LoveDA is a land-cover dataset built around a domain shift

The dataset is designed around a domain shift

LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation, from the NeurIPS 2021 Datasets and Benchmarks track, by Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu and Yanfei Zhong.

The four highlights tell you what the dataset is for. There are 5,987 high spatial resolution, 0.3 m, remote sensing images from Nanjing, Changzhou and Wuhan. There is a focus on different geographical environments between Urban and Rural. It advances both semantic segmentation and domain adaptation tasks. And it carries three considerable challenges: multi-scale objects, complex background samples and inconsistent class distributions.

That second highlight is the design decision everything else follows from. The point is not that the dataset is hard, it is that the training distribution and the test distribution differ in geography. A model that learns to segment urban imagery and is then applied to rural imagery has to generalise across a real change of environment, which is the situation domain adaptation research is about. The three challenges are the same phenomenon seen from three angles: rural imagery contains objects at different scales, more complicated backgrounds and different class frequencies.

The topic list on the repository reflects that: domain-adaptation, pytorch, remote-sensing and semantic-segmentation. The repository layout has two top level directories, Semantic_Segmentation/ and Unsupervised_Domian_Adaptation/, with the spelling as it is, which is the clearest sign that the two task families are the reason the repository exists.

The labels, the no-data regions, and the data loader

The label set is small and deliberately so, which is what makes a seven class problem tractable for a domain shift study.

Category labels: background is 1, building is 2, road is 3, water is 4, barren is 5, forest is 6 and agriculture is 7. And the no-data regions were assigned 0 which should be ignored. A data loader comes with the dataset and will help you construct your pipeline.

Seven classes plus a zero for regions with no label is a very different proposition from the dozens of classes in most land-cover benchmarks, and it is a deliberate choice for a domain adaptation study, where the difficulty is supposed to come from the shift rather than from a long tail of rare classes.

The zero class deserves attention because it is the kind of detail that changes a result. Regions with no label are not background, they are simply unknown, and a pipeline that treats them as background is measuring the wrong thing. That they are separated out and explicitly marked to be ignored is a small piece of design that saves a lot of confusion.

The README also says you can develop your models on the Train and Validation sets, and that you should feel free to design your own models.

The data loader is mentioned as being provided, but the repository listing does not show it at the top level, so it is likely inside the two task directories, and the MMsegmentation integration note in the news section points to the same place for the details.

How to get the data, and the two mirrors

The dataset is released at Zenodo with the DOI 10.5281/zenodo.5706578, and the repository also links a Baidu Drive copy with the access code 27vc.

The Baidu mirror is there because the imagery is large and a mirror closer to the authors' network is a practical matter, and the code in the link is what opens that copy. Zenodo is the citable copy, and the README's citation block includes a dataset entry for it, with the doi 10.5281/zenodo.5706578, alongside the paper entry.

The link structure is worth noting for anyone automating the download. The paper is on ResearchGate, there is a video on SlideShare, the dataset is on Zenodo, the BibTeX entry points at the NeurIPS datasets and benchmarks proceedings, and the two leaderboards are on Codabench, one for the segmentation task and one for the unsupervised domain adaptation task.

That last pair matters. The README says to submit your test results on the LoveDA Semantic Segmentation Challenge and the LoveDA Unsupervised Domain Adaptation Challenge, and that you will get your Test scores smoothly. In other words, the local train and validation splits are for development, and the scored result comes from submitting to the challenge servers, where the test data is held out. The news entries show the evaluation platform was moved to these two challenges on 2021/11/30 and then the platform was moved again on 2026/01/22, so the current locations are the Codabench competitions numbered 13030 and 13037.

What has changed since 2021

The News section is worth reading in order, because it shows the dataset has not stood still even though the release is from 2021.

On 2021/11/11, LoveDA was included in MMsegmentation, and the note in the README says the Semantic Segmentation task can be prepared following the dataset_prepare.md instructions in that repository. On 2021/12/10 it was included in Torchgeo, as a dataset module. On 2021/12/13 the pre-trained URLs for HRNet were updated.

The two more recent entries are larger changes. On 2024/05/12, a new version of the LoveDA dataset was released through the EarthVQA dataset, which is a visual question answering dataset built on the same imagery. On 2026/01/08, the scope of LoveDA was extended from China to a global scale through the EarthVL dataset, and on 2026/01/22 the evaluation platform moved to the two Codabench challenges.

So the original Chinese three-city release is now the starting point of a wider effort. The extension to a global scale is a natural response to the same criticism that applies to any single-region benchmark, and EarthVL and EarthVQA are the venues where that work is happening.

The last push to this repository was on 2026-08-29 and the repository is not archived. There are no GitHub releases, so the commit history is the record, and the code license is recorded as unknown in the repository metadata while the data licence is stated in the README.

The licence, and the Google Earth constraint

The licensing is the part of this dataset that most affects whether you can use it, and it is stricter than the code licence of most research datasets.

The owners of the data and of the copyright on the data are RSIDEA at Wuhan University. Use of the Google Earth images must respect the Google Earth terms of use. And the key sentence: all images and their associated annotations in LoveDA can be used for academic purposes only, but any commercial use is prohibited.

The README points at the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 deed, so the terms are CC BY-NC-SA 4.0, with the non-commercial and share-alike parts spelled out in words rather than left to the licence identifier.

The share-alike part has a consequence that is easy to miss. If you derive a dataset from LoveDA, the derivative inherits the requirement, and the attribution requirement applies to the original images as well as the annotations. A team building a proprietary model on top of this data therefore has two problems: the commercial prohibition, and the share-alike obligation on anything derived from the imagery.

The Google Earth clause is a separate constraint on top of the Creative Commons terms, and it is the one most likely to be overlooked because the README mentions it once. It is also the reason the licence says the images can be used for academic purposes only, since the underlying imagery comes from a service whose own terms govern its use.

For a paper or a public benchmark result, all of this is workable. For anything a company wants to ship, it is not, and the practical alternative is to look at a dataset with a commercial licence.

Where LoveDA fits, and where it stops

The comparison worth making is with the other remote sensing land-cover datasets, and with the synthetic domain shift setups.

Against a single-domain land-cover dataset, the difference is that LoveDA makes the shift the point rather than an accident. Train on one city's imagery and test on another's, and the evaluation measures what the model actually does with unfamiliar terrain rather than how well it memorised a particular distribution. For domain adaptation work, that is the whole experiment in one dataset.

Against a synthetic domain shift, where you apply style transfer or corruptions to a single source dataset to manufacture a target domain, the difference is authenticity. A synthetic shift is easy to control and easy to criticise, because the transformation may not reflect any real variation. LoveDA's shift is geographic and therefore real, at the cost of being harder to analyse, because you cannot tune the difficulty.

The limits are clear from the README. The class set is seven plus an ignored no-data class, which is a small vocabulary by land-cover standards and means a project needing finer distinctions is looking elsewhere. The imagery is Google Earth based, so its licence travels with it in a way that most academic datasets do not. The original release covers three Chinese cities, and while the 2026 extension to EarthVL widens the scope, the dataset that most published work uses is the Chinese one.

For the tooling, the position is comfortable rather than risky. Torchgeo has a dataset module for it, MMsegmentation has a documented preparation flow, and the two Codabench challenges give a scored evaluation rather than a self-reported number. The HRNet pre-trained URLs are provided and were refreshed in December 2021, so the standard baseline is available without training from scratch.

The last thing to weigh is the licence, and it is not a footnote. A dataset whose images and annotations are academic only, with share-alike obligations on derivatives, is a good benchmark and a bad foundation for a product, and that is worth knowing before you plan a research project around it.

Editorial conclusion

LoveDA suits a researcher who wants a benchmark where the domain shift between training and test data is real rather than simulated, since the urban and rural split is inherent in where the imagery was collected, and the seven label classes and the ignored no-data regions are simple enough to start with. It is the wrong tool for a commercial model, because the licence permits academic purposes only and explicitly prohibits commercial use, and it is the wrong tool if you need a wide set of land-cover classes. Before building against it, read the licence section, since the data owners are RSIDEA at Wuhan University and the Google Earth terms apply, and use the Codabench leaderboards rather than only the local train and validation splits, because the test scores are the point of the dataset.

Frequently asked questions

Official sources

1. Issues
2. Junjue-Wang/LoveDA on GitHub
3. README