Conference PaperPDF Available

LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation


Abstract and Figures

Deep learning approaches have shown promising results in remote sensing high spatial resolution (HSR) land-cover mapping. However, urban and rural scenes can show completely different geographical landscapes, and the inadequate generalizability of these algorithms hinders city-level or national-level mapping. Most of the existing HSR land-cover datasets mainly promote the research of learning semantic representation, thereby ignoring the model transferability. In this paper, we introduce the Land-cOVEr Domain Adaptive semantic segmentation (LoveDA) dataset to advance semantic and transferable learning. The LoveDA dataset contains 5987 HSR images with 166768 annotated objects from three different cities. Compared to the existing datasets, the LoveDA dataset encompasses two domains (urban and rural), which brings considerable challenges due to the: 1) multi-scale objects; 2) complex background samples; and 3) inconsistent class distributions. The LoveDA dataset is suitable for both land-cover semantic seg-mentation and unsupervised domain adaptation (UDA) tasks. Accordingly, we benchmarked the LoveDA dataset on eleven semantic segmentation methods and eight UDA methods. Some exploratory studies including multi-scale architectures and strategies, additional background supervision, and pseudo-label analysis were also carried out to address these challenges. The code and data are available at
Content may be subject to copyright.
LoveDA: A Remote Sensing Land-Cover Dataset for
Domain Adaptive Semantic Segmentation
Junjue Wang
, Zhuo Zheng
, Ailong Ma, Xiaoyan Lu, Yanfei Zhong
State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing
Wuhan University, Wuhan 430074, China
Deep learning approaches have shown promising results in remote sensing high
spatial resolution (HSR) land-cover mapping. However, urban and rural scenes
can show completely different geographical landscapes, and the inadequate gener-
alizability of these algorithms hinders city-level or national-level mapping. Most
of the existing HSR land-cover datasets mainly promote the research of learn-
ing semantic representation, thereby ignoring the model transferability. In this
paper, we introduce the Land-cOVEr Domain Adaptive semantic segmentation
(LoveDA) dataset to advance semantic and transferable learning. The LoveDA
dataset contains 5987 HSR images with 166768 annotated objects from three dif-
ferent cities. Compared to the existing datasets, the LoveDA dataset encompasses
two domains (urban and rural), which brings considerable challenges due to the:
1) multi-scale objects; 2) complex background samples; and 3) inconsistent class
distributions. The LoveDA dataset is suitable for both land-cover semantic seg-
mentation and unsupervised domain adaptation (UDA) tasks. Accordingly, we
benchmarked the LoveDA dataset on eleven semantic segmentation methods and
eight UDA methods. Some exploratory studies including multi-scale architectures
and strategies, additional background supervision, and pseudo-label analysis were
also carried out to address these challenges. The code and data are available at
1 Introduction
With the continuous development of society and economy, the human living environment is gradually
being differentiated, and can be divided into urban and rural zones [
]. High spatial resolution (HSR)
remote sensing technology can help us to better understand the geographical and ecological environ-
ment. Specifically, land-cover semantic segmentation in remote sensing is aimed at determining the
land-cover type at every image pixel. The existing HSR land-cover datasets such as the Gaofen Image
Dataset (GID) [
], DeepGlobe [
], Zeebruges [
], and Zurich Summer [
] contain large-scale
images with pixel-wise annotations, thus promoting the development of fully convolutional networks
(FCNs) in the field of remote sensing [
]. However, these datasets are designed only for semantic
segmentation, and they ignore the diverse styles among geographic areas. For urban and rural areas, in
particular, the manifestation of the land cover is completely different, in the class distributions, object
scales, and pixel spectra. In order to improve the model generalizability for large-scale land-cover
mapping, appropriate datasets are required.
In this paper, we introduce an HSR dataset for Land-cOVEr Domain Adaptive semantic segmentation
(LoveDA) for use in two challenging tasks: semantic segmentation and UDA. Compared with the
Equal contribution.
Corresponding author.
35th Conference on Neural Information Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks.
UDA datasets [
] that use simulated images, the LoveDA dataset contains real urban and rural
remote sensing images. Exploring the use of deep transfer learning methods on this dataset will
be a meaningful way to promote large-scale land-cover mapping. The major characteristics of this
dataset are summarized as follows:
1) Multi-scale objects.
The HSR images were collected from
18 complex urban and rural scenes, covering three different cities in China. The objects in the same
category are in completely different geographical landscapes in the different scenes, which increases
the scale variation.
2) Complex background samples.
The remote sensing semantic segmentation
task is always faced with the complex background samples (i.e., land-cover objects that are not
of interest) [
], which is particularly the case in the LoveDA dataset. The high-resolution
and different complex scenes bring more rich details as well as larger intra-class variance for the
background samples.
3) Inconsistent class distributions.
The urban and rural scenes have different
class distributions. The urban scenes with high population densities contain lots of artificial objects
such as buildings and roads. In contrast, the rural scenes include more natural elements, such as
forest and water. Compared with UDA datasets [
] in general computer vision, the LoveDA
dataset focuses on the style differences of the geographical environments. The inconsistent class
distributions pose a special challenge for the UDA task.
As the LoveDA dataset was built with two tasks in mind, both advanced semantic segmentation and
UDA methods were evaluated. Several exploratory experiments were also conducted to solve the par-
ticular challenges inherent in this dataset, and to inspire further research. A stronger representational
architecture and UDA method are needed to jointly promote large-scale land cover mapping.
2 Related Work
Table 1: Comparison between LoveDA and the main land-cover semantic segmentation datasets.
Image level Resolution (m) Dataset Year Sensor Area (km2)Classes Image width Images Task
Meter level 10 LandCoverNet [1] 2020 Sentinel-2 30000 7 256 1980 X
4 GID [38] 2020 GF-2 75900 5 48006300 150 X
Sub-meter level
0.250.5 [2] 2020 Airborne 216.27 3 42009500 41 X
0.6 Zurich Summer [43] 2015 QuickBird 9.37 8 6221830 20 X
0.5 DeepGlobe [9] 2018 WorldView-2 1716.9 7 2448 1146 X
0.05 Zeebruges [24] 2018 Airborne 1.75 8 10000 7 X
0.05 ISPRS Potsdam 32013 Airborne 3.42 6 6000 38 X
0.09 ISPRS Vaihingen 42013 Airborne 1.38 6 18873816 33 X
0.07 AIRS [6] 2019 Airborne 475 2 10000 1047 X
0.5 SpaceNet [41] 2017 WorldView-2 2544 2 406439 6000 X
0.3 LoveDA (Ours) 2021 Spaceborne 536.15 7 1024 5987 X X
The abbreviations are: SS – semantic segmentation, UDA – unsupervised domain adaptation.
2.1 Land-cover semantic segmentation datasets
Land-cover semantic segmentation, as a long-standing research topic, has been widely explored
over the past decades. The early research relied on low- and medium-resolution datasets, such as
MCD12Q1 [
], the National Land Cover Database (NLCD) [
], GlobeLand30 [
], LandCoverNet
], etc. However, these studies all focused on large-scale mapping and analysis from a macro-level.
With the advancement of remote sensing technology, massive HSR images are now being obtained
on a daily basis from both spaceborne and airborne platforms. Due to the advantages of the clear
geometrical structure and fine texture, HSR land-cover datasets are tailored for specific scenes at a
micro-level. As is shown in Table 1, datasets such as ISPRS Potsdam
, ISPRS Vaihingen
, Zurich
Summer [
], and Zeebruges [
] are designed for urban parsing. These datasets only contain a small
number of annotated images and cover limited areas. In contrast, DeepGlobe [
] and
[2] focus on rural areas with a larger scale, in which the homogeneous areas contain few man-made
structures. The GID dataset[38] was collected with Gaofen-2 satellite from different cities in China.
Although LandCoverNet and GID datasets contain both urban and rural areas, the geo-locations of
these released images are private. Therefore, the urban and rural areas are not able to be divided.
In addition, the identifications of cities in released GID images have been already removed so it is
hard to perform UDA tasks. Considering limited coverage and annotation cost, the existing HSR
datasets mainly promote the research of improving land-cover segmentation accuracy, ignoring its
transferability. Compared with land-cover datasets, the iSAID dataset[
] focuses on key objects
semantic segmentation. The different study objects bring different challenges for different remote
sensing tasks.
These HSR land-cover datasets have all promoted the development of semantic segmentation, and
many variants of FCNs [
] have been evaluated [
]. Recently, some UDA methods
have been developed from the combination of two public datasets [
]. However, directly utilizing
combined datasets may result in two problems: 1) Insufficient common categories. Different datasets
are designed for different purposes, and the insufficient common categories limit further exploration.
2) Inconsistent annotation granularity. The different spatial resolutions and labeling styles lead to
different annotation granularities, which can result in unreliable conclusions. Compared with existing
datasets, LoveDA dataset encompasses two domains (urban and rural), representing a novel UDA
task for land-cover mapping.
2.2 Unsupervised domain adaptation
For natural images, UDA is aimed at transferring a model trained on the source domain to the target
domain. Some conventional image classification studies [
] have directly minimized the
discrepancy of the feature distributions to extract domain-invariant features. The recent works have
mainly proceeded in two directions, i.e., adversarial training and self-training.
Adversarial training
. In adversarial training, the architecture includes a feature extractor and a
discriminator. The extractor aims to learn domain-invariant features, while the discriminator attempts
to distinguish these features. For semantic segmentation, Tsai et al. [
] considered the semantic
outputs containing spatial similarities between the different domains, and adapted the structured
output space for segmentation (AdaptSeg) with adversarial learning. Luo et al. [
] introduced a
category-level adversarial network (CLAN) to align each class with an adaptive adversarial loss.
Differing from the binary discriminators, Wang et al. [
] proposed a fine-grained adversarial learning
framework for domain adaptive semantic segmentation (FADA), aligning the class-level features.
From the aspect of structure, the transferable normalization (TransNorm) method [
] was proposed
to enhance the transferability of the FCN-based feature extractors. All these advanced adversarial
learning methods were implemented on the LoveDA dataset for evaluation.
. Self-training involves alternately generating pseudo-labels on the target data and
fine-tuning the model. Recently, the self-training UDA methods have focused on improving the
quality of the pseudo-labels [
]. Lian et al. [
] designed the self-motivated pyramid curriculum
(PyCDA) to observe the target properties, and fused multi-scale features. Zou et al. [
] proposed a
class-balanced self-training (CBST) strategy to sample pseudo-labels, thus avoiding the dominance
of the large classes. Mei et al. [
] used an instance adaptive self-training (IAST) selector for sample
balance. In addition to testing these self-training methods on the LoveDA dataset, we also performed
the pseudo-label analysis for the CBST.
UDA in the remote sensing community
. The early UDA methods focused on scene classification
tasks [
]. Recently, adversarial training [
] and self-training [
] have been studied for UDA
land-cover semantic segmentation. These methods follow the general UDA approach in the computer
vision field, with some improvements. However, with only the public datasets, the advancement
of the UDA algorithms has been limited by the insufficient shared categories and the inconsistent
annotation granularity. To this end, the LoveDA dataset is proposed for a more challenging benchmark,
promoting future research of remote sensing UDA algorithms and applications.
3 Dataset Description
3.1 Image Distribution and Division
The LoveDA dataset was constructed using 0.3 m images obtained from Nanjing, Changzhou and
Wuhan in July 2016, totally covering 536.15
(Figure 1). The historical images were obtained
from the Google Earth platform. As each research area has its own planning strategy, the urban-rural
ratio is inconsistent [52].
Figure 1: Overview of the dataset distribution. The images were collected from Nanjing, Changzhou,
and Wuhan cities, covering 18 different administrative districts.
Data from the rural and urban areas were collected referring to the “Urban and Rural Division Code
issued by the National Bureau of Statistics. There are nine urban areas selected from different
economically developed districts, which are all densely populated (
) [
]. The
other nine rural areas were selected from undeveloped districts. The spatial resolution is 0.3 m, with
red, green, and blue bands. After geometric registration and pre-processing, each area is covered by
1024 ×1024
images, without overlap. Considering Tobler’s First Law, i.e., everything is related to
everything else, but near things are more related than distant things [
], the training, validation, and
test sets were split so that they were spatially independent (Figure 1), thus enhancing the difference
between the split sets. There are two tasks that can be evaluated on the LoveDA dataset:
1) Semantic
. There are eight areas for training, and the others are for validation and testing.
The training, validation, and test sets cover both urban and rural areas.
2) Unsupervised domain
. The UDA process considers two cross-domain adaptation sub-tasks: a)
The images from the Qinhuai, Qixia, Jianghan, and Gulou areas are included in the source training
set. The images from Liuhe and Huangpi are included in the validation set. The Jiangning, Xinbei,
and Liyang images included in the test set. The Oracle setting is designed to test the upper limit of
accuracy in a single domain [
]. Hence, the training images were collected from the Pukou, Lishui,
Gaochun, and Jiangxia areas. b)
Urban. The images from the Pukou, Lishui, Gaochun,
and Jiangxia areas are included in the source training set. The images from Yuhuatai and Jintan are
used for the validation set. The Jiangye, Wuchang, and Wujin images are used for the test set. In the
Oracle setting, the training images cover the Qinhuai, Qixia, Jianghan, and Gulou areas.
With the division of these images, a comprehensive annotation pipeline was adopted, including
professional annotators and strict inspection procedures [
]. Further details of the data division and
annotation can be found in §A.1.
3.2 Statistics for LoveDA
Some statistics of the LoveDA dataset are analyzed in this section. With the collection of public HSR
land-cover datasets, the number of labeled objects and pixels has been counted. As is shown in the
Figure 2(a), our proposed LoveDA dataset contains the largest number of labeled pixels as well as
land-cover objects, which shows the advantage in data diversity. There are a lot of buildings because
urban scenes have large populations (Figure 2(b)). As is shown in Figure 2(c), the background class
contains the most pixels with complex samples [
]. The
complex background samples
larger intra-class variance in the complex scenes and cause serious false alarms.
(a) (b) (c)
Figure 2: Statistics for the pixels and objects in LoveDA dataset. (a) Number of objects vs. number
of pixels. The radius of the circles represents the number of classes. (b) Histogram of the number of
objects for each class. (c) Histogram of the number of pixels for each class.
(a) Class distributions (b) Spectral values (c) Building scales
Figure 3: Statistics for the urban and rural scenes in Nanjing City. (a) Class distribution. (b) Spectral
statistics. The mean and standard deviation (
) for 5 urban and 5 rural areas are reported. (c)
Distribution of the building sizes. The Jianye (urban) and Lishui (rural) scenes are reported.
3.3 Differences Between Urban and Rural Scenes
During the process of urbanization, cities differentiate into rural and urban forms. In this section,
we list the main differences, which reveal the meaning and challenges of the remote sensing UDA
task. For the Nanjing City, the main differences come from the shape, layout, scale, spectra, and class
distribution. As is shown in Figure 1, the buildings in the urban area are neatly arranged, with various
shapes, while the buildings in the rural area are disordered, with simpler shapes. The roads are wide
in the urban scenes. In contrast, the roads are narrow in the rural scenes. Water is often presented in
the form of large-scale rivers or lakes in the urban scenes, while small-scale ponds and ditches are
common in the rural scenes. The agricultural is found in the gaps between the buildings in the urban
scenes, but occurs in a large-scale and continuously distributed form in the rural scenes.
For the class distribution, spectra, and scale, the related statistics are reported in Figure 3. The urban
areas always contain more man-made objects such as buildings and roads due to their high population
density (Figure 3(a)). In contrast, the rural areas have more agricultural land. The
inconsistent class
between the urban and rural scenes increases the difficulty of model generalization.
For the spectral statistics, the mean values are similar (Figure 3(b)). Because of the large-scale
homogeneous geographical areas, such as agriculture and water, the rural images have lower standard
deviations. As is shown in Figure 3(c), most of the buildings have relatively small scales in the rural
areas, representing the “long tail” phenomenon. However, the buildings in the urban scenes have a
larger size variance. Scale differences also exist in the other categories, as shown in Figure 1. The
multi-scale objects
require the models to have multi-scale capture capabilities. When faced with
large-scale land cover mapping tasks, the differences between urban and rural scenes bring new
challenges to the model transferability.
4 Experiments
4.1 Semantic Segmentation
For the semantic segmentation task, the general architectures as well as their variants, and particularly
those most often used in remote sensing, were tested on the LoveDA dataset. Specifically, the selected
networks were: UNet[
], UNet++[
], LinkNet[
], DeepLabV3+[
], PSPNet[
], FCN8S[
], Semantic-FPN[
], HRNet[
], FarSeg[
], and FactSeg[
]. Following the common
], we use the intersection over union (IoU) to report the semantic segmentation
accuracy. With respect to the IoU for each class, the mIoU represents the mean of the IoUs over all
the categories. The inference speed is reported with a single
512 ×512
input (repeated
using frames per second (FPS).
Table 2: Semantic segmentation results obtained on the Test set of LoveDA.
Method Backbone IoU per category (%) mIoU (%) Speed (FPS)
Background Building Road Water Barren Forest Agriculture
FCN8S [19] VGG16 42.60 49.51 48.05 73.09 11.84 43.49 58.30 46.69 86.02
DeepLabV3+ [5] ResNet50 42.97 50.88 52.02 74.36 10.40 44.21 58.53 47.62 75.33
PAN [17] ResNet50 43.04 51.34 50.93 74.77 10.03 42.19 57.65 47.13 61.09
UNet [32] ResNet50 43.06 52.74 52.78 73.08 10.33 43.05 59.87 47.84 71.35
UNet++ [55] ResNet50 42.85 52.58 52.82 74.51 11.42 44.42 58.80 48.20 27.22
Semantic-FPN [16] ResNet50 42.93 51.53 53.43 74.67 11.21 44.62 58.68 48.15 73.98
PSPNet [53] ResNet50 44.40 52.13 53.52 76.50 9.73 44.07 57.85 48.31 74.81
LinkNet [3] ResNet50 43.61 52.07 52.53 76.85 12.16 45.05 57.25 48.50 67.01
FarSeg [54] ResNet50 43.09 51.48 53.85 76.61 9.78 43.33 58.90 48.15 66.99
FactSeg [23] ResNet50 42.60 53.63 52.79 76.94 16.20 42.92 57.50 48.94 65.58
HRNet [45] W32 44.61 55.34 57.42 73.96 11.07 45.25 60.88 49.79 16.74
Table 3: Multi-Scale augmentation during Training and Testing (MSTrTe).
Method mIoU(%)
Baseline +MSTr +MSTrTe
DeepLabV3+ 47.62 49.97 51.18
UNet 48.00 50.21 51.13
SFPN 48.15 50.80 51.82
HRNet 49.79 51.51 52.14
Implementation details
. The data splits followed the Table 8in §A.1. During the training, we used
the Stochastic Gradient Descent (SGD) optimizer with a momentum of
and a weight decay of
. The learning rate was initially set to
, and a ‘poly’ schedule with power
was applied.
The number of training iterations was set to
with a batch size of
. For the data augmentation,
512 ×512
patches were randomly cropped from the raw images, with random mirroring and rotation.
The backbones used in all the networks were pre-trained on ImageNet.
Multi-scale architectures and strategies
. As ground objects show considerable scale variance,
especially in complex scenes (§3.3), we have analyzed the multi-scale architectures and strategies.
There are three noticeable observations from Table 2: 1) UNet++ outperforms UNet due to its nested
cross-scale connections between different scales. 2) Among the different fusion strategies, UNet++,
Semantic-FPN, LinkNet and HRNet outperform DeepLabV3+. This demonstrates that the cross-layer
fusion works better than the in-module fusion. 3) HRNet outperforms the other methods, due to its
sophisticated architecture, where the features are repeatedly exchanged across different scales. As
is shown in Table 3, multi-scale augmentation (with
scale ={0.5,0.75,1.0,1.25,1.5,1.75}
) was
(a) Semantic-FPN (b) HRNet
Figure 4: Representative confusion matrices for the semantic segmentation experiments.
conducted during the training (MSTr), significantly improving the performance of different methods.
In the implementation, the multi-scale inference adopts multi-scale inputs and ensembles the rescaled
multiple outputs using a simple mean function. With further use in the testing process, all methods
were further improved. As for multi-scale fusion, hierarchical multi-scale architecture search [
may also become an effective solution.
Additional background supervision
. The complex background samples cause serious false alarms
in HRS imagery semantic segmentation [
]. As is shown in Figure 4, the confusion matrices
show that lots of objects were misclassified into background, which is consistent with our analysis
in §3.2. Based on Semantic-FPN, we designed the additional background supervision to address
this problem. Dice loss [
] and binary cross-entropy loss were utilized with the corresponding
modulation factors. We calculated the total loss as:
Ltotal =Lce +αLbce +βLdice
, where
denotes the original cross-entropy loss. Table 4and Table 5additionally report the precision (P),
recall (R) and F1-score (F1) of the background class with varying modulation factors. Besides, the
standard deviations are reported after
runs. Table 4shows that the addition of binary cross-entropy
loss improves the background accuracy and the overall performance. The combination of Ldice and
performs well because they optimize the background class from different directions. In the
future, the spatial attention mechanism [27] may improve the background with adaptive weights.
Table 4: Varied αfor Lbce
αBackground mIoU (%)
P (%) R (%) F1(%)
0 55.46 61.01 59.86 48.15 ±0.17
0.2 57.70 63.36 60.39 48.50 ±0.13
0.5 56.92 65.86 61.06 48.85 ±0.15
0.7 57.73 64.62 61.98 48.74 ±0.19
0.9 57.30 64.05 60.48 48.26 ±0.14
1.0 58.43 62.64 60.46 48.14 ±0.18
Table 5: Varied βfor Ldice (w. optimal α)
β α Background mIoU (%)
P (%) R (%) F1(%)
0.2 0.5 56.68 64.82 60.47 48.97 ±0.16
0.5 0.5 56.88 65.16 60.96 49.23 ±0.09
0.7 0.5 57.13 65.31 60.93 49.68 ±0.14
0.2 0.7 56.91 66.03 61.13 49.69 ±0.17
0.5 0.7 57.14 66.21 61.34 50.08 ±0.15
0.7 0.7 56.68 65.52 60.78 49.48 ±0.13
. Some representative results are shown in Figure 5. With the shallow backbone
(VGG16), FCN8S can hardly recognize the road due to its lack of feature extraction capability.
The other methods which utilize deep layers can produce better results. Because of the disorderly
arrangement and varied scales, the edges of the buildings are hard to extract accurately. Some
small-scale objects such as buildings and scattered trees are easy to miss. In contrast, water class
achieves higher accuracies for all methods. This because water have strong spectral homogeneity
and low intra-class variance [
]. The forest is easy to misclassify into agriculture because these
classes have similar spectra. Because of the high-resolution retention and multi-scale fusion, HRNet
produces the best visualization result, especially in the details.
(a) Image (b) Ground truth (c) FCN8S (d) DeepLabV3+ (e) PSPNet (f) UNet
(g) UNet++ (h) PAN
(i) Semantic-
(j) LinkNet (k) FarSeg (l) HRNet
Figure 5: Semantic segmentation results on images from the LoveDA
set in the Liuhe (
area. Some small-scale objects such as buildings and scattered trees are hard to recognize. The forest
and agricultural classes are easy to misclassify due to their similar spectra.
4.2 Unsupervised Domain Adaptation
The advanced UDA methods were evaluated on the LoveDA dataset. In addition to the original
metric-based approach of DDC [
], two mainstream UDA approaches were tested, i.e., adversarial
training (AdaptSeg [
], CLAN [
], TransNorm [
], FADA [
]) and self-training (CBST [
PyCDA [18], IAST [25]).
Table 6: Unsupervised domain adaptation results obtained on the Test set of the LoveDA dataset.
Domain Method Type IoU (%) mIoU(%)
Background Building Road Water Barren Forest Agriculture
Oracle - 48.18 52.14 56.81 85.72 12.34 36.70 35.66 46.79
Source only - 43.30 25.63 12.70 76.22 12.52 23.34 25.14 31.27
DDC [40] - 43.60 15.37 11.98 79.07 14.13 33.08 23.47 31.53
AdaptSeg [39] AT 42.35 23.73 15.61 81.95 13.62 28.70 22.05 32.68
FADA [44] AT 43.89 12.62 12.76 80.37 12.70 32.76 24.79 31.41
CLAN [22] AT 43.41 25.42 13.75 79.25 13.71 30.44 25.80 33.11
TransNorm [47] AT 38.37 5.04 3.75 80.83 14.19 33.99 17.91 27.73
PyCDA [18] ST 38.04 35.86 45.51 74.87 7.71 40.39 11.39 36.25
CBST [56] ST 48.37 46.10 35.79 80.05 19.18 29.69 30.05 41.32
IAST [25] ST 48.57 31.51 28.73 86.01 20.29 31.77 36.50 40.48
Oracle - 37.18 52.74 43.74 65.89 11.47 45.78 62.91 45.67
Source only - 24.16 37.02 32.56 49.42 14.00 29.34 35.65 31.74
DDC [40] - 25.61 44.27 31.28 44.78 13.74 33.83 25.98 31.36
AdaptSeg [39] AT 26.89 40.53 30.65 50.09 16.97 32.51 28.25 32.27
FADA [44] AT 24.39 32.97 25.61 47.59 15.34 34.35 20.29 28.65
CLAN [22] AT 22.93 44.78 25.99 46.81 10.54 37.21 24.45 30.39
TransNorm [47] AT 19.39 36.30 22.04 36.68 14.00 40.62 3.30 24.62
PyCDA [18] ST 12.36 38.11 20.45 57.16 18.32 36.71 41.90 32.14
CBST [56] ST 25.06 44.02 23.79 50.48 8.33 39.16 49.65 34.36
IAST [25] ST 29.97 49.48 28.29 64.49 2.13 33.36 61.37 38.44
The abbreviations are: AT – adversarial training methods. ST – self-training methods.
Implementation details.
All the UDA methods adopted the same feature extractor and discriminator,
following the common practice [
]. Specifically, DeepLabV2 [
] with ResNet50 was utilized
as the extractor, and the discriminator was constructed by fully convolutional layers [
]. For the
adversarial training (AT), the classification and discriminator learning rates were set to
, respectively. The Adam optimizer was used for the discriminator with the momentum of
. The number of training iterations was set to
, with a batch size of
. The eight
source images and eight target images were alternatively input. The other settings are the same in the
semantic segmentation. and the learning schedule is the same as in semantic segmentation settings.
For the self-training (ST), the classification learning rate was set to
. Full implementation details
are provided in the §A.4.
Benchmark results.
As is shown in Table 6, the Oracle setting obtains the best overall performances.
However, DeepLabV2 has lost its effectiveness due to the domain divergence, referring to the result of
Source only setting. In the
Urban experiments, the accuracies of artificial classes (building
and road) drop more than natural classes (forest and agricultural). Because of the inconsistent class
distribution, the
Rural experiments show the opposite results. The transfer learning methods
relatively improve the model transferability. Noticeably, TransNorm obtains the lowest mIoUs. This
is because the source and target images were obtained by the same sensor, and their spectral statistics
are similar (Figure 3(2)). These rural and urban domains require similar normalization weights, so
that the adaptive normalization can lead to optimization conflicts (more analysis are provided in §A.6).
The ST methods achieve better performances because they address the class imbalance problem with
pseudo-label generation.
Inconsistent class distributions.
It is noticeable to find that the ST methods surpass AT methods in
cross-domain adaptation experiments. We conclude that the main reason for this is the extremely
inconsistent class distribution (Figure 3(a)). The rural scenes only contain a few artificial samples
and large-scale natural objects. In contrast, the urban scenes have a mixture of buildings and roads
with few natural objects. The AT methods cannot address this difficulty, so that they report lower
accuracies. However, differing from the AT methods, the ST methods generate pseudo-labels on
the target images. With the addition of diverse target samples, the class distribution divergence is
eliminated during the training. Overall, the ST methods show more potential in the UDA land-cover
semantic segmentation task. In
Rural experiments, all UDA methods show negative
transfer effects for the road class. Hence, more tailored UDA methods are worth exploring faced with
these special challenges.
The qualitative results for the
Urban experiments are shown in Figure 6.
The Oracle result successfully recognizes the buildings and roads, and is the closest to the ground
truth. According to the Table 2, it can be further improved by using a more robust backbone. The
ST methods (j)–(l) produce better results than AT methods (f)–(i), but there is still much room for
improvement. The large-scale mapping visualizations are provided in §A.7.
(a) Image (b) Ground truth (c) Oracle (d) Source only (e) DDC (f) AdaptSeg
(g) CLAN (h) TransNorm (i) FADA (j) PyCDA (k) CBST (l) IAST
Figure 6: Visual results for the
Urban experiments. (f)–(i) and (j)–(l) were obtained from
the AT and ST methods, respectively. The ST methods produce better results than the AT methods.
Pseudo-label analysis for CBST.
As pseudo samples are important for addressing inconsistent class
distribution problem, we varied the target class proportion in CBST, which is a hyper-parameter
controlling the number of pseudo samples. The mean F1-score (mF1) and mIoU are reported in
Table 7. Without pseudo-label learning (
t= 0
), the model degenerated into Source only setting and
achieved low accuracy. The optimal range of
is relatively large (
0.05 t0.5
), which proves that
it is not sensitive to the remote sensing UDA task.
Table 7: Varied pfor target class proportion (Rural Urban)
t0. 0.01 0.05 0.1 0.5 0.7 0.9 1.0
mF1(%) 46.81 45.24 48.50 50.93 56.30 51.23 51.03 49.43
mIoU(%) 32.94 32.18 34.46 36.84 41.32 37.12 37.02 35.47
5 Conclusion
The differences between urban and rural scenes limit the generalization of deep learning approaches
in land-cover mapping. In order to address this problem, we built an HSR dataset for Land-cOVEr
Domain Adaptive semantic segmentation (LoveDA). The LoveDA dataset reflects three challenges in
large-scale remote sensing mapping, including multi-scale objects, complex background samples,
and inconsistent class distributions. The state-of-the-art methods were evaluated on the LoveDA
dataset, revealing the challenges of LoveDA. In addition, multi-scale architectures and strategies,
additional background supervision and pseudo-label analysis were conducted to find alternative ways
to address these challenges.
6 Broader Impact
This work offers a free and open dataset with the purpose of advancing land-cover semantic segmen-
tation in the area of remote sensing. We also provide two benchmarked tasks with three considerable
challenges. This will allow other researchers to easily build on this work and create new and enhanced
capabilities. The authors do not foresee any negative societal impacts of this work. A potential
positive societal impact may arise from the development of generalizable models that can produce
large-scale high-spatial-resolution land-cover mapping accurately. This could help to reduce the
manpower and material resource consumption of surveying and mapping.
7 Acknowledgments
This work was supported by National Key Research and Development Program of China under Grant
No. 2017YFB0504202, National Natural Science Foundation of China under Grant Nos. 41771385,
41801267, and the China Postdoctoral Science Foundation under Grant 2017M622522. This work
was supported by the Nanjing Bureau of Surveying and Mapping.
H. Alemohammad and K. Booth. LandCoverNet: A global benchmark land cover classification training
dataset. arXiv preprint arXiv:2012.03111, 2020.
A. Boguszewski, D. Batorski, N. Ziemba-Jankowska, T. Dziedzic, and A. Zambrzycka.
Dataset for automatic mapping of buildings, woodlands, water and roads from aerial imagery. In Pro-
ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1102–1110,
A. Chaurasia and E. Culurciello. Linknet: Exploiting encoder representations for efficient semantic
segmentation. In 2017 IEEE Visual Communications and Image Processing (VCIP), pages 1–4. IEEE,
L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille. Deeplab: Semantic image
segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions
on pattern analysis and machine intelligence, 40(4):834–848, 2017.
L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam. Encoder-decoder with atrous separable
convolution for semantic image segmentation. In Proceedings of the European conference on computer
vision (ECCV), pages 801–818, 2018.
Q. Chen, L. Wang, Y. Wu, G. Wu, Z. Guo, and S. L. Waslander. Aerial imagery for roof segmentation:
A large-scale dataset towards automatic mapping of buildings. ISPRS Journal of Photogrammetry and
Remote Sensing, 147:42–55, 2019.
W. Chen, Z. Jiang, Z. Wang, K. Cui, and X. Qian. Collaborative global-local networks for memory-efficient
segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 8924–8933, 2019.
U. S. Commission et al. A recommendation on the method to delineate cities, urban and rural areas for
international statistical comparisons. European Commission, 2020.
I. Demir, K. Koperski, D. Lindenbaum, G. Pang, J. Huang, S. Basu, F. Hughes, D. Tuia, and R. Raskar.
Deepglobe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition Workshops, pages 172–181, 2018.
Y. Dong, T. Liang, Y. Zhang, and B. Du. Spectral–spatial weighted kernel manifold embedded distribution
alignment for remote sensing image classification. IEEE Transactions on Cybernetics, 51(6):3185–3197,
Y. Duan, H. Huang, Z. Li, and Y. Tang. Local manifold-based sparse discriminant learning for feature
extraction of hyperspectral image. IEEE transactions on cybernetics, 2020.
M. Everingham, S. A. Eslami, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual
object classes challenge: A retrospective. International journal of computer vision, 111(1):98–136, 2015.
J. Iqbal and M. Ali. Weakly-supervised domain adaptation for built-up region segmentation in aerial and
satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 167:263–275, 2020.
S. Jin, C. Homer, J. Dewitz, P. Danielson, and D. Howard. National land cover database (nlcd) 2016
science research products. In AGU Fall Meeting Abstracts, volume 2019, pages B11I–2301, 2019.
[15] C. Jun, Y. Ban, and S. Li. Open access to earth land-cover map. Nature, 514(7523):434–434, 2014.
A. Kirillov, R. Girshick, K. He, and P. Dollár. Panoptic feature pyramid networks. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399–6408, 2019.
H. Li, P. Xiong, J. An, and L. Wang. Pyramid attention network for semantic segmentation. arXiv preprint
arXiv:1805.10180, 2018.
Q. Lian, F. Lv, L. Duan, and B. Gong. Constructing self-motivated pyramid curriculums for cross-domain
semantic segmentation: A non-adversarial approach. In Proceedings of the IEEE/CVF International
Conference on Computer Vision, pages 6758–6767, 2019.
J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In
Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431–3440, 2015.
M. Long, Y. Cao, J. Wang, and M. Jordan. Learning transferable features with deep adaptation networks.
In International conference on machine learning, pages 97–105. PMLR, 2015.
X. Lu, T. Gong, and X. Zheng. Multisource compensation network for remote sensing cross-domain scene
classification. IEEE Transactions on Geoscience and Remote Sensing, 58(4):2504–2515, 2019.
Y. Luo, L. Zheng, T. Guan, J. Yu, and Y. Yang. Taking a closer look at domain shift: Category-level
adversaries for semantics consistent domain adaptation. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pages 2507–2516, 2019.
A. Ma, J. Wang, Y. Zhong, and Z. Zheng. Factseg: Foreground activation-driven small object semantic
segmentation in large-scale remote sensing imagery. IEEE Transactions on Geoscience and Remote
Sensing, 2021.
D. Marcos, M. Volpi, B. Kellenberger, and D. Tuia. Land cover mapping at very high resolution with
rotation equivariant cnns: Towards small yet accurate models. ISPRS journal of photogrammetry and
remote sensing, 145:96–107, 2018.
K. Mei, C. Zhu, J. Zou, and S. Zhang. Instance adaptive self-training for unsupervised domain adaptation. In
Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings,
Part XXVI 16, pages 415–430. Springer, 2020.
F. Milletari, N. Navab, and S.-A. Ahmadi. V-net: Fully convolutional neural networks for volumetric
medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565–571.
IEEE, 2016.
L. Mou, Y. Hua, and X. X. Zhu. A relation-augmented fully convolutional network for semantic segmen-
tation in aerial scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 12416–12425, 2019.
E. Othman, Y. Bazi, F. Melgani, H. Alhichri, N. Alajlan, and M. Zuair. Domain adaptation network for
cross-scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(8):4441–4456, 2017.
J. Pang, C. Li, J. Shi, Z. Xu, and H. Feng.
-CNN: Fast tiny object detection in large-scale remote
sensing images. IEEE Transactions on Geoscience and Remote Sensing, 57(8):5512–5524, 2019.
X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang. Moment matching for multi-source
domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages
1406–1415, 2019.
X. Peng, B. Usman, N. Kaushik, D. Wang, J. Hoffman, and K. Saenko. Visda: A synthetic-to-real
benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition Workshops, pages 2021–2026, 2018.
O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation.
In International Conference on Medical image computing and computer-assisted intervention, pages 234–
241. Springer, 2015.
D. Sulla-Menashe and M. A. Friedl. User guide to collection 6 modis land cover (mcd12q1 and mcd12c1)
product. USGS: Reston, VA, USA, pages 1–18, 2018.
B. Sun and K. Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European
conference on computer vision, pages 443–450. Springer, 2016.
O. Tasar, Y. Tarabalka, A. Giros, P. Alliez, and S. Clerc. Standardgan: Multi-source domain adaptation for
semantic segmentation of very high resolution satellite images by data standardization. In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 192–193, 2020.
C. Tian, C. Li, and J. Shi. Dense fusion classmate network for land cover classification. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 192–196, 2018.
W. R. Tobler. A computer movie simulating urban growth in the Detroit region. Economic geography,
46(sup1):234–240, 1970.
X.-Y. Tong, G.-S. Xia, Q. Lu, H. Shen, S. Li, S. You, and L. Zhang. Land-cover classification with
high-resolution remote sensing images using transferable deep models. Remote Sensing of Environment,
237:111322, 2020.
Y.-H. Tsai, W.-C. Hung, S. Schulter, K. Sohn, M.-H. Yang, and M. Chandraker. Learning to adapt structured
output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 7472–7481, 2018.
E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for
domain invariance. arXiv preprint arXiv:1412.3474, 2014.
A. Van Etten, D. Lindenbaum, and T. M. Bacastow. Spacenet: A remote sensing dataset and challenge
series. arXiv preprint arXiv:1807.01232, 2018.
H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan. Deep hashing network for unsupervised
domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pages 5018–5027, 2017.
M. Volpi and V. Ferrari. Semantic segmentation of urban scenes by learning local class interactions. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1–9,
H. Wang, T. Shen, W. Zhang, L.-Y. Duan, and T. Mei. Classes matter: A fine-grained adversarial approach
to cross-domain semantic segmentation. In European Conference on Computer Vision, pages 642–659.
Springer, 2020.
J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y. Zhao, D. Liu, Y. Mu, M. Tan, X. Wang, et al. Deep
high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and
machine intelligence, 2020.
J. Wang, Y. Zhong, Z. Zheng, A. Ma, and L. Zhang. RSNet: the search for remote sensing deep neural
networks in recognition tasks. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2520–2534,
X. Wang, Y. Jin, M. Long, J. Wang, and M. I. Jordan. Transferable normalization: Towards improving
transferability of deep neural networks. In Advances in Neural Information Processing Systems, pages
1953–1963, 2019.
S. Waqas Zamir, A. Arora, A. Gupta, S. Khan, G. Sun, F. Shahbaz Khan, F. Zhu, L. Shao, G.-S. Xia, and
X. Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28–37, 2019.
Y. Xiao, X. Su, Q. Yuan, D. Liu, H. Shen, and L. Zhang. Satellite video super-resolution via multiscale
deformable convolution alignment and temporal grouping projection. IEEE Transactions on Geoscience
and Remote Sensing, pages 1–19, 2021.
L. Yan, B. Fan, H. Liu, C. Huo, S. Xiang, and C. Pan. Triplet adversarial domain adaptation for pixel-level
classification of vhr remote sensing images. IEEE Transactions on Geoscience and Remote Sensing,
58(5):3558–3573, 2019.
Q. ZHANG, J. Zhang, W. Liu, and D. Tao. Category anchor-guided unsupervised domain adaptation for
semantic segmentation. Advances in Neural Information Processing Systems, 32:435–445, 2019.
H. Zhao. National urban population and construction land in 2016 (by cities). China Statistics Press, 2016.
H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE
conference on computer vision and pattern recognition, pages 2881–2890, 2017.
Z. Zheng, Y. Zhong, J. Wang, and A. Ma. Foreground-aware relation network for geospatial object seg-
mentation in high spatial resolution remote sensing imagery. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition, pages 4096–4105, 2020.
Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical
image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical
decision support, pages 3–11. Springer, 2018.
Y. Zou, Z. Yu, B. Kumar, and J. Wang. Unsupervised domain adaptation for semantic segmentation via
class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages
289–305, 2018.
Y. Zou, Z. Yu, X. Liu, B. Kumar, and J. Wang. Confidence regularized self-training. In Proceedings of the
IEEE/CVF International Conference on Computer Vision, pages 5982–5991, 2019.
A Appendix
A.1 Annotation Procedure and Data Division
The seven common land-cover types were developed according to the “Data Regulations and Collec-
tion Requirements for the General Survey of Geographical Conditions”, i.e., buildings, road, water,
forest, agriculture, and background classes. Based on the advanced ArcGIS geo-spatial software ,
all the images were annotated by professional remote sensing annotators. With the division of these
images, a comprehensive annotation pipeline was adopted referring to [
]. The annotators labeled all
objects belonging to six categories (except background) using polygon features. As for the 18 selected
areas, it took approximately 24.6 h to finish the single-area annotations, resulting in a time cost of
442.8 man hours in total. After the first round of labeling, self-examination and cross-examination
were conducted, correcting the false labels, missing objects, and inaccurate boundaries. The team
supervisors then randomly sampled 600 images for quality inspection. The unqualified annotations
were then refined by the annotators. Finally, several statistics (e.g. object numbers per image,
object areas, etc.) were computed to double check the outliers. Based on DeepLabV3, preliminary
experiments were conducted to ensure the validity of the annotations.
Table 8: The division of the LoveDA dataset
Domain City Region #Images Train Val Test
Qixia 320 X
Gulou 320 X
Qinhuai 336 X
Yuhuatai 357 X
Jianye 357 X
Changzhou Jintan 320 X
Wujin 320 X
Wuhan Jianghan 180 X
Wuchang 143 X
Pukou 320 X
Gaochun 336 X
Lishui 336 X
Liuhe 320 X
Jiangning 336 X
Changzhou Liyang 320 X
Xinbei 320 X
Wuhan Jiangxia 374 X
Huangpi 672 X
Total 5987 2522 1669 1796
A.2 Top Performances Compared with Other Datasets
In order to support the "challeangability" of the proposed dataset compared to other land-cover
datasets. By investigating the current researches, the top performances on different datasets have
been reported in Table 9. The advanced method (HRNet) only achieved the lowest performance on
the LoveDA dataset, showing the difficulty of this dataset
Table 9: Top performances compared with other datasets
Dataset Top mIoU (%)
GID [46] 93.54
DeepGlobe [36] 52.24
ISPRS Potsdam [27] 82.38
ISPRS Vaihingen [27] 79.76
LoveDA 49.79
A.3 Instance Differences Between Urban and Rural Areas
For the LoveDA dataset, the differences between urban and rural areas at the instance level are shown
in the Figure 7. Similar with the pixel analysis in §3.3, the instances across domains are imbalanced.
Specifically, the urban areas have more buildings and fewer instances of agricultural land. The rural
areas have more instances of agricultural land. This also highlights the inconsistent class distribution
problem between different domains.
Figure 7: Instance differences between urban and rural areas.
A.4 Implementation Details
All the networks were implemented under the PyTorch framework, using an NVIDIA 24 GB RTX
TITAN GPU. The backbones used in all the networks were pre-trained on ImageNet. The number of
training iterations was set to
with a batch size of
. The eight source images and eight target
images were alternately input. The other settings were the same as in the semantic segmentation. As
for self-training (ST), the pseudo-generation hyper-parameters remained the same as in the original
literature. The classification learning rate was set to
. All the ST-based networks were trained
steps including two stages: 1) for the first
steps, the models were trained only on the
source images for initialization; and 2) the pseudo-labels were then updated every
steps during
the remaining training process. Considering the training stability, IAST method was set
steps for
initialization in the Urban Rural experiments.
All the networks were then re-implemented following the original literature. The segmentation
models followed the default settings in [
], including a modified ResNet50 and atrous spatial
pyramid pooling (ASPP)[
]. By using dilated convolutions, the stride of the last two convolution
layers was modified from 2to 1. The final output stride of the feature map was 16.
Following [
], the discriminator was made up of five convolutional layers with a kernel of
a stride of
, where the channel numbers were
, respectively. Each convolution
was followed with a Leaky ReLU, and the parameter was set to
. Bilinear interpolation was used
for re-scaling the output to the size of the input.
As for the hyperparameter settings, the adversarial scale factor
was set to
following [
With respect to the two segmentation outputs in [
were set to
respectively. The weight discrepancy loss was used in CLAN[
], and the default settings were
adopted, i.e.,
λw= 0.01
λlocal = 10
, and
= 0.4
. FADA [
] adopts the temperature
encourage a soft probability distribution over the classes, which was set to
by default. The
confidence of pseudo-label
in PyCDA[
] was set to
by default. The pseudo-label related
hyperparameters for IAST remained the same as in [
]. The target proportion
in CBST was set to
0.1and 0.5when transferring to the rural and urban domains, respectively.
A.5 Error Bar Visualization for the UDA Experiments
In order to make the results more convincing and reproducible, we ran all UDA methods five times
using a random seed. The error bar visualization for the UDA experiments is shown in Figure 8. The
adversarial training methods achieve smaller error fluctuations than the self-training methods. This
is because the self-training methods assign and update the pseudo-labels alternately, which brings
greater randomness. Hence, for the self-training methods, we suggest that three times more repeats
are preferred to provide more convincing results.
Figure 8: Error bar visualization for the UDA experiments.
A.6 Batch Normalization Statistics in the Different Domains
The batch normalization (BN) statistics are shown in Figure 9. We observe that in the Oracle source
and target settings, the model has similar BN statistics in both mean and variance. This demonstrates
that the gap between the source and target domains does not lie in the BNs, which is different from
the conclusion in [
]. Hence, the modification of the BN statistics may have a negative effect, as in
], where the target BN statistics are far different from those of the Oracle target model.
This observation is consistent with the results listed in Table 6. We speculate that the cause of this
failure in the combined simulation dataset UDA experiments[
] is that the source and target
domains have large spectral differences, and thus require domain-specific BN statistics. However, the
LoveDA dataset is real data obtained from the same sensor at the same time. The spectral difference
in the source and target domains is very small (Figure 3(b)), so the BN statistics are very similar
(Figure 9).
(a) Layer1’s RM (b) Layer2’s RM (c) Layer3’s RM (d) Layer4’s RM
(e) Layer1’s RV (f) Layer2’s RV (g) Layer3’s RV (h) Layer4’s RV
Figure 9: Statistics of the running mean (RM) and running var (RV) of the batch normalization in the
different layers of ResNet50. Two Oracle models and TransNorm in the
Rural experiments
are shown.
A.7 Large-scale Visualizations on UDA Test Set
The large-scale visualizations are shown in the Figure 10. Compared with the baseline, CBST can
produce better results on large-scale mapping, which highlights the importance of developing UDA
methods. However, CBST still has a lot of room for improvement. More tailored UDA algorithms
requires to be developed on the LoveDA dataset.
(a) Baseline on Wujin area
(b) CBST on Wujin area
Figure 10: Large-scale visualizations on UDA Test set (Rural Urban).
... Considering the similar resolution, data source, and large image size, we choose the LoveDA [49] and DeepGlobe dataset [11] to implement the cross-dataset generalization evaluation. Following the data split of [49,18], we use DeepLab v3+-Res-50 as the segmentation model to train the models and evaluate the cross-dataset performance. ...
... Considering the similar resolution, data source, and large image size, we choose the LoveDA [49] and DeepGlobe dataset [11] to implement the cross-dataset generalization evaluation. Following the data split of [49,18], we use DeepLab v3+-Res-50 as the segmentation model to train the models and evaluate the cross-dataset performance. ...
... In addition, our PCL can detect small and slim rivers, as shown in (n) of Figure 7. Qualitative results about cross-dataset generalization evaluation. Figures 8 and 9 display the visualization results on the DeepGlobe [10] and LoveDA [49] test sets of models trained on different datasets, respectively. As shown in the quantitative results presented in our paper (IoU: 74.43% vs. 79.1% and 68.15% vs. 69.27%), ...
Full-text available
Global surface water detection in very-high-resolution (VHR) satellite imagery can directly serve major applications such as refined flood mapping and water resource assessment. Although achievements have been made in detecting surface water in small-size satellite images corresponding to local geographic scales, datasets and methods suitable for mapping and analyzing global surface water have yet to be explored. To encourage the development of this task and facilitate the implementation of relevant applications, we propose the GLH-water dataset that consists of 250 satellite images and manually labeled surface water annotations that are distributed globally and contain water bodies exhibiting a wide variety of types (e.g., rivers, lakes, and ponds in forests, irrigated fields, bare areas, and urban areas). Each image is of the size 12,800 $\times$ 12,800 pixels at 0.3 meter spatial resolution. To build a benchmark for GLH-water, we perform extensive experiments employing representative surface water detection models, popular semantic segmentation models, and ultra-high resolution segmentation models. Furthermore, we also design a strong baseline with the novel pyramid consistency loss (PCL) to initially explore this challenge. Finally, we implement the cross-dataset and pilot area generalization experiments, and the superior performance illustrates the strong generalization and practical application of GLH-water. The dataset is available at
... In the field of aerial semantic segmentation there have been numerous advances based on the use of deep learning models [2], [3], [6], [10], [11] trained on large public datasets with annotated images [9], [12], which have led to remarkable levels of performance. However, this performance generally does not carry over when these models are set to operate on images that come from a distribution (target domain) different from the data experienced during training (source domain). ...
... We evaluate HIUDA on the LoveDA benchmark [12], the only dataset designed for evaluating unsupervised domain adaptation in aerial segmentation, where we exceed the current state-of-the-art. We further provide a comprehensive ablation study to assess the impact of the proposed solutions. ...
... where y c i represents the ground truth annotation for the pixel i and class c. While alternative functions or a combination of them, such as cross-entropy and Dice loss [12], could be adopted in the aerial domain, this work concentrates on the UDA task. In this context, the objective function is not the main focus, and the cross-entropy provides a fair comparison with other approaches [16]. ...
Full-text available
We investigate the task of unsupervised domain adaptation in aerial semantic segmentation observing that there are some shortcomings in the class mixing strategies used by the recent state-of-the-art methods that tackle this task: (i) they do not account for the large disparity in the extension of the semantic categories that is common in the aerial setting, which causes a domain imbalance in the mixed image; (ii) they do not consider that aerial scenes have a weaker structural consistency in comparison to the driving scenes for which the mixing technique was originally proposed, which causes the mixed images to have elements placed out of their natural context; (iii) the source model used to generate the pseudo-labels may be susceptible to perturbations across domains, which causes inconsistent predictions on the target images and can jeopardize the mixing strategy. We address these shortcomings with a novel aerial semantic segmentation framework for UDA, named HIUDA, which is composed of two main technical novelties: (i) a new mixing strategy for aerial segmentation across domains called Hierarchical Instance Mixing (HIMix), which extracts a set of connected components from each semantic mask and mixes them according to a semantic hierarchy and, (ii) a twin-head architecture in which two separate segmentation heads are fed with variations of the same images in a contrastive fashion to produce finer segmentation maps. We conduct extensive experiments on the LoveDA benchmark, where our solution outperforms the current state-of-the-art.
... To verify the performance of ST-DASegNet, we conduct extensive experiments on benchmark datasets (Potsdam/Vaihingen [22] and LoveDA [23] ...
... Wu et al. [57] also focus on extracting domain-invariant features. Instructed by this principle, they propose a deep covariance alignment (DCA) module, which achieves competitive results on famous benchmark dataset, LoveDA [23]. Li et al. [58] propose a step-wise RS segmentation networks with covariate shift alleviation close the gap between source and target domains. ...
... On LoveDA dataset, we design two cross-domain RS semantic segmentation tasks, which are urban-to-rural (urban → rural) and rural-to-urban (rural → urban) task. All our experiments setting are followed [23], [57]. Evaluation Metric. ...
Deep convolutional neural networks (DCNNs) based remote sensing (RS) image semantic segmentation technology has achieved great success used in many real-world applications such as geographic element analysis. However, strong dependency on annotated data of specific scene makes it hard for DCNNs to fit different RS scenes. To solve this problem, recent works gradually focus on cross-domain RS image semantic segmentation task. In this task, different ground sampling distance, remote sensing sensor variation and different geographical landscapes are three main factors causing dramatic domain shift between source and target images. To decrease the negative influence of domain shift, we propose a self-training guided disentangled adaptation network (ST-DASegNet). We first propose source student backbone and target student backbone to respectively extract the source-style and target-style feature for both source and target images. Towards the intermediate output feature maps of each backbone, we adopt adversarial learning for alignment. Then, we propose a domain disentangled module to extract the universal feature and purify the distinct feature of source-style and target-style features. Finally, these two features are fused and served as input of source student decoder and target student decoder to generate final predictions. Based on our proposed domain disentangled module, we further propose exponential moving average (EMA) based cross-domain separated self-training mechanism to ease the instability and disadvantageous effect during adversarial optimization. Extensive experiments and analysis on benchmark RS datasets show that ST-DASegNet outperforms previous methods on cross-domain RS image semantic segmentation task and achieves state-of-the-art (SOTA) results. Our code is available at
... We first introduce the datasets, evaluation metrics, and implementation details and then conduct ablation studies to validate the effectiveness of our framework. Finally, we compare the proposed network with several state-of-the-art methods on ISPRS Vaihingen, Potsdam [32], and LoveDA Urban [33] datasets. ...
... LoveDA Urban dataset: The LoveDA dataset is constructed by Wang et al. [33]. The historical images were obtained from the Google Earth platform. ...
Full-text available
Semantic segmentation of high-resolution remote sensing images (HRSI) is significant, yet challenging. Recently, several research works have utilized the self-attention operation to capture global dependencies. HRSI have complex scenes and rich details, and the implementation of self-attention on a whole image will introduce redundant information and interfere with semantic segmentation. The detail recovery of HRSI is another challenging aspect of semantic segmentation. Several networks use up-sampling, skip-connections, parallel structure, and enhanced edge features to obtain more precise results. However, the above methods ignore the misalignment of features with different resolutions, which affects the accuracy of the segmentation results. To resolve these problems, this paper proposes a semantic segmentation network based on sparse self-attention and feature alignment (SAANet). Specifically, the sparse position self-attention module (SPAM) divides, rearranges, and resorts the feature maps in the position dimension and performs position attention operations (PAM) in rearranged and restored sub-regions, respectively. Meanwhile, the proposed sparse channel self-attention module (SCAM) groups, rearranges, and resorts the feature maps in the channel dimension and performs channel attention operations (CAM) in the rearranged and restored sub-channels, respectively. SPAM and SCAM effectively model long-range context information and interdependencies between channels, while reducing the introduction of redundant information. Finally, the feature alignment module (FAM) utilizes convolutions to obtain a learnable offset map and aligns feature maps with different resolutions, helping to recover details and refine feature representations. Extensive experiments conducted on the ISPRS Vaihingen, Potsdam, and LoveDA datasets demonstrate that the proposed method precedes general semantic segmentation- and self-attention-based networks.
... Producing a VHR national-scale LC map of China can help us to investigate the environment, development, and future trend of the country in detail. With the updating of remote sensing platforms, the available LC products have been through the trends of coarse to fine [1]. Nevertheless, due to the low orbit and small visual field of the VHR image captured platforms, the corresponding LC products generally have a smaller coverage when they have a higher spatial resolution. ...
Nowadays, many large-scale land-cover (LC) products have been released, however, current LC products for China either lack a fine resolution or nationwide coverage. With the rapid urbanization of China, there is an urgent need for creating a very-high-resolution (VHR) national-scale LC map for China. In this study, a novel 1-m resolution LC map of China covering $9,600,000 km^2$, called SinoLC-1, was produced by using a deep learning framework and multi-source open-access data. To efficiently generate the VHR national-scale LC map, firstly, the reliable LC labels were collected from three 10-m LC products and Open Street Map data. Secondly, the collected 10-m labels and 1-m Google Earth imagery were utilized in the proposed low-to-high (L2H) framework for training. With weak and self-supervised strategies, the L2H framework resolves the label noise brought by the mismatched resolution between training pairs and produces VHR results. Lastly, we compare the SinoLC-1 with five widely used products and validate it with a sample set including 10,6852 points and a statistical report collected from the government. The results show the SinoLC-1 achieved an OA of 74\% and a Kappa of 0.65. Moreover, as the first 1-m national-scale LC map for China, the SinoLC-1 shows overall acceptable results with the finest landscape details.
Full-text available
This paper presents a fine-grained and multi-sourced dataset for environmental determinants of health collected from England cities. We provide health outcomes of citizens covering physical health (COVID-19 cases, asthma medication expenditure, etc.), mental health (psychological medication expenditure), and life expectancy estimations. We present the corresponding environmental determinants from four perspectives, including basic statistics (population, area, etc.), behavioural environment (availability of tobacco, health-care services, etc.), built environment (road density, street view features, etc.), and natural environment (air quality, temperature, etc.). To reveal regional differences, we extract and integrate massive environment and health indicators from heterogeneous sources into two unified spatial scales, i.e., at the middle layer super output area (MSOA) and the city level, via big data processing and deep learning. Our data holds great promise for diverse audiences, such as public health researchers and urban designers, to further unveil the environmental determinants of health and design methodology for a healthy, sustainable city.
Semantic segmentation is an extremely challenging task in high-resolution remote sensing (HRRS) images as objects have complex spatial layouts and enormous variations in appearance. Convolutional neural networks (CNNs) have excellent ability to extract local features and have been widely applied as the feature extractor for various vision tasks. However, due to the inherent inductive bias of convolution operation, CNNs inevitably have limitations in modeling long-range dependencies. Transformer can capture global representations well, but unfortunately ignores the details of local features and has high computational and spatial complexity in processing high-resolution feature maps. In this paper, we propose a novel hybrid architecture for HRRS image segmentation, termed EMRT, to exploit the advantages of convolution operations and Transformer to enhance multi-scale representation learning. We incorporate the deformable self-attention mechanism in the Transformer to automatically adjust the receptive field, and design an encoder-decoder architecture accordingly to achieve efficient context modeling. Specifically, the CNN is constructed to extract feature representations. In the encoder, local features and global representations at different resolutions are extracted by the CNN and Transformer, respectively, and fused in an interactive manner. Moreover, a separate spatial branch is designed to extract multi-scale contextual information as queries, and global dependencies between features at different scales are efficiently established by the decoder. Extensive experiments on three public remote sensing datasets demonstrate the superiority of EMRT and indicate that the overall performance of our method outperforms state-of-the-art methods. Code is available at
High spatial resolution (HSR) remote sensing images contain complex foreground-background relationships, which makes the remote sensing land cover segmentation a special semantic segmentation task. The main challenges come from the large-scale variation, complex background samples and imbalanced foreground-background distribution. These issues make recent context modeling methods sub-optimal due to the lack of foreground saliency modeling. To handle these problems, we propose a Remote Sensing Segmentation framework (RSSFormer), including Adaptive TransFormer Fusion Module, Detail-aware Attention Layer and Foreground Saliency Guided Loss. Specifically, from the perspective of relation-based foreground saliency modeling, our Adaptive Transformer Fusion Module can adaptively suppress background noise and enhance object saliency when fusing multi-scale features. Then our Detail-aware Attention Layer extracts the detail and foreground-related information via the interplay of spatial attention and channel attention, which further enhances the foreground saliency. From the perspective of optimization-based foreground saliency modeling, our Foreground Saliency Guided Loss can guide the network to focus on hard samples with low foreground saliency responses to achieve balanced optimization. Experimental results on LoveDA datasets, Vaihingen datasets, Potsdam datasets and iSAID datasets validate that our method outperforms existing general semantic segmentation methods and remote sensing segmentation methods, and achieves a good compromise between computational overhead and accuracy. Our code is available at .
Full-text available
As a new earth observation tool, satellite video has been widely used in remote-sensing field for dynamic analysis. Video super-resolution (VSR) technique has thus attracted increasing attention due to its improvement to spatial resolution of satellite video. However, the difficulty of remote-sensing image alignment and the low efficiency of spatial-temporal information fusion make poor generalization of the conventional VSR methods applied to satellite videos. In this article, a novel fusion strategy of temporal grouping projection and an accurate alignment module are proposed for satellite VSR. First, we propose a deformable convolution alignment module with a multiscale residual block to alleviate the alignment difficulties caused by scarce motion and various scales of moving objects in remote-sensing images. Second, a temporal grouping projection fusion strategy is proposed, which can reduce the complexity of projection and make the spatial features of reference frames play a continuous guiding role in spatial-temporal information fusion. Finally, a temporal attention module is designed to adap-tively learn the different contributions of temporal information extracted from each group. Extensive experiments on Jilin-1 satellite video demonstrate that our method is superior to current state-of-the-art VSR methods.
Full-text available
The small object semantic segmentation task is aimed at automatically extracting key objects from high-resolution remote sensing (HRS) imagery. Compared with the large-scale coverage areas for remote sensing imagery, the key objects such as cars, ships, etc. in HRS imagery often contain only a few pixels. In this paper, to tackle this problem, the foreground activation (FA) driven small object semantic segmentation (FactSeg) framework is proposed from perspectives of structure and optimization. In the structure design, FA object representation is proposed to enhance the awareness of the weak features in small objects. The FA object representation framework is made up of a dual-branch decoder and collaborative probability (CP) loss. In the dual-branch decoder, the FA branch is designed to activate the small object features (activation), as well as suppress the largescale background, and the semantic refinement (SR) branch is designed to further distinguish small objects (refinement). The CP loss is proposed to effectively combine the activation and refinement outputs of the decoder under the CP hypothesis. During the collaboration, the weak features of the small objects are enhanced with the activation output, and the refined output can be viewed as the refinement of the binary outputs. In the optimization stage, small object mining (SOM) based network optimization is applied to automatically select effective samples, to refine the direction of the optimization, while addressing the imbalanced sample problem between the small objects and the large-scale background. The experimental results obtained with two benchmark HRS imagery segmentation datasets demonstrate that the proposed framework outperforms the state-of-the-art semantic segmentation methods, and achieves a good tradeoff between accuracy and efficiency.
Conference Paper
Full-text available
The divergence between labeled training data and unlabeled testing data is a significant challenge for recent deep learning models. Unsupervised domain adaptation (UDA) attempts to solve such a problem. Recent works show that self-training is a powerful approach to UDA. However, existing methods have difficulty in balancing scalability and performance. In this paper, we propose an instance adaptive self-training framework for UDA on the task of semantic segmentation. To effectively improve the quality of pseudo-labels, we develop a novel pseudo-label generation strategy with an instance adaptive selector. Besides, we pro- pose the region-guided regularization to smooth the pseudo-label region and sharpen the non-pseudo-label region. Our method is so concise and efficient that it is easy to be generalized to other unsupervised domain adaptation methods. Experiments on ‘GTA5 to Cityscapes’ and ‘SYN-THIA to Cityscapes’ demonstrate the superior performance of our approach compared with the state-of-the-art methods. Codes:
Full-text available
Deep learning algorithms, especially convolutional neural networks (CNNs), have recently emerged as a dominant paradigm for high spatial resolution remote sensing (HRS) image recognition. A large amount of CNNs have already been successfully applied to various HRS recognition tasks, such as land-cover classification, scene classification, etc. However, they are often modifications of the existing CNNs derived from natural image processing, in which the network architecture is inherited without consideration of the complexity and specificity of HRS images. In this paper, the remote sensing deep neural network (RSNet) framework is proposed using an automatically search strategy to find the appropriate network architecture for HRS image recognition tasks. In RSNet, the hierarchical search space is first designed to include module-level and transition-level spaces. The module-level space defines the basic structure block, where a series of lightweight operations as candidates, including depthwise separable convolutions, are proposed to ensure the efficiency. The transition-level space controls the spatial resolution transformations of the features. In the hierarchical search space, a gradient-based search strategy is used to find the appropriate architecture. In RSNet, the task-driven architecture training process can acquire the optimal model parameters of the switchable recognition module for HRS image recognition tasks. The experimental results obtained using four benchmark datasets for land-cover classification and scene classification tasks demonstrate that the searched RSNet can achieve a satisfactory accuracy with a high computational efficiency, and hence provides an effective option for the processing of HRS imagery.
Recently, the convolutional neural network has brought impressive improvements for object detection. However, detecting tiny objects in large-scale remote sensing images still remains challenging. First, the extreme large input size makes the existing object detection solutions too slow for practical use. Second, the massive and complex backgrounds cause serious false alarms. Moreover, the ultratiny objects increase the difficulty of accurate detection. To tackle these problems, we propose a unified and self-reinforced network called remote sensing region-based convolutional neural network (R <sup xmlns:mml="" xmlns:xlink="">2</sup> -CNN), composing of backbone Tiny-Net, intermediate global attention block, and final classifier and detector. Tiny-Net is a lightweight residual structure, which enables fast and powerful features extraction from inputs. Global attention block is built upon Tiny-Net to inhibit false positives. Classifier is then used to predict the existence of target in each patch, and detector is followed to locate them accurately if available. The classifier and detector are mutually reinforced with end-to-end training, which further speed up the process and avoid false alarms. Effectiveness of R <sup xmlns:mml="" xmlns:xlink="">2</sup> -CNN is validated on hundreds of GF-1 images and GF-2 images that are 18000 × 18 192 pixels, 2.0-m resolution, and 27 620 × 29 200 pixels, 0.8-m resolution, respectively. Specifically, we can process a GF-1 image in 29.4 s on Titian X just with single thread. According to our knowledge, no previous solution can detect the tiny object on such huge remote sensing images gracefully. We believe that it is a significant step toward practical real-time remote sensing systems.
This paper proposes a novel domain adaptation algorithm to handle the challenges posed by the satellite and aerial imagery, and demonstrates its effectiveness on the built-up region segmentation problem. Built-up area estimation is an important component in understanding the human impact on the environment, effect of public policy and in general urban population analysis. The diverse nature of aerial and satellite imagery (capturing different geographical locations, terrains and weather conditions) and lack of labeled data covering this diversity makes machine learning algorithms difficult to generalize for such tasks, especially across multiple domains. Re-training for new domain is both computationally and labor expansive mainly due to the cost of collecting pixel level labels required for the segmentation task. Domain adaptation algorithms have been proposed to enable algorithms trained on images of one domain (source) to work on images from other dataset (target). Unsupervised domain adaptation is a popular choice since it allows the trained model to adapt without requiring any ground-truth information of the target domain. On the other hand, due to the lack of strong spatial context and structure, in comparison to the ground imagery, application of existing unsupervised domain adaptation methods results in the sub-optimal adaptation. We thoroughly study limitations of existing domain adaptation methods and propose a weakly-supervised adaptation strategy where we assume image level labels are available for the target domain. More specifically, we design a built-up area segmentation network (as encoder-decoder), with image classification head added to guide the adaptation. The devised system is able to address the problem of visual differences in multiple satellite and aerial imagery datasets, ranging from high resolution (HR) to very high resolution (VHR), by investigating the latent space as well as the structured output space. A realistic and challenging HR dataset is created by hand-tagging the 73.4 sq-km of Rwanda, capturing a variety of build-up structures over different terrain. The developed dataset is spatially rich compared to existing datasets and covers diverse built-up scenarios including built-up areas in forests and deserts, mud houses, tin and colored rooftops. Extensive experiments are performed by adapting from the single-source domain datasets, such as Massachusetts Buildings Dataset, to segment out the target domain. We achieve high gains ranging 11.6–52% in IoU over the existing state-of-the-art methods.