 Source only TroRf →\rightarrow SubMs 1.8 5.5 41.1 0.0 2.7 10.2 0.1 AdaptSeg ( Tsai et al., 2018 ) 28.9 75.3 72.2 22.8 60.1 51.9 1.2 CBST ( Zou et al., 2018 ) 42.9 77.5 72.5 32.0 27.8 50.5 2.1 CLAN ( Luo et al., 2019 ) 27.0 76.8 70.8 26.6 48.2 49.9 0.9 PyCDA ( Lian et al., 2019 ) 35.2 69.5 42.3 4.0 8.0 31.8 0.7 FADA ( Wang et al., 2020b ) 24.5 74.2 70.9 18.4 38.9 45.4 0.5 ASA ( Zhou et al., 2020 ) 35.5 63.7 71.4 0.0 1.6 34.4 1.9 DAFormer* ( Hoyer et al., 2022 ) 50.4 80.6 77.0 45.1 59.9 62.6 1.7Domain IoU (%) mIoU (%) ±\pm (%) Source →\rightarrow Target Bkgd Bldg Frst Rd Wtr TroMs →\rightarrow TroMs 68.9 59.8 78.5 14.0 61.4 56.5 0.2 TemMs →\rightarrow TroMs 1.2 1.1 48.8 0.2 3.0 10.9 0.8 59.5 53.1 75.7 2.9 36.6 45.6 3.8 70.1 53.2 81.0 2.3 39.5 49.2 2.6 64.0 56.2 81.9 3.5 42.6 49.7 0.3 59.5 20.3 47.8 0.6 19.1 29.5 0.6 62.5 48.8 75.4 4.5 29.8 44.2 1.1 50.9 19.8 70.3 0.5 51.7 38.6 1.6 73.2 61.4 84.0 34.2 73.0 65.2 1.4 SubMs →\rightarrow TroMs 1.1 4.8 53.3 0.1 4.7 12.8 0.4 71.3 55.5 81.1 10.6 72.4 58.2 1.2 72.1 54.0 82.0 16.5 68.7 58.7 7.7 72.3 56.3 81.2 13.6 67.0 58.1 1.0 70.8 29.1 70.2 1.2 52.7 44.8 2.0 72.5 50.1 80.6 7.9 46.5 51.6 0.5 56.6 30.4 67.1 0.3 16.7 34.2 2.8 69.2 67.8 77.8 37.2 74.6 65.3 1.5 TroRf →\rightarrow TroMs 2.4 1.8 43.9 0.1 6.1 10.9 0.4 65.7 56.1 79.2 7.0 74.8 56.5 0.4 66.3 61.2 78.9 6.8 68.9 56.4 2.1 66.3 56.7 78.7 5.4 76.6 56.8 0.4 64.0 16.6 46.6 0.4 57.4 37.0 2.7 52.0 42.9 73.6 6.5 69.2 48.8 1.0 52.5 28.2 66.1 5.0 33.5 36.1 2.7 63.1 68.6 73.7 33.0 79.4 63.5 1.8 TroRf →\rightarrow TroRf 15.0 70.0 92.8 10.8 56.4 49.0 0.1 TemMs →\rightarrow TroRf 0.4 0.6 58.7 0.1 1.2 12.2 2.5 7.5 66.2 83.3 5.7 14.5 35.4 1.1 6.5 66.0 58.1 18.7 5.6 31.0 2.0 8.6 64.3 81.5 8.9 28.0 38.2 0.5 4.8 31.4 46.8 0.8 4.9 17.8 4.7 7.1 48.0 67.4 8.1 11.4 28.4 0.7 7.6 29.8 83.0 1.8 0.1 24.5 2.1 17.5 53.4 87.1 31.9 42.2 46.4 1.6 SubMs →\rightarrow TroRf 0.4 2.6 70.1 0.1 3.5 15.3 0.7 26.2 70.5 91.3 11.0 21.1 44.0 0.9 24.6 71.6 92.2 25.4 20.4 46.8 1.2 26.1 71.6 93.2 14.7 34.8 48.1 1.7 3.9 17.5 45.7 1.5 9.7 15.6 0.9 20.0 65.7 90.5 12.5 28.0 43.3 0.2 12.5 50.1 87.0 0.5 11.9 32.4 1.2 31.9 75.2 94.0 40.4 75.5 63.4 1.7 TroMs →\rightarrow TroRf 0.7 0.4 41.2 0.0 2.9 9.1 0.4 17.9 58.2 85.0 18.2 29.1 41.7 5.1 17.1 56.9 85.1 22.1 22.0 40.6 3.5 18.2 69.9 85.4 20.3 29.0 44.6 0.9 4.6 18.6 53.6 1.5 17.2 19.1 2.0 19.0 65.2 87.4 19.9 22.3 42.8 0.1 9.1 50.4 81.5 2.6 23.8 33.5 2.3 25.2 74.2 91.2 41.1 57.5 57.8 1.5
Table 8: Unsupervised domain adaptation results on the test set of the CASID dataset. Bkgd: Background, Bldg: Building, Frst: Forest, Rd: Road, Wtr: Water. All experiments are repeated five times, and the mean (mIoU) and variance ( ±\pm ) results are presented. The "*" indicates DAFormer ( Hoyer et al., 2022 ) using stronger SegFormer ( Xie et al., 2021 ) as the segmentation model. All other methods, including "target/source only," use the standard DeepLabV2 ( Chen et al., 2017 ) model.
Figure 15: Visual results for the unsupervised domain adaptation experiments. The “A →\rightarrow B" indicates that the model is trained on domain A and evaluated on domain B. The top three rows are evaluated on the test set of TemMs, and the bottom three rows are evaluated on the test set of SubMs.
Figure 16: Visual results for the unsupervised domain adaptation experiments. The “A →\rightarrow B" indicates that the model is trained on domain A and evaluated on domain B. The top three rows are evaluated on the test set of TroMs, and the bottom three rows are evaluated on the test set of TroRf.
ClimateAreaImagesTrain/Test split
100% Temperate monsoonHokkaido, Japan21Train: 201
(TemMs-100%)Seoul, South Korea55
Zhangjiakou, Hebei, China35
Weifang, Shandong, China48
Daqing, Heilongjiang, China42
50% Temperate monsoonHokkaido, Japan21Train: 98
Table 9: The protocol for reducing source domain samples to reduce variance.
ClimateAreaImagesTrain/Test split
50% Temperate monsoonHokkaido, Japan10Train: 98
(TemMs-50%-U)Seoul, South Korea28
Zhangjiakou, Hebei, China18
Weifang, Shandong, China22
Daqing, Heilongjiang, China20
25% Temperate monsoonHokkaido, Japan6Train: 56
Table 10: The protocol for uniformly reducing source domain samples.
MethodTemMs-25%TemMs-25%-UTemMs-50%TemMs-50%-UTemMs-100%
σ=0.32\sigma=0.32σ=0.60\sigma=0.60σ=0.45\sigma=0.45σ=0.62\sigma=0.62σ=0.64\sigma=0.64
mIoUDicemIoUDicemIoUDicemIoUDicemIoUDice
Target only (TroRf →\rightarrow TroRf)--------49.059.3
Source only (TemMs →\rightarrow TroRf)7.711.612.816.612.015.212.716.912.215.7
AdaptSeg ( Tsai et al., 2018 )30.039.130.739.632.341.532.941.635.459.0
CBST ( Zou et al., 2018 )28.139.632.442.330.640.732.744.231.041.4
Table 11: Experiments for intra-domain variance (IDV). The source domain is TemMs, and the target domain is TroRf. TemMs-25/50% is created by reducing sampling areas, resulting in lower diversity and lower IDV compared to TemMs-25/50%-U which is uniformly sampled on the whole TemMs. A higher standard variance σ\sigma indicates a higher IDV. As a result, methods trained on TemMs-25/50% exhibit lower performance than TemMs-25/50%-U. This indicates that higher IDV is beneficial for increasing performance.
Figure 17: Failure UDA examples on the CASID.
Figure 18: Visualization of the data representations for the source and the target domains. We presented pairwise two-dimensional t-SNE ( Van der Maaten and Hinton, 2008 ) visualizations of 4 domains in the CASID. Different colors indicate different domains, Blue: TemMs, Green: SubMs, Yellow: TroMs, Red: TroRf. The symbol “ ∘\circ " indicates the marginal data, which are predicted by the k-Nearest Neighbors (k-NN) algorithm. If a data point’s nearest neighbor belongs to a different category, then that data point can be considered one of the marginal points, i.e ., they are at the boundary of different domains. (a) TemMs & SubMs. (b) TemMs & TroMs. (c) TemMs & TroRf. (d) SubMs & TroMs. (e) SubMs & TroRf. (f) TroMs & TroRf. Zoom for better details.
More in this topic

Remote Sensing Image Segmentation

Remote sensing image segmentation labels land cover and objects in satellite, aerial, radar, and multispectral imagery.

See all 220 papers

Explore similar work

CardsList
Jun 18, 2026

BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe

Earth observation imagery plays a critical role in environmental monitoring, urban planning, disaster assessment, and climate analysis. While multi-spectral sensors are increasingly available, true-color (RGB) imagery remains widely used due to the power, cost, and deployment constraints of many satellite and aerial platforms. However, existing land-cover segmentation datasets are often limited in geographic coverage, scale, or public accessibility. To bridge this gap, we introduce BELDE (Building a Large-scale Earth-observation Land-cover Dataset for Europe), a publicly available dataset tailored for RGB-based remote sensing semantic segmentation. Constructed from Sentinel-2 true-color images and ESA WorldCover data annotations, BELDE contains 1,088,385 curated image-segmentation map pairs spanning Europe with 7 land-cover classes at 10 m spatial resolution, making it one of the largest publicly available RGB land-cover segmentation datasets for Earth observation. To facilitate cross-region generalization studies, we additionally introduce BELDE-K (16,607 pairs) covering the Republic of Korea and BELDE-CA-NV (88,155 pairs) covering California and Nevada in the United States. We establish baseline results using multiple semantic segmentation architectures and evaluate both in-domain and cross-domain performance. Models trained on BELDE achieve an F1 score of 83.0% on the European test set, while performance decreases to 66.4% on BELDE-CA-NV and 58.3% on BELDE-K, highlighting the challenges posed by out-of-distribution geographic domain shift. By providing a continental-scale RGB segmentation and evaluation benchmark, BELDE supports the development of robust and transferable Earth observation models. The dataset and benchmark resources will be publicly released.

Ümit Mert Çağlar, Alptekin Temizel

Graduate School of Informatics, Middle East Technical University, Ankara, Turkey

Remote Sensing Image SegmentationDistribution ShiftLand Cover ClassificationRemote SensingRemote Sensing Benchmarks
AskPDF
Oct 21, 2026

OpenEarth-OVS44: A Multimodal Remote Sensing Dataset for Open-Vocabulary Segmentation

In remote sensing, open-vocabulary segmentation is particularly important for flexible and scalable land-cover understanding across diverse regions, as it improves generalization to novel categories without requiring exhaustive pixel-level annotations. Despite its great promise, progress in this area remains largely constrained by existing datasets with limited category coverage and overly coarse taxonomies, and therefore struggles to support fine-grained semantic understanding required in real-world applications such as urban functional zoning, carbon accounting, and post-disaster assessment. To address these challenges, we present OpenEarth-OVS44, a multimodal benchmark with 290,587 expert-verified segments from 2,832 RGB images and 1,341 strictly registered RGB-SAR pairs at sub-meter spatial resolution, covering 40 cities across France, Japan, and the United States under a unified 44-class taxonomy organized into 9 base categories. Three evaluation tracks are provided: an RGB-only fine-grained benchmark, a multimodal track with diverse synthetic cloudy conditions for weather-robust evaluation, and a held-out cross-region track. Extensive experiments on representative open-vocabulary segmentation models demonstrate that training on OpenEarth-OVS44 substantially improves fine-grained segmentation performance over the strongest training-free baseline, with improvements of up to 7.1 mIoU points on unseen categories and 27.6 mIoU points on seen categories. The dataset and code will be released.

Yimin Wei, Hongruixuan Chen, Hao Zheng +5

The University of Tokyo · RIKEN AIP · East China Normal University +1

Remote Sensing Image SegmentationOpen-Vocabulary Semantic SegmentationRemote Sensing BenchmarksMultimodal Semantic SegmentationFine-Grained Semantic Segmentation
AskPDF
Apr 10, 2026

Low-Data Supervised Adaptation Outperforms Prompting for Cloud Segmentation Under Domain Shift

Adapting vision-language models to remote sensing imagery presents a fundamental challenge: both the visual and linguistic distributions of satellite data lie far outside natural image pretraining corpora. Despite this, prompting remains the dominant deployment paradigm, driven by the assumption that domain-specific language can guide frozen model representations toward specialized tasks. We test this assumption directly on a domain where the mismatch is prominent: cloud segmentation for satellite imagery. Using CLIPSeg on the CloudSEN12+ cloud segmentation benchmark, we evaluate 60 prompt variants spanning simple labels, domain terminology, appearance descriptors, and contextual cues, finding that every variant underperforms the zero-shot baseline (0.255 mIoU), with engineered prompts scoring as low as 0.07 mIoU. No amount of linguistic refinement bridges the gap between CLIP's natural image representations and satellite spectral imagery. In contrast, supervised fine-tuning with just 0.1% labeled data (~8 images) surpasses zero-shot performance overall, and 5-10% data recovers ~85% of maximum achievable mIoU. Full fine-tuning consistently outperforms low-rank adaptation by 0.03-0.09 mIoU, with the largest gaps for spectrally ambiguous classes, and at 0.5 to 1% labeled data, fine-tuning temporarily degrades performance on these classes before recovering, a supervision dip that aggregate mIoU can mask. For practitioners adapting vision-language models to specialized imagery, our results deliver a clear message: labeled data is not the expensive alternative to prompting; it is the worthwhile path.

Harshith Kethavath, Weiming Hu

University of Georgia, USA

VLM AdaptationRemote Sensing VLMsFine-TuningImage SegmentationRemote Sensing Image Segmentation
AskPDF