=Paper= {{Paper |id=Vol-2009/fmt-proceedings-2017-paper1 |storemode=property |title=Towards Automated Real Estate Assessment from Satellite Images with CNNs |pdfUrl=https://ceur-ws.org/Vol-2009/fmt-proceedings-2017-paper1.pdf |volume=Vol-2009 |authors=Valentin Muhr,Miroslav Despotovic,David Koch,Mario Döller,Matthias Zeppelzauer |dblpUrl=https://dblp.org/rec/conf/fmt/MuhrDKDZ17 }} ==Towards Automated Real Estate Assessment from Satellite Images with CNNs== https://ceur-ws.org/Vol-2009/fmt-proceedings-2017-paper1.pdf

Towards Automated Real Estate Assessment from
Satellite Images with CNNs
Valentin Muhr Miroslav Despotovic David Koch
Kufstein University of Applied Sciences Kufstein University of Applied Sciences Kufstein University of Applied Sciences
Email: v.muhr@a1.net Email: miroslav.despotovic@fh-kufstein.ac.at Email: david.koch@fh-kufstein.ac.at

Mario Döller Matthias Zeppelzauer
Kufstein University of Applied Sciences St.Pölten University of Applied Sciences
Email: mario.doeller@fh-kufstein.ac.at Email: matthias.zeppelzauer@fhstp.ac.at

Abstract—A driving factor for real estate prices is the location [5], [6], [7]. Due to these comprehensive data former limita-
quality. Models for location quality are usually built from tions to recognize fine-grained structural patterns of objects
available price information and distinct GIS information. In (e.g. buildings) and background are dissolved [8], and thereby
this paper, we present a first approach towards the automated
assessment of location quality from satellite images using com- enable a number of novel approaches, as the one presented in
puter vision. For this purpose, we first introduce a novel dataset this work.
generated from publicly available data sources with suitable A rich literature on satellite image analysis and land cover
ground-truth annotations for location assessment. Next, we adapt classification exists [9], [10] and numerous datasets have been
a state-of-the-art convolutional neural network (CNN) and adapt introduced [11], [12] for developing automatic classification
it to predict different land covers and objects from satellite
images. Finally, we feed information derived from the recognized and detection algorithms. Furthermore, recently, deep learn-
land covers into a regression-based price model which acts as a ing has made a significant impact on the field, especially
proxy for the assessment of location quality. Our results show Convolutional Neural Networks (CNNs) [13] have improved
that (i) land cover classification can be performed with high classification performance of satellite imagery significantly
accuracy and demonstrates that automatic classification could [9], [14], [15], [16], [17]. Although there is work on many
further be used in the future for the detection of mis-aligned
and erroneous GIS data; (ii) our adapted network reaches state- different tasks in the context of satellite image analysis, such
of-the-art performance in much less training time compared to as building detection, road network analysis, terrain classifica-
our reference network; (iii) the automatically extracted visual tion, flood detection, and even poverty detection, the detection
information improves the prediction of real estate prices and of location quality has, to the knowledge of the authors, not
thereby shows clear potential for the description of location been investigated so far [18], [19], [20], [21], [22].
quality.
For a proper development and implementation of our re-
search objectives, we had to build our own satellite image
I. I NTRODUCTION
dataset, which suits the requirements for location quality as-
One of the most important criteria for assessing real estate sessment. For this purpose, we used publicly available satellite
is its location and its neighborhood. The by the authors of image data from the Tyrolean Tiris Database1 and the freely
this paper specified research objective is to assess the quality available Open Government Data2 (OGD). For location quality
and livability of urban geographical locations to support the assessment we have defined a set of eight relevant land covers,
automatic assessment of real-estate objects and properties. To which reflect geological characteristics and also considerable
realize this goal, we aim at leveraging the capabilities of infrastructure in Austria. Based on this novel dataset, we
automated land cover analysis in satellite images. designed a CNN for location assessment and compared it to
Land Cover Analysis focuses on the extraction of infor- an existing network (VGG network model) [23]. The networks
mation from remote sensing satellite image data and GIS were designed (and adapted respectively) to classify the land
data with the goal to study and monitor geological resources covers and objects in our dataset.
and their dynamic changes [1]. In this aspect, remote sensing This work represents a first step towards our main research
has become more and more important due to the increasing objective to assess location quality by providing the necessary
amount of available geographical data [2], provided through basis information for this task, namely pre-classified land
satellite recordings world-wide and comprehensive descriptive 1 Orthofoto Tirol - data.gv.at. [Online] Available:
information collections from GIS. These data allow better https://www.data.gv.at/katalog/dataset/35691b6c-9ed7-4517-b4b3-
understanding and development of the earths global and phys- 688b0569729a. Accessed on: Aug. 30, 2017
2 Datenauftritt Land Tirol — data.gv.at. [Online] Avail-
ical processes [3] and in combination with GIS have fostered
able: https://www.data.gv.at/auftritte/?organisation=land-
several applications and meaningful results in the fields of tirol&katFilter=geographie-und-planung#showresults. Accessed on: Aug
agriculture, environment, and eco-environment assessment [4], 30, 2017

14
Towards Automated Real Estate Assessment from Satellite Images with CNNs

covers, terrains and objects in RGB satellite images. The [35], [36], [37], road networks [38], [39], [40], and the
training of regression models for location quality based on segmentation or classification of different land covers, such
this analysis data is currently under investigation. The hedonic as grassland, forest, and water [41], [42], [43], [44]. Satellite
pricing method, provides the theoretical background in real images are usually cut into smaller patches before analysis,
estate and is widely used in housing research and appraisals in order to provide local spatial information (at a particular
[24]. Originally developed for automobiles by [25], hedonic location) together with its surrounding context (neighborhood)
price models describe how the quantity and quality of these to the subsequent analysis. Afterwards, traditionally feature
characteristics determine its price in a particular market. In a extraction is employed followed by machine learning [28].
formal way, the general hedonic price function takes the form Deep Learning, in contrast, combines feature extraction and
classification into one process, which is demonstrated by an
Pi = f (Si , Li , Ni ) (1)
early work by Mnih and Hinton [45]. This work is based on a
where Pi is the log of the price or rent of house i, Si is a neural network with multiple local and fully-connected layers,
vector of structural housing characteristics, Li is a vector of designed to detect roads and road networks. In recent years
location variables and Ni is the neighborhood characteristics. many different methods for feature learning based on neural
In the field of hedonic pricing, only the aspects that are networks have been introduced [46], [47], [41], [31], [48].
available as quantifiable data, e.g. for location variables, such In these methods, the extracted features (neural activations
as distances to schools, public transport, etc., are considered. at intermediate layers of the networks) are further processed
For instance, if one wants to evaluate sociodemographic data by other classifiers, such as Support Vector Machines (SVM)
within the scope of hedonic pricing on a certain geographical to obtain a final classification. In other works the networks
level, an appropriate statement can be developed only on that combine feature extraction and learning into one supervised
particular aggregated level. Therefore, the exact location or the learning process by adding a classification layer on top of the
immediate neighborhood can be valuated only conditionally in network [45], [22], [44], [49], [50], [51], [32], [43].
a fully automatic way. Furthermore, the demarcation of micro- Higher-level tasks often build upon a basic land cover
residential areas based on e.g. predefined spatial units, and classification or segmentation and include, for example build-
price similarity cannot always lead to an adequate qualitative ing type classification, i.e., the classification of residential
distinction. Pattern recognition from satellite images circum- buildings, apartment buildings, and industrial and factory
vents this problem, as the factual delineation of the macro- building [52], [53] to support urban development and proper
location is ignored and only the features within selected micro- planning. Higher-level tasks exist for different scaling levels.
location are being investigated. An example for a very fine-grained analysis is the segmen-
tation of rooftops of buildings into areas of different angles
II. R ELATED W ORK [35] and thereby the evaluation of the suitability of these
The major objective of image analysis is to extract discrim- rooftops for photo-voltaic systems [54], [55]. Other works
inative visual features from images [26]. An early approach, focus on higher-level tasks at a much larger scale, such
which is sometimes still applied today, is the application of as the analysis of abstract demographic and socio-economic
manually designed visual features based on color- and edge- parameters from satellite images. An example is the work of
descriptors [27], [28], e.g. histograms and correlograms. Later [56] where the authors try to derive regions of poverty through
local features (e.g. SIFT, SURF, ORB) were introduced and satellite image data. They exploit the light intensity in areas
became popular within the Bag-of-Visual-Word model for of settlement in night satellite images as an indicator. Based
image classification [29]. During the last years, increasingly on this information they apply transfer learning to identify
more automated methods for feature extraction and entirely characteristical visual features which indicate poverty. In a
end-to-end learned approaches based on neural networks (deep follow-up work, the authors were able to explain 75% of the
learning, DL) have been introduced and achieved remarkable variation of economic outcomes from satellite images with a
improvements in image analysis and classification [30], [31], CNN [19].
[32]. The assessment of location quality is a new high-level task
In remote sensing different image types are utilized, ranging for satellite image analysis introduced in this paper, which
from grayscale and RGB images, to multi-spectral images, and builds upon low-level land cover classification. To facilitate the
Synthetic Aperture Radar (SAR) images, as well as measuring characterization of location quality from land covers, quality-
procedures such as Light Detection and Ranging (LiDAR) relevant land covers need to be selected and recognized in
[33] enabling 3-dimensional geographical surface scanning. the images. We have performed a survey on satellite image
RGB satellite images are the most widely distributed resources datasets to identify suitable annotated data collections for
in remote sensing, since region-wide captures are broadly location assessment.
available. One of the most recent dataset is the AID dataset. It has
In RGB satellite image analysis we can differentiate be- been introduced in early 2017 by [57] and covers 10.000 object
tween low-level tasks and (more sophisticated) higher-lever images divided into 30 individual categories. The set entails
tasks, which often build upon low-level tasks. Low-level tasks high intra-class variations, since visible lighting conditions
focus on detection of certain objects, such as buildings [34], and time or season differ from sample to sample. Also due

15
Towards Automated Real Estate Assessment from Satellite Images with CNNs

to multi-resolution collections, the images range widely in
their spatial resolution of 50 cm and up to eight meters. For
our initial experiments on location assessment, we aim at a
more controlled dataset with less degrees of freedom (i.e.
different scales and different seasons) to reduce the number
of influencing factors. We consider the AID dataset as a
challenging dataset for future experiments, once a first location
assessment method has been established.
The RSSCN7 dataset, introduced by [58], contains 2800
satellite images with eight different annotated land cover
classes. With 400 images in the categories of grassland, forest,
farmland, parking lot, residential region, industrial region, and (a) extracted satellite image with ground-truth
river and lake, these classes are closely resembled to the
introduced dataset of current reading. Challenging is the fact
that the samples are captured under different seasons, weather
conditions and resolutions. The overall size, however, is much
smaller than our dataset (10k images per class vs. 400 images
per class).
The UC Merced Land Use Dataset by [59], is one of the
most utilized datasets [60], [57], [61], [62], [63] and has been (b) patch to class assignment (c) sliding window ap-
proach
introduced in 2010. This set covers 21 object classes with 100
images each, in a spatial resolution of 30 cm. Although, this Fig. 1. (a) (left) extracted satellite image; (right) extracted ground truth. (b)
dataset covers mostly all of the, for this research determined, Patch extraction algorithm, localizing the centered object and identifying its
corresponding class. (c) Sliding window approach, visualizing the right and
land covers and objects, the research project of this reading downward movement by 90px.; Credit: Land Tirol - data.tirol.gv.at
requires far more image samples per category.
Due to the reasons mentioned above, we have created a new
satellite dataset from publicly available satellite and GIS data hardly or non-settled (covered settlement area 11.8% of whole
to better address our research objectives. Tyrol [64]) we have selected those tiles which cover settlement
areas, resulting in a total of 21,076 tiles. Each obtained
III. DATASET tile (4,050x4,050 px) covers 500x500 m2 . The tiles have a
The satellite image dataset for our investigation has ground resolution of 8.1 cm2 per pixel. The vector layers were
been retrieved from the orthophoto map of the Tyrolean rasterized with the same resolution as the satellite image tiles
Tiris Database in combination with Open Government Data for further processing. Figure 1(a) shows a part of an image
(OGD). Images together with the OGD have been loaded into tile with the corresponding ground-truth. The dataset has then
the GIS software QGIS (version 2.18.2), which enables to been partitioned randomly into a training set (80%), including
align the satellite images and the vector layers of the OGD a hold-out validation set (10%), and a test set (20%). This
and to process them together. The OGD covers multiple initial split guarantees a clear separation of training and test
different land cover and object categories. From the rich data in our experiments.
available data, we have selected those categories that are Next, every satellite image and ground truth layer has been
most relevant for our research purpose of location quality processed in parallel by a sliding window operation to cut
assessment. A total of eight object and land cover categories out patches from the images. The patch size is determined by
has been selected: 4050px/15 = 270px. A single patch thus covers 478 m2 of
landscape and thereby provides a significant amount of visual
Residential Buildings Local Roads context for the detection of land covers and objects.
Non-Residential Buildings Grassland For each image patch a ground-truth label has been com-
Highway Water puted. The ground-truth label is derived from the center area
Federal Highway Forest of a patch (90x90 px) and is specified as the label of the most
frequently occurring land cover in this area. The remaining
These eight classes are assumed to be most descriptive area of the patch is considered as contextual information and
and provisioning for the visual representation of local land is not used to define the class label. Figures 1(b) illustrates
covers and infrastructures in Austria. Together they represent the labeling process of an image patch, showing a residential
the ground truth classes of our dataset. By using the scripting building.
capabilities of QGIS we have exported large-scale image tiles The patch extraction is performed with a step size of 90 px
from the satellite images at the highest available resolution. for the sliding window operation. In this way a dense coverage
The area from which tiles were extracted covers the entire of the input tiles is obtained, see Figure 1(c). Thereby, each
region of Tyrol in Austria. Since large areas in Tyrol are tile is cut into 1,849 patches.

16
Towards Automated Real Estate Assessment from Satellite Images with CNNs

Fig. 2. The process workflow of database construction.

TABLE I
N UMBERS AND PERCENTAGES OF MIS - ASSIGNED PATCHES INSIDE EACH
CATEGORY DUE TO INACCURATE OGD.

Category % mis-assigned
Residential Buildings 1% (132/15,000)
Non-Residential Buildings 1% (147/15,000)
Highway 57% (8,578/15,000)
Federal Highway 20% (2,991/15,000)
Local Road 6% (913/15,000)
Grassland 22.5% (3390/15,000)
Water 35.5% (5,332/15,000)
Forest 10% (1,489/15,000)

Fig. 3. An OGD polygon which is indicating the shape of a federal highway.; neural networks (CNNs). As described in Section III, the
Credit: Land Tirol - data.tirol.gv.at
satellite images have been pre-processed, i.e. cut into patches
by applying a sliding window operation to the large-scale
During construction of the dataset it showed that the OGD image tiles. The patches are stored in a HDF5 image database
shape files (vector layers) frequently exhibit poor accuracy. and fed into the employed networks
Distorted and inaccurate polygons occur frequently, which do
A. Center Patch Approach
not match with the satellite image. Figure 3, demonstrates
an example of an inaccurate OGD polygon. The visualized The center patch approach is inspired by the work of [22]
polygon should indicate the shape of a federal highway but and has been considered already during database construction.
actually overlaps to a large degree with the nearby forest. The idea of this approach is that the networks should learn to
Similar observations have been made in all other categories. classify only the center area of a given patch and to use the
To clean the dataset, we have manually investigated the input surrounding context information as additional input. For this
tiles and the patches and removed those with inaccurate reason, we label each patch according to the majority class in
OGD polygons. For each category 15,000 patches have been the center of the patch (see Section III). The visual context
manually investigated. The amount of patches with inaccurate is used for making more accurate decisions and is not being
ground-truth are listed in Table I. The percentage reaches from directly predicted. By following this approach the networks
6% for local roads to 57% for highway. The only exception are tuned to classify the image center by using additional
are the two building classes with a tolerable deviation of only information from the neighborhood. This is reasonable since
1%. Since category of highway and water consisted of so many some objects are more frequently surrounded by certain land
wrong images, additional images to the previous 15,000 have covers than others, e.g. gardens around residential buildings,
been investigated. The final training set consists of 10,000 woods and bushes next to highways and water. Additionally,
patches per category and the test set contains another 2,000 context information helps to disambiguate larger objects, that
patches per category. Thus, there are 80,000 patches in the cover the entire center area. For the network training this
training set and 16,000 patches in the. The complete process means that the entire patch is fed in as input and the label for
of dataset generation is summarized in Figure 2. Finally, the center region is learned, independent of the land covers
the training dataset is converted to HDF5 format for easier present in the neighborhood.
processing.
B. Network Models
IV. M ETHOD For the training process, two network models have been em-
For the classification of the different land covers and objects ployed for land cover classification. First, the VGG Network
in our dataset we propose - in accordance with the current [23], which acts as a baseline and second our own model,
state of the art - end-to-end learning by deep convolutional called SatNet-8 in the following. The VGG model is used

17
Towards Automated Real Estate Assessment from Satellite Images with CNNs

with its original architecture. Only minor adaptations have TABLE II
been made, i.e. increase image input size to 270x270, set COMPARISON OF ARCHITECTURE BETWEEN VGG AND S AT N ET-8

number of output neurons in the last fully-connected layer
to the number of classes in our dataset and application of Layer Grp. VGG SatNet-8
the ’adam’ optimizer instead of ’RMSProp’. SatNet-8 is a Group 1 conv1 - 64 filters conv1 - 128 filters
conv2 - 64 filters conv2 - 128 filters
variation of VGG with the same input and output dimensions. max pool max pool
For training, the same parameters and number of epochs are batch normalization
used to enable a fair performance comparison. Group 2 conv3 - 128 filters conv3 - 128 filters
conv4 - 128 filters conv4 - 128 filters
1) VGG Network: The VGG network is a CNN with 16 max pool max pool
layers, i.e. 13 convolutional (conv) layers and three fully- batch normalization
connected (fc) layers, which have been formed to logical Group 3 conv5 - 256 filters conv5 - 256 filters
conv6 - 256 filters conv6 - 256 filters
groups. The first group represents two conv layers and one max conv7 - 256 filters
pooling operation. The conv layers operate with 64 filters and max pool max pool
a filter size of 3 and ReLU as activation function. Max pooling batch normalization
Group 4 conv8 - 512 filters conv7 - 256 filters
is set to a stride of 2. In the second group the number of filters conv9 - 512 filters conv8 - 256 filters
per conv layer increases to 128. The third group consists of conv10 - 512 filters
three conv layers with 256 filters each and one max pooling max pool max pool
batch normalization
operation. The fourth and the fifth group exhibit 512 filters for Group 5 conv11 - 512 filters conv9 - 512 filters
every conv layer. After the conv layers, two fc layers follow. conv12 - 512 filters conv10 - 512 filters
These fc layers consist of 4096 neurons each and employ conv13 - 512 filters
max pool max pool
again ReLU for activation. Followed by a dropout function batch normalization
with a 50% dropout rate after the first and second fc layer. fc1 - 4096 nodes fc1 - 1024 nodes
Dropout helps to avoid overfitting by randomly suppressing dropout - 50% rate dropout 50% rate
fc2 - 4096 nodes fc2 - 1024 nodes
weight updates during learning [65]. The third and final fc dropout - 50% rate dropout 50% rate
layer defines the output and has been changed to a number fc3 - 8 nodes fc3 - 8 nodes
of K = 8 neurons to be compatible to our dataset. This layer regression function regression function
uses a softmax activation function. See Table II for a overview # parameters: 180.8 M # parameters: 43.8 M
of the architecture. Finally a logistic regression function is
applied.
many object types in our dataset are characterized by simple
2) SatNet-8: During the development of SatNet-8 the VGG
shapes, such as lines (e.g. boarders of buildings and roads).
Network has been used as a reference model. Multiple struc-
The max pooling is set to a stride of 2 and thereby always
tural modifications have been made and each training session
halves the input dimension. The following two groups exhibit
has been investigated in relation to its previous performance
256 filters, which is again doubled to a total of 512 in the fifth
and structural design. Our basic assumption for the opti-
group. The first and the second fc layers consist of only 1048
mization is that we assume satellite images to have a lower
neurons with ReLU activation. Since, the number of classes
overall complexity than arbitrary (object-related or scene-
in our dataset is much smaller than that used for the original
related) images, which were originally used to train VGG
VGG network, this reduction in complexity is reasonable. To
and similar networks. Thus, we expect that the complexity
avoid overfitting dropout functions with a 50% dropout rate are
of the network can be reduced without loosing discriminative
employed for the fc layers. The final output layer is similar to
power. Therefore, most modifications have been made with
VGG Networks output layer, including the logistic regression
the goal to reduce the models complexity and to keep training
function. The adaptations performed for the SatNet-8 network
performance at a high level.
lead to a reduction of parameters by approximately a factor of
After evaluating several different modifications and strate-
four. Both networks are initialized with random weights and
gies for the reduction of complexity, we decided for the follow-
trained from scratch in our experiments.
ing architecture. The SatNet-8 consists of only 10 conv layers
(instead of 13 as in VGG) and three smaller fc layers, which C. Implementation
have been arranged into five groups. Every group represents We employ Tensorflow with GPU support by CUDA Toolkit
two conv layers and a max pooling operation, followed by a (version 7.5) and cuDNN (version 5.1) for training. On top
batch normalization (bn) layer. This layer acts as a regularizer of Tensorflow, the higher-lever API TFLearn3 has been em-
for the distribution of the inputs to a given layer. It normalizes ployed, which allows for rapid prototyping of experimental
the layers input for each batch during the training process. As a setups while still retaining full access to Tensorflows capaci-
result, it allows higher learning rates and improves accuracy. ties. The training hardware has been an Intel Core i7-7700K
The first two groups of conv layers operate with 128 filters with an Nvidia GeForce GTX 1080 Ti.
with a filter size of 3 and ReLU activation. Since the first conv
layers of a network model recognize simple edge and shape 3 A. Damien, TFLearn — TensorFlow Deep Learning Library. [Online]
information, the number of filters has been increased because Available: http://tflearn.org. Accessed on: Aug 30, 2017

18
Towards Automated Real Estate Assessment from Satellite Images with CNNs

Fig. 4. Training and validation accuracy over the number of batch iterations.

network models reach a similar performance level after 10
epochs. Interestingly, however, SatNet-8 faster learns, espe-
V. E XPERIMENTS & R ESULTS cially in early epochs. The VGG takes longer and more
A. Training Setup iterations to reach the same accuracy as the SatNet-8. Also
Prior to training, the networks’ input layers have been the SatNet-8 seems to be more robust on the validation set.
adjusted to an input dimension of 270x270 px (the size of The main reason for the faster learning of SatNet-8 is the
the image patches in our dataset) and three color channels.We significantly lower number of parameters. This is also reflected
performed different pre-processing steps on the input images. by the training times. The training of the VGG network has
First, we zero-center the images to normalize them regarding taken 122.6 hours, whereas the SatNet-8 finished training after
illumination and contrast variations. We zero center each im- approx. 79.6 hours. As shown in Figure 4, both networks have
age by the measured mean across all three RGB color channels potentially not reached their best performance capabilities after
only. This normalization facilitates the networks to converge 10 epochs as the validation accuracy continuously increases
faster. Second, we apply augmentation to the training images. which further shows that the network does not overfit on
Augmentation adds additional variation to the training set by the data. Anyway, due to the total amount of training time,
performing image transformations on the training patches. We we stoped training after epoch 10. The final result is an
employ two transforms to the patches: rotations by 0, 90, 180, overall training accuracy of VGG of 86.55% and SatNet-8
and 270 degree and vertical flipping. While the training data of 84.98%. The validation accuracy of VGG is 84.51% and
is loaded into the system, the methods randomly rotate every that of SatNet-8 is 84.48%. There is no significant difference
image and subsequently perform vertical flipping. between the performance measures, except for the significantly
The learning rate is a key parameter in training and in- lower training time of SatNet-8.
fluences the convergence of the training significantly. A too
high rate can lead to an unstable training while a too low rate B. Classification Results
unnecessarily increases the required training time. We employ After both networks have been trained on our dataset, we
stochastic gradient descent (adam optimizer) for learning with have evaluated them on the so far unseen test set. Therefore,
an initial learning rate of 0.0001. This optimizer computes every test image patch has been passed through the network
adaptive learning rates for each parameter during the training and the most probable class from the eight possible classes
process[66] and yields good and stable results in practice. To has been taken as the final prediction for the patch. To assess
measure classification loss, we utilize the categorical cross- the performance, we have computed the overall accuracy as
entropy loss function. well as a confusion matrix. The confusion matrix provides
To monitor the training process, we have employed a insight into the quality of classification as it displays mis-
validation set. This set has contained 10% of all training classifications between all categories and thereby reveals the
samples and has been chosen randomly thereof. These samples networks weaknesses and strengths. Figure 5 and Figure 6
serve the network to validate its current performance after a present the confusion matrices with the percentages of cor-
training iteration and are never used for training itself. The rectly and incorrectly classified test image patches. The top
training batch size has been adjusted to 32 images. A value of row indicates the predicted categories (pc) and the left row
32 represents a good performance trade-off for the training provides the true label (tl). The diagonal contains the achieved
hardware employed in our experiments. We train both the classification accuracy in percent (in bold letters). The off-
networks from scratch for a number of 10 epochs. diagonal values represent mis-classifications. The column
Figure 4 visualizes the learning performance of both net- ’false positive’ contains the percentage of false predictions be-
works over the entire training. In direct comparison, both tween the two building-related categories (residental building

19
Towards Automated Real Estate Assessment from Satellite Images with CNNs

TABLE III
P ERFORMANCE COMPARISON BETWEEN VGG AND S AT N ET-8

VGG SatNet-8
average accuracy 84.75% 84.73%
building detection 95,85% 96,2%
building classification 74,7% 75,375%
road detection 93,08% 94,58%
road classification 80,1% 79,21%

Network achieved 73.35%. However, the SatNet-8 performs
Fig. 5. Confusion Matrix of the VGG Network. Values in percentage terms.
notable better in the categories of residential building 81.7%
’false positive’ describes percentage of false predictions between related object and local road 83.15%. The overall prediction accuracy and
classes. ’% mis-classification is the sum of all wrongly predicted images. predicted values of related object groups reveal that both
networks perform similar and at a very high performance level,
see Table III.
For location quality assessment the building-related cate-
gories have an especially high relevance. Thus, we want to
compare the performance level of our building type classi-
fication (see Table III) to related works. An approach for
building classification is presented in [53]. The authors classify
buildings from Light Detection and Ranging (LiDAR) data
[67] and consider three building types: single-family houses,
multiple-family houses, and non-residential buildings. A over-
all accuracy of 70% is achieved for all three classes which is
slightly lower than our accuracy of 74.7% to 75.4%. This may
Fig. 6. Confusion Matrix of the SatNet-8. Values in percentage terms. ’false on the one hand be due to the larger number of classes and
positive’ describes percentage of false predictions between related object thus the more complex tasks but on the other hand LIDAR
classes. ’% mis-classification is the sum of all wrongly predicted images.
data provides 3D information which is not available for our
method. Thus, the performance level achieved by our method
and non-residental building, first two rows) and the percentage can be considered similar. Another approach for building
of false predictions among the three street categories (highway, classification is presented in [52]. The authors investigate
federal highway, and local road, rows 3-5). Column ’% of mis- the classification of buildings into three main classes: Resi-
classifications’ sums up the percentages of wrongly predicted dential/Small Buildings, Apartment Buildings, and Industrial
images per class. The VGG Network achieved its best results and Factory Building by a Random Forest (RF) classifier
on the classes grassland, water, and forest. With an accuracy and report an average F1 score of 69.56% over all three
of 94.75% to 97.6%. VGG is capable of predicting 1,895 / classes. Although F1 score cannot be directly compared to
2,000 grassland samples, 1,918 / 2,000 water samples, and classification accuracy, it shows that the performance level is
1,952 / 2,000 forest samples correctly. When it comes to the similar.
prediction of street types, the category of highway achieved
91.05% accuracy, where only some samples tend to be mis- C. Detection of Erroneous GIS Data
classified as federal highway. Since in some areas these two As mentioned already in Section III, we have observed
street types look quiet similar, these errors are comprehensible. partly poor accuracy of the OGD polygons with the visual
A similar pattern are observed for federal highway and local content in the satellite images, which has led to some serious
road. The building-related classes can be detected with a lower complications during dataset generation. The detection of
accuracy of 74% and 75,4% for residential building and non- wrong OGD polygons could be automated by applying a
residential building, respectively. A total amount of 1,480 / reliable land cover classification or segmentation algorithm
2,000 and 1,508 / 2,000 samples have been correctly allocated to satellite image and by finding inconsistencies between the
to their corresponding class. Mis-classifications of buildings, prediction and the polygons. Thereby, mis-aligned polygons
however, mostly occur between the two building categories. of buildings, roads, rivers, lakes, forest, and other object
If we join both categories, an overall accuracy for building categories could be identified and in a further step corrected
detection of approximately 96% would be reached. (semi-)automatically.
The SatNet-8 has achieved similar results to the VGG Figure 7 provides an example image with the ground-truth
Network with slightly different prediction accuracies. The information derived from the available OGD polygons. It can
biggest difference is the class of federal highway, which is only be observed that a number of buildings are missing in the
predicted correctly in 61.85% of all cases, whereas the VGG ground-truth which is maybe due to the fact that the polygons

20
Towards Automated Real Estate Assessment from Satellite Images with CNNs

(a) original satellite image (b) original OGD GT (c) SatNet-8 predicted labels (d) color legend

Fig. 7. (a) Original satellite image from the Tiris Database; (b) Ground-truth derived from the original GIS Information (OGD polygons); (c) Predicted
patch-wise labels via SatNet-8. (d) Color legend for land cover classes. (The yellow boundaries indicate buildings and land cover objects, which have been
detected correctly by the SatNet-8, but have not been displayed accurately in the OGD GT.); Credit: Land Tirol - data.tirol.gv.at

TABLE IV of covered area) of each category within a 100 meter radius
H EDONIC P RICE M ODEL : F OR CONVENIENCE WE SHOW ONLY from each residential building. The result of the regression is
PARAMETERS FROM THE S AT N ET-8 PREDICTIONS . C OLUMN TWO SHOWS
THE EFFECT OF THE COEFFICIENTS ON THE LOGARITHMUS OF THE PRICE . displayed in Table IV. For convenience we cut out the location
R2 IST THE ADJUSTED COEFFICIENT OF DETERMINATION AND RMSE coefficients. A ll m unicipality d ummies a re s ignificant and
MEANS THE ROOT- MEAN - SQUARE ERROR reflect the expected magnitude, but are not in the focus of this
Hedonic Price Model paper. The categories from the SatNet-8 reflect t he expected
Residential Buildings 0.004∗ magnitude and most are statistically significant. F or example
Non-Residential Buildings 0.005 the category grassland shows that a higher portion of grassland
Highway −0.012∗∗
Federal Highway −0.007∗∗ results in a higher price. On the other side, the presence
Local Road −0.001 of highway or federal highway in the near neighborhood
Water 0.001 indicates a negative impact which is related to a decline in
Grassland 0.002∗∗∗
Forest 0.002∗∗
price. This is in line with real estate theory. Our investigation
Adj. R2 0.66 provides first i nsights t hat c onfirm a po sitive co ntribution of
Num. observations 2739 the automatically extracted visual parameters and demonstrates
RMSE 0.34
∗∗∗ that this novel type of modeling location quality has promising
p < 0.001, ∗∗ p < 0.01, ∗ p < 0.05
potential.
VI. C ONCLUSION
are rather old. In Figure 7(c) the patch-wise prediction of the In this paper, we presented a first a pproach t owards the
SatNet-8 is depicted. We can see that several buildings that automated assessment of location quality from satellite images.
were not present in the ground-truth were correctly detected We adapted the SatNet-8 to predict different land covers
(highlighted in yellow boundaries). We thus, see a great and objects from our novel dataset. There out, we have fed
potential for the automatic detection of inaccurate GIS data information derived from the recognized land covers into a
by automated land cover classification. regression-based price model which acts as a proxy for the
assessment of location quality. Our results show that (i) land
D. Hedonic Pricing
cover classification c an b e p erformed w ith h igh a ccuracy and
Location quality and livability are usually approximated demonstrate that automatic classification could further be used
by the price or price-level of a location. This means that in the future for the detection of mis-aligned and erroneous
by predicting the price of a location, conclusions about the GIS data; (ii) our adapted SatNet-8 reaches state-of-the-art
location quality can be drawn. A popular approach for the performance in much less training time compared to the ref-
modeling of prices of real estate is hedonic pricing (which erence VGG Network; (iii) the automatically extracted visual
we briefly described in the introduction). We evaluated the information improves the prediction of real estate prices and
utility of our land cover classifications for hedonic price thereby shows clear potential for the description of location
modeling by feeding parameters, derived from the SatNet-8 quality. Future work will include the extension of the proposed
predictions, to the hedonic (regression) model as additional approach to larger areas including urban areas like cities and
inputs. Therefore, we used 2739 valuations of residential additional land covers.
buildings, as we focus only on the location and neighborhood
characteristics. In our model, for the location variables, we ACKNOWLEDGMENT
used only the municipalities in Tyrol, Austria. Additional to This work was supported by the Austrian Research Pro-
this information, we used the information from our eight land motion Agency (FFG), Project No. 855784 and Project No.
cover clusters. Therefore, we calculated the share (in terms 856333.

21
Towards Automated Real Estate Assessment from Satellite Images with CNNs

R EFERENCES https://pdfs.semanticscholar.org/1b3a/c4b4187a3dbc9373869e7774b1dc6
3f748d2.pdf
[1] C. Berlanga-Robles and A. Ruiz-Luna, “Land use mapping and change [20] L. J. Zhao, P. Tang, and L. Z. Huo, “Land-use scene classification using
detection in the coastal zone of northwest mexico using remote sensing a concentric circle-structured multiscale bag-of-visual-words model,”
techniques,” vol. 18, pp. 514–522, 06 2002. IEEE Journal of Selected Topics in Applied Earth Observations and
[2] Z. Wang and L. Liu, “Assessment of coarse-resolution land cover Remote Sensing, vol. 7, no. 12, 2014.
products using casi hyperspectral data in an arid zone in northwestern [21] Y. Cao, X. Niu, and Y. Dou, “Region-based convolutional neural
china,” Remote Sensing, vol. 6, no. 4, pp. 2864–2883, 2014. [Online]. networks for object detection in very high resolution remote sensing im-
Available: http://www.mdpi.com/2072-4292/6/4/2864 ages,” in 2016 12th International Conference on Natural Computation,
[3] A. T. Hudak and C. A. Wessman, “Textural analysis of historical aerial Fuzzy Systems and Knowledge Discovery, ICNC-FSKD 2016, 2016.
photography to characterize woody plant encroachment in south african [22] S. Saito and Y. Aoki, “Building and road detection from large aerial
savanna,” Remote sensing of environment, vol. 66, no. 3, pp. 317–330, imagery,” vol. 9405, 02 2015.
1998. [23] K. Simonyan and A. Zisserman, “Very deep convolutional networks
[4] X. Li and A. G. O. Yeh, “Principal component analysis of stacked multi- for large-scale image recognition,” CoRR, vol. abs/1409.1556, 2014.
temporal images for the monitoring of rapid urban expansion in the pearl [Online]. Available: http://arxiv.org/abs/1409.1556
river delta,” vol. 19, pp. 1501–1518, 05 1998. [24] W.-C. Liao and X. Wang, “Hedonic house prices and spatial quantile
[5] H. Chang and W. S. Yoon, “Improving the classification of landsat regression,” Journal of Housing Economics, vol. 21, no. 1, pp. 16–27,
data using standardized principal components analysis,” KSCE Journal 2012.
of Civil Engineering, vol. 7, no. 4, pp. 469–474, Jul 2003. [Online]. [25] A. T. Court, “Hedonic Price Indexes with Automotive Examples,” in The
Available: https://doi.org/10.1007/BF02895842 Dynamics of Automobile Demand, C. F. Roos, Ed. New York: General
Motors, 1939, pp. 99–117.
[6] H. Long, X. Wu, W. Wang, and G. Dong, “Analysis of urban-rural
[26] G. Kumar and P. K. Bhatia, “A detailed review of feature extraction
land-use change during 1995-2006 and its policy dimensional driving
in image processing systems,” in 2014 Fourth International Conference
forces in chongqing, china,” Sensors, vol. 8, no. 2, pp. 681–699, 2008.
on Advanced Computing Communication Technologies, Feb 2014, pp.
[Online]. Available: http://www.mdpi.com/1424-8220/8/2/681
5–12.
[7] P. K. Mallupattu and J. R. S. Reddy, “Analysis of land use/land cover
[27] J. A. dos Santos, O. A. B. Penatti, and R. da Silva Torres, “Evaluating
changes using remote sensing data and gis at an urban area, tirupati,
the potential of texture and color descriptors for remote sensing image
india,” vol. 2013, p. 268623, 05 2013.
retrieval and classification.” in VISAPP (2), 2010, pp. 203–208.
[8] J. Han, D. Zhang, G. Cheng, L. Guo, and J. Ren, “Object detection
[28] Y.-W. Seo, C. Urmson, and D. Wettergreen, “Exploiting Publicly Avail-
in optical remote sensing images based on weakly supervised learning
able Cartographic Resources for Aerial Image Analysis.”
and high-level feature learning,” IEEE Transactions on Geoscience and
[29] J. Sivic and A. Zisserman, “Video google: A text retrieval approach to
Remote Sensing, vol. 53, no. 6, pp. 3325–3337, June 2015.
object matching in videos,” in null. IEEE, 2003, p. 1470.
[9] G. Cheng and J. Han, “A survey on object detection in optical remote [30] K. Nogueira, O. A. B. Penatti, and J. A. Dos Santos, “Towards Better
sensing images.” ISPRS Journal of Photogrammetry & Remote Sensing, Exploiting Convolutional Neural Networks for Remote Sensing Scene
2016. Classification,” 2016.
[10] D. Lu, Q. Weng, E. Moran, G. Li, and S. Hetrick, Remote sensing image [31] O. A. B. Penatti, K. Nogueira, and J. A. dos Santos, “Do deep features
classification. CRC Press/Taylor and Francis: Boca Raton, FL, USA, generalize from everyday objects to remote sensing and aerial scenes
2011. domains?” in 2015 IEEE Conference on Computer Vision and Pattern
[11] C. Homer, J. Dewitz, L. Yang, S. Jin, P. Danielson, G. Xian, J. Coul- Recognition Workshops (CVPRW), June 2015, pp. 44–51.
ston, N. Herold, J. Wickham, and K. Megown, “Completion of the [32] K. Nogueira, W. O. Miranda, and J. A. D. Santos, “Improving spatial
2011 national land cover database for the conterminous united states– feature representation from aerial scenes by using convolutional net-
representing a decade of land cover change information,” Photogram- works,” in 2015 28th SIBGRAPI Conference on Graphics, Patterns and
metric Engineering & Remote Sensing, vol. 81, no. 5, pp. 345–354, Images, Aug 2015, pp. 289–296.
2015. [33] J. Zhao and S. You, “Road network extraction from airborne lidar data
[12] G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu, using scene context,” in 2012 IEEE Computer Society Conference on
“Aid: A benchmark data set for performance evaluation of aerial scene Computer Vision and Pattern Recognition Workshops, June 2012, pp.
classification,” IEEE Transactions on Geoscience and Remote Sensing, 9–16.
2017. [34] J. P. Cohen, W. Ding, C. Kuhlman, A. Chen, and L. Di, “Rapid building
[13] Y. LeCun, Y. Bengio et al., “Convolutional networks for images, speech, detection using machine learning,” CoRR, vol. abs/1603.04392, 2016.
and time series,” The handbook of brain theory and neural networks, [Online]. Available: http://arxiv.org/abs/1603.04392
vol. 3361, no. 10, p. 1995, 1995. [35] Y. E. Merabet, C. Meurie, Y. Ruichek, A. Sbihi, and R. Touahni,
[14] T. Ishii, R. Nakamura, H. Nakada, Y. Mochizuki, and H. Ishikawa, “Building roof segmentation from aerial images using a line-and region-
“Surface object recognition with CNN and SVM in Landsat 8 images,” based watershed segmentation technique,” in Sensors, 2015.
in Proceedings of the 14th IAPR International Conference on Machine [36] A. O. Ok, C. Senaras, and B. Yuksel, “Automated detection of arbi-
Vision Applications, MVA 2015, 2015. trarily shaped buildings in complex environments from monocular vhr
[15] P. N. Druzhkov and V. D. Kustikova, “A survey of deep learning methods optical satellite imagery,” IEEE Transactions on Geoscience and Remote
and software tools for image classification and object detection,” Pattern Sensing, vol. 51, no. 3, pp. 1701–1717, March 2013.
Recognition and Image Analysis, vol. 26, no. 1, pp. 9–15, 2016. [Online]. [37] F. Dornaika, A. Moujahid, Y. E. Merabet, and Y. Ruichek, “Building
Available: http://link.springer.com/10.1134/S1054661816010065 detection from orthophotos using a machine learning approach: An
[16] D. Tuia, C. Persello, and L. Bruzzone, “Domain adaptation for the empirical study on image segmentation and descriptors,” Expert Systems
classification of remote sensing data: An overview of recent advances,” with Applications, vol. 58, pp. 130 – 142, 2016. [Online]. Available:
2016. http://www.sciencedirect.com/science/article/pii/S0957417416301154
[17] G. Cheng, P. Zhou, and J. Han, “Learning Rotation-Invariant Con- [38] J. D. Wegner, J. A. Montoya-Zegarra, and K. Schindler, “A higher-order
volutional Neural Networks for Object Detection in VHR Optical crf model for road network extraction,” in 2013 IEEE Conference on
Remote Sensing Images,” IEEE Transactions on Geoscience and Remote Computer Vision and Pattern Recognition, June 2013, pp. 1698–1705.
Sensing, 2016. [39] J. A. Montoya-Zegarra, J. D. Wegner, L. Ladickỳ, and K. Schindler,
[18] S. Ahmed, M. Liwicki, M. Weber, and A. Dengel, “Mind the gap: modeling local and global context in (road) networks,”
“Improved Automatic Analysis of Architectural Floor Plans,” in German Conference on Pattern Recognition. Springer, 2014, pp.
in 2011 International Conference on Document Analysis and 212–223.
Recognition. IEEE, sep 2011, pp. 864–869. [Online]. Available: [40] C. Poullis, “Tensor-cuts: A simultaneous multi-type feature extractor
http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=6065434 and classifier and its application to road extraction from
[19] N. Jean, M. Burke, M. Xie, W. M. Davis, D. B. Lobell, and satellite images,” ISPRS Journal of Photogrammetry and Remote
S. Ermon, “Combining satellite imagery and machine learning to Sensing, vol. 95, pp. 93 – 108, 2014. [Online]. Available:
predict poverty,” Science, vol. 353, no. 6301, 2016. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0924271614001543

22
Towards Automated Real Estate Assessment from Satellite Images with CNNs

[41] F. Hu, G.-S. Xia, J. Hu, and L. Zhang, “Transferring deep convolutional [59] Y. Yang and S. Newsam, “Bag-of-visual-words and spatial extensions
neural networks for the scene classification of high-resolution remote for land-use classification,” in Proceedings of the 18th SIGSPATIAL
sensing imagery,” Remote Sensing, vol. 7, no. 11, pp. 14 680–14 707, International Conference on Advances in Geographic Information
2015. [Online]. Available: http://www.mdpi.com/2072-4292/7/11/14680 Systems, ser. GIS ’10. New York, NY, USA: ACM, 2010, pp. 270–
[42] S. Basu, S. Ganguly, S. Mukhopadhyay, R. DiBiano, M. Karki, 279. [Online]. Available: http://doi.acm.org/10.1145/1869790.1869829
and R. R. Nemani, “Deepsat - A learning framework for satellite [60] G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classifica-
imagery,” CoRR, vol. abs/1509.03602, 2015. [Online]. Available: tion: Benchmark and state of the art,” Proceedings of the IEEE, vol. PP,
http://arxiv.org/abs/1509.03602 no. 99, pp. 1–19, 2017.
[43] M. Castelluccio, G. Poggi, C. Sansone, and L. Verdoliva, “Land [61] O. A. B. Penatti, K. Nogueira, and J. A. Dos Santos, “Do Deep Features
use classification in remote sensing images by convolutional neural Generalize from Everyday Objects to Remote Sensing and Aerial Scenes
networks,” CoRR, vol. abs/1508.00092, 2015. [Online]. Available: Domains?”
http://arxiv.org/abs/1508.00092 [62] K. Nogueira, W. O. Miranda, and J. A. D. Santos, “Improving Spatial
[44] M. Längkvist, A. Kiselev, M. Alirezaie, and A. Loutfi, “Classification Feature Representation from Aerial Scenes by Using Convolutional
and segmentation of satellite orthoimagery using convolutional neural Networks,” in Brazilian Symposium of Computer Graphic and Image
networks,” vol. 8, p. 329, 04 2016. Processing, 2015.
[45] V. Mnih and G. E. Hinton, Learning to Detect Roads in High-Resolution [63] M. Castelluccio, G. Poggi, C. Sansone, and L. Verdoliva, “Land Use
Aerial Images. Berlin, Heidelberg: Springer Berlin Heidelberg, 2010, Classification in Remote Sensing Images by Convolutional Neural
pp. 210–223. Networks,” Arxiv preprint arXiv:1508.00092, aug 2015. [Online].
[46] J. Wang, J. Song, M. Chen, and Z. Yang, “Road network extraction: Available: http://arxiv.org/abs/1508.00092
A neural-dynamic framework based on deep learning and a finite state [64] Amt der Tiroler Landesregierung. Flaechennutzung. Accessed:
machine,” International Journal of Remote Sensing, vol. 36, no. 12, pp. 2017-08-29. [Online]. Available: https://www.tirol.gv.at/statistik-
3144–3169, 2015. budget/statistik/flaechennutzung/
[47] A. Lagrange and B. Le Saux, “Convolutional neural networks for [65] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
semantic labeling,” Tech. Rep., Onera–The French Aerospace Lab, Tech. R. Salakhutdinov, “Dropout: a simple way to prevent neural networks
Rep., 2015. from overfitting.” Journal of machine learning research, vol. 15, no. 1,
[48] A. Lagrange, B. L. Saux, A. Beaupère, A. Boulch, A. Chan-Hon- pp. 1929–1958, 2014.
Tong, S. Herbin, H. Randrianarivo, and M. Ferecatu, “Benchmarking [66] S. Ruder, “An overview of gradient descent optimization
classification of earth-observation data: From learning explicit features algorithms,” CoRR, vol. abs/1609.04747, 2016. [Online]. Available:
to convolutional networks,” in 2015 IEEE International Geoscience and http://arxiv.org/abs/1609.04747
Remote Sensing Symposium (IGARSS), July 2015, pp. 4173–4176. [67] S. Reutebuch, H. Andersen, and R. Mcgaughey, “Light detection and
[49] S. Paisitkriangkrai, J. Sherrah, P. Janney, and A. V.-D. Hengel, “Effective ranging (lidar): An emerging tool for multiple resource inventory,” vol.
semantic pixel labelling with convolutional networks and conditional 103, pp. 286–292, 09 2005.
random fields,” in 2015 IEEE Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW), June 2015, pp. 36–43.
[50] O. Firat, G. Can, and F. T. Y. Vural, “Representation learning for
contextual object and region detection in remote sensing,” in 2014 22nd
International Conference on Pattern Recognition, Aug 2014, pp. 3708–
3713.
[51] Q. Lv, Y. Dou, X. Niu, J. Xu, and B. Li, “Classification of land cover
based on deep belief networks using polarimetric radarsat-2 data,” in
2014 IEEE Geoscience and Remote Sensing Symposium, July 2014, pp.
4679–4682.
[52] M. Belgiu, I. Tomljenovic, T. J. Lampoltshammer, T. Blaschke, and
B. Höfle, “Ontology-based classification of building types detected
from airborne laser scanning data,” Remote Sensing, vol. 6, no. 2,
pp. 1347–1366, 2014. [Online]. Available: http://www.mdpi.com/2072-
4292/6/2/1347
[53] Z. Lu, J. Im, J. Rhee, and M. Hodgson, “Building type
classification using spatial and landscape attributes derived
from lidar remote sensing data,” Landscape and Urban
Planning, vol. 130, pp. 134 – 148, 2014. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0169204614001601
[54] N. Srećković, N. Lukač, B. Žalik, and G. Štumberger, “Determining
roof surfaces suitable for the installation of PV (photovoltaic)
systems, based on LiDAR (Light Detection And Ranging) data,
pyranometer measurements, and distribution network configuration,”
Energy, vol. 96, pp. 404–414, feb 2016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0360544215017235
[55] S. Szabó, P. Enyedi, M. Horváth, Z. Kovács, P. Burai, T. Csoknyai,
and G. Szabó, “Automated registration of potential locations
for solar energy production with Light Detection And Ranging
(LiDAR) and small format photogrammetry,” Journal of Cleaner
Production, vol. 112, pp. 3820–3829, jan 2016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0959652615010483
[56] M. Xie, N. Jean, M. Burke, D. Lobell, and S. Ermon, “Transfer Learning
from Deep Features for Remote Sensing and Poverty Mapping,” p. 16,
2015. [Online]. Available: http://arxiv.org/abs/1510.00098
[57] G. S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu,
“AID: A benchmark data set for performance evaluation of aerial scene
classification,” IEEE Transactions on Geoscience and Remote Sensing,
2017. [Online]. Available: https://arxiv.org/pdf/1608.05167.pdf
[58] Q. Zou, L. Ni, T. Zhang, and Q. Wang, “Deep learning based feature
selection for remote sensing scene classification,” vol. 12, pp. 1–5, 11
2015.