IMPROVING DATA QUALITY AND MANAGEMENT FOR REMOTE SENSING ANALYSIS: USE-CASES AND EMERGING RESEARCH QUESTIONS

Abstract

During the last decades satellite remote sensing has become an emerging technology producing big data for various application fields every day. However, data quality checking as well as the long-time management of data and models are still issues to be improved. They are indispensable to guarantee smooth data integration and the reproducibility of data analysis such as carried out by machine learning models. In this paper we clarify the emerging need of improving data quality and the management of data and models in a geospatial database management system before and during data analysis. In different use cases various processes of data preparation and quality checking, integration of data across different scales and references systems, efficient data and model management, and advanced data analysis are presented in detail. Motivated by these use cases we then discuss emerging research questions concerning data preparation and data quality checking, data management, model management and data integration. Finally conclusions drawn from the paper are presented and an outlook on future research work is given.

Full text

IMPROVING DATA QUALITY AND MANAGEMENT
FOR REMOTE SENSING ANALYSIS:
USE-CASES AND EMERGING RESEARCH QUESTIONS
Martin Breunig a, Paul Kuper a*, Friederike Reitze a, Steven Landgraf a
Mulhim Al-Doori b, Emmanuel Stefanakis c, Hussein Abdulmuttalibd, Zsófia Kugler e
a Geodetic Institute, Karlsruhe Institute of Technology, Germany –
(martin.breunig, paul.kuper, friederike.reitze, steven.landgraf)@kit.edu
b University of Science & Technology of Fujairah, College of Engineering and Technology, U.A.E. – m.aldoori@ustf.ac.ae
cDepartment of Geomatics Engineering, University of Calgary, Canada – emmanuel.stefanakis@ucalgary.ca
dGIS Department, Dubai Municipality, U.A.E. – husseinma@dm.gov.ae
eDept. of Photogramm. and Geoinformatics, Budapest Univ. of Technology and Economics, Hungary – kugler.zsofia@emk.bme.hu
ISPRS Commission IV: WG IV/4 and ICWG III/IVb
KEY WORDS: Geospatial Data Preparation, Geospatial Data Modelling, Geospatial Data Management, Geospatial Data Processing
and Analysis, Big Geospatial Data, Remote Sensing Data Quality.
ABSTRACT:
During the last decades satellite remote sensing has become an emerging technology producing big data for various application fields
every day. However, data quality checking as well as the long-time management of data and models are still issues to be improved.
They are indispensable to guarantee smooth data integration and the reproducibility of data analysis such as carried out by machine
learning models. In this paper we clarify the emerging need of improving data quality and the management of data and models in a
geospatial database management system before and during data analysis. In different use cases various processes of data preparation
and quality checking, integration of data across different scales and references systems, efficient data and model management, and
advanced data analysis are presented in detail. Motivated by these use cases we then discuss emerging research questions concerning
data preparation and data quality checking, data management, model management and data integration. Finally conclusions drawn
from the paper are presented and an outlook on future research work is given.
1. INTRODUCTION
During the last decades satellite remote sensing has become an
emerging technology producing big data every day rasing the
question of appropriate data preparation, data quality checking
and efficient data management. NASA® and COPERNICUS®
platforms, for example, are highly attractive for a multitude of
possible users in science as well as in the public and in the private
sector. Access to the data is provided by means of data and
information access services, which provide basic functionalities
to download the data and to process them to some degree.
However, as experience shows, typical data preparation
processes still consist of many single steps and advanced skills in
data handling are needed to manually extract data of a given
region in parallel for subsequent scenes or to extract data of
different regions in parallel for the same scene. Therefore, data
preparation - as an important step before writing the data into the
database - still is expensive in operator time and hinders a fast
exploitation of the data. That is why tailored geospatial database
operations to support data preparation should be provided to
achieve efficient data handling and data analysis. During data
preparation, also the quality of the data has to be checked
covering different aspects of data quality. This means e.g. that
errors in the data have to be detected and corrected and the
reliability of the data has to be documented by the authors. Also
accuracy dimensions should be considered. Furthermore, it is
important to provide an efficient long term management of big
remote sensing data and their corresponding analysis models in a
* Corresponding author
geospatial database management system. Thus data analysis and
data management should be closer integrated so that the analysis
models have direct access to the operations of the geospatial
database management system and integration of various data
sources is supported adequately.
The paper is structured as follows: In section 2, we refer to related
work followed by section 3, describing use cases for data
preparation to improve data quality, integration of elevation data
across various scales and reference systems, data management
and model integration for data analysis as well as advanced
neural network based data analysis. In section 4, emerging
research questions are derived from the use cases to improve the
preparation, quality checking, management and integration of
remote sensing data and models. Finally, section 5 presents the
conclusions drawn from the paper and gives an outlook on future
research.
2. RELATED WORK
In the context of big data analysis and geospatial data
management the improvement of data-driven workflows has
been extensively discussed (Laney, 2001; Chen et al., 2014;
Cheng et al., 2014; Lee and Kang, 2015; Breunig et al., 2016; Li
et al., 2016; Werner and Chiang, 2021). In particular, parallel
query support (Hahn et al., 2002) based on parallel hardware and
software architectures (Xiaoqiang and Yuejin, 2010; Taylor,
2010; Lenka et al., 2017; SpatialHadoop, 2023) has been
investigated. Intensive research has also been carried out in the
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
41

field of raster databases focusing on the efficient storage of raster
data (Baumann et al., 1997) and services (Baumann, 2010) to
improve the access on raster data and operations (Zhong et al.,
2011; Ouyang et al., 2013; Hu et al., 2018). The appropriateness
of existing database management systems to handle geospatial
big data, has been examined by (Amirian et al., 2014; Mazroob
et al., 2020) and other authors. A “tailored approach” to manage
raster data, considering heterogeneous data models, was
introduced by (Baumann et al., 2016). (Baumann et al., 2018)
have proven specialized data cubes as a suitable concept to
provide raster data interfaces for spatial and temporal data
analysis. Here the code is “shipped to the data” to minimize the
communication costs when transporting the data from one tool to
another. As an example of a scalable geospatial data analytics
cloud platform, the “Physical Analytics Integrated Repository
and Services” (PAIRS) homogenize archived and real-time
spatial data (Klein et al., 2015). This approach is empowered by
Hadoop® holding a parallelized structure based on MapReduce
(Klein et al., 2015). Parallel system architectures such as
Hadoop® and Spark® distribute the computation actions to a
computer cluster. They work on the basis of the Map-Reduce
model (Dean and Ghemawat, 2008), which automatically
distributes (Map) the calculation steps to the existing computers
to execute there and merge (Reduce) the intermediate results of
the map step into a solution. Concerning data analysis in remote
sensing, Artificial Intelligence is a pregnant technology to
support data handling (Lary, 2010; Lary et al. 2016; Mathieu and
Aubrecht, 2018). Supervised or unsupervised machine learning
algorithms, especially neural networks (NNs), have been
frequently used for regression and classification (Bishop, 1995),
image recognition and object detection (LeCun et al., 2015).
Multiple radar applications, ground- and satellite-based have
been proven to work with neural networks (NNs) (Qin et al.,
2004; Lombacher et al., 2016). Zhu et al. use machine learning
methods to develop algorithms from signal processing and
Artificial Intelligence to improve the extraction of geospatial
information from satellite data (Zhu et al., 2017).
However, until the present time, the data preparation and quality
checking, data selection, integration and analysis of satellite data
for scientific use are very time-consuming processes. Further
research is necessary to support data scientists and experts from
various disciplines adequately.
3. USE CASES
The following use-cases show examples how to improve the data
processing workflow during data quality checking, data and
model management, data integration, and data analysis in remote
sensing scenarios.
3.1 Data preparation and data quality: Cleansing of
Sentinel-1 SAR Data
Correctly unwrapping interferograms in Interferometric SAR
(InSAR) approaches still poses a challenge due to its ill-defined
nature and is still present in Differential InSAR (DInSAR) data
used for time series analysis (Yu et al., 2019). Hence, the
approach presented here is concerned with the automation of
finding and mending phase unwrapping errors in Sentinel-1 data.
In contrast to previous works, the problem is approached in a
data-driven manner using semi-supervised active learning
methods. In the proposed workflow, two distinct model types are
trained, namely detection and correction model. The former is
trained to detect an erroneous time series, and the latter is trained
to suggest corrections on time steps to a human observer. The
human observer inspects the proposed corrections using a
graphical user interface and adapts them if needed. These newly
labeled time series are used for further training of the models (see
Figure 1). As the provided DInSAR data is spatially sparse, i. e.
not structured in a dense raster, and provides sequential structure
in time, Long Short-Term Memory (LSTM) neural networks are
used to compensate processing errors in Sentinel-1 SAR data sets
of the Upper Rhine Graben (Oberrheingraben, ORG) and Landau
region, Germany.
The considered data were provided by (Heck, 2019) and include
two sets of processed Sentinel-1 time series. Singel Look
Complex (SLC) SAR data products of the Sentinel-1 mission of
the European Space Agency (ESA, 2023) were preprocessed and
turned into interferograms using SNAP (Leskovec and Sosič,
2016). Afterward, Persistent Scatterer DInSAR processing was
conducted using the StaMPS software (Hooper et al., 2012) with
the 3D phase unwrapping technique introduced by (Hooper et al.
2007) and (Hooper, 2010).
Four training iterations were conducted (see Figure 1): First, the
models were trained to imitate a simple, heuristic initializing
model. Second, the models’ current error detections in the
Landau data set were adapted by a human observer and used as
new labeled data for further training. In the last two iterations,
the models were trained on labeled data of the ORG data set, first
in Landau region and then on four subsets distributed over the
ORG data set, where each iteration’s labeled data include current
predictions adapted by a human observer. All training iterations
were evaluated on validation and test sets separated from the
Figure 1: Workflow of error detection and correction, general
overview with correction of example time series (left), and
interaction loop with the user (right).
Figure 2: Example of predictions and evaluation of detection
model on Landau evaluation site. TN, FP, FN, TP pixels.
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
42

training data regarding the true positives, false positives, false
negatives, and true negatives as well as the derived metrics
precision, recall, F1-score, and balanced accuracy. One example
of the trained detection model prediction on the Landau
evaluation site can be seen in Figure 2. A final evaluation
regarding those metrics was conducted on three evaluation sites:
Landau, Staufen i. B. and Lahr, for which labels were created
manually beforehand.
Regarding the detection model precision, recall, F1-score, and
balanced accuracy ranged from 0.08 to 0.28, 0.90 to 1.0, 0.14 to
0.42, and 0.93 to 0.99 respectively over the evaluation sites.
Regarding the correction model, the average metrics over all
correction classes for precision ranged from 0.46 to 0.66, the
recall from 0.79 to 0.92, the F1-score from 0.4 to 0.59 and
balanced accuracy from 0.72 to 0.82.
The results show that phase unwrapping errors can be detected
with high recall and low to moderate precision, reducing the
number of pixels a human observer must inspect. Few pixel time
series are affected by PU errors and even fewer are affected in
more than three time steps. Errors affecting the slope of the
displacement time series are scarce and mainly found near the
geothermal site at Landau where higher deformation occurs. The
detection model can reduce the amount of time series required to
be regarded to 3 % to 9 % of the total amount of pixels in a data
set, depending on the considered area. This provides an
improvement in working efficiency, as currently PU errors in PS-
DInSAR data are mainly found by carefully inspecting the whole
data set.
The correction is more challenging: It provides a high number of
falsely corrected time steps, but shows good results in case of
truly erroneous pixel time series. The portability between the two
data sets is limited, but considering the highly different provided
data sets, it is possible that the trained models are still able to
extend to more similar data sets. Furthermore, new training
iterations with training samples from another data set can be used
to transfer the model to another site. In addition, different models
can be stored and trained individually for different area types. In
the current state the models cannot provide an automatized
correction of DInSAR time series. Nonetheless, this work is a
step towards the automation of error prediction and correction in
geospatial data.
3.2 Integration of elevation data across various scales and
reference systems
The availability of multiple diverse elevation datasets – with
different coverage, vertical datum, horizontal datum and
resolution, accuracies, and consideration of waterbodies is very
problematic for end-users who are often required to choose
amongst multiple elevation datasets for a study area. This task
usually entails heavy data pre-processing, while it can lead to
biased and/or inconsistent decisions, with a negative impact to
the corresponding project outcomes.
In Canada, terrain datasets released by Natural Resources Canada
(NRCan) mainly include the Canadian Digital Elevation Model
(CDEM; NRCan, 2013) and the High Resolution Digital
Elevation Model (HRDEM; NRCan, 2019). The CDEM
collection is part of the NRCan’s altimetry system. The coverage
and resolution of the CDEM mosaic varies according to latitude
and extent of the study area. With reference to the NAD83 CSRS
datum, the mosaic covers the whole country at resolutions that
range from 0.75 to 12 arcsec along the latitudes. The elevation
values are expressed in integer meters referenced to the Canadian
Geodetic Vertical Datum of 1928 (CGVD28) and can be either
ground or reflective surface elevations. As a part of the
CanElevation Series, the HRDEM largely improves the accuracy
and spatial resolution of Canadian terrain data. The HRDEM
collection consists of high-resolution DEMs derived from
LiDAR and remote sensing imagery produced by separate
projects. HRDEM includes a Digital Terrain Model (DTM), a
Digital Surface Model (DSM) and other derived data (e.g., slope,
aspect, shaded relief, color-relief, and color-shaded relief maps).
HRDEM is available only over the corresponding projects
footprints. In the southern part of the country, HRDEM
collections include DTM and DSM datasets at a 1m or 2m
resolution and projected to the UTM NAD83 (CSRS) coordinate
system and the corresponding zones. In the northern part of the
country, HRDEM collections include DSM datasets at a 2m
resolution projected in the Polar Stereographic North coordinate
system referenced to WGS84 horizontal datum or UTM NAD83
(CSRS) coordinate system. HRDEM elevation values are
referenced to the Canadian Geodetic Vertical Datum of 2013
(CGVD2013), which is now the reference standard for heights
across Canada.
In Li et al. (2021) we have recently explored the adoption of
Discrete Global Grid Systems (DGGS; OGC, 2017) as an
integration platform for Canadian terrain datasets to improve the
coverage and elevation data quality in various study areas across
the nation (Figure 3). Various algorithms were introduced to
integrate the CDEM and HRDEM by direct quantization at
various granularities and to aggregate the modelled elevations by
mean, maximum, and minimum statistics across the resolution
levels to meet the needs of different applications. For example,
the minimum elevation helps to determine stream channel areas,
while the maximum elevation is useful for calculating the height
of vertical obstructions (Danielson and Gesch, 2011). This study
set the stage for a national elevation service across various scales
for Canada.
Figure 3: Representation of a. world landmass, b. Canada, c.
Ontario, and d. study areas in the ISEA3H DGGS (adopted by Li
et al., 2021).
3.3 Data management and model integration for CNN-
based image analysis
In the following use case precipitation data of Hyderabad, India,
from 1991 – 2011 is used as Landsat 5 multi-spectral data. For
categorization self-organizing neural networks (Babu, 1997), and
especially Convolutional Neural Networks (CNNs) are used as
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
43

an appropriate method for image classification of satellite data
(Kanellopoulos and Wilkinson, 1997). To produce a straight-
forward joint workflow of data analysis and data management,
and to avoid unnecessary data loading, in the following use case
it is outlined how CNN-based data analysis and raster-based data
and model management can work together using TensorFlow®
data analysis tool and rasdaman® array-DBMS, respectively.
The dataset consists of 94 images of 7 spectral bands of 1710 x
1750 with a resolution of 30m from the years 1991 to 2011
(excluding 2002-2003). Using rasdaman´s filter functions for
irregular time intervals, all data have been considered presenting
less than 10 percent clouds in each image (Yang, 2022). Using
rasdaman, spatio-temporal database queries have been used in
two different ways: first, attributes of a coverage such as bands
of the 7-band hyperspectral Landsat 5 dataset have been selected
via OGC´s Web Coverage Service. Secondly, OGC´s Web
Coverage Processing Service have been used to execute raster
operations such as special filter operations. For example, a 3D
data cube with the three dimensions “band”, “time step” and “2D
region” has been created as a result of a spatio-temporal database
query (see figure 4). Here bands 2-7 are selected in the region of
Hyderabad from 1991 to 2011, presenting one time step in figure
4.
Figure 4: Result of a spatio-temporal database query represented
in the three dimensions “band”, “time step” and “2D region” of a
3D data cube (from: Yang, 2022).
A database query filter operation has been used to generate False
Color Composite (FCC) images facilitating the detection of
vegetation changes (Yang, 2022). The user controls such queries
via rasql, rasdaman´s raster query language. An example of such
a database result showing two FCC images is presented in figure
5.
Figure 5: Result of a spatio-temporal database query computing
False Color Composite images of the precipitation data in the
region of Hyderabad, March and October 2011 (from: Yang,
2022).
To achieve reproducible CNN models during data analysis, the
different variants of CNNs produced during the training process
of the data have to be stored together with the data in the
geospatial database for long term use. From a technical point of
view, a direct connection is established between the data analysis
tool (such as TensorFlow®) and the array-based DBMS (such as
rasdaman®). This means that TensorFlow® then has direct
access to the data and models stored in rasdaman®. As an
alternative, graph-based DBMS may be used to store the CNNs.
In both cases, a DBMS-supported image analysis workflow then
consists of the following steps: Data preparation and quality
checking => Data management and integration => Model
management and integration => Data analysis with continuous
access to data and model management. This workflow not only
saves time, but supports the continuous management of data and
CNN models during data analysis.
3.4 Advanced data analysis: CNN-based semantic
segmentation of swimming pools
As application of advanced data analysis for remote sensing data
we refer to the identification and delineation of man-made
structures in aerial or satellite imagery. In this use case, aerial
images from Baden-Württemberg, Germany, have been used for
semantic segmentation of swimming pools and subsequent
comparison with publicly available information from
OpenStreetMap.
In general, such an application could enable the estimation of
residential property values or monitoring of swimming pool
maintenance and safety. However, for the development of robust
deep learning models, a variety of requirements and challenges
must be addressed:
- Large image sizes: Images need to be resized, cropped, or
tiled to be usable by regular deep learning models. In this
case, the images were 5000 x 5000 pixels.
- Small objects: In this case, a typical swimming pool mask
has the size of approx. 100-500 pixels within a total pixel
quantity of 25,000,000 pixels per image.
- Unbalanced dataset: Even after filtering out images without
any swimming pools, the vast number of pixels still do not
belong to the desired class.
- Geo-referencing: If the results of the semantic segmentation
are supposed to be used for an application where the
geospatial context is needed, the geo-referencing of the
original images must be preserved.
Figure 6: Top left: Original image. Top: right: Corresponding
ground truth mask. Bottom left: Random crop of original image.
Bottom right: Corresponding ground truth mask.
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
44

As a first step, a dataset of 155 aerial images with manually
created ground truth masks was developed. As shown in Figure
6, the masks classify each pixel as either belonging to a
swimming pool or background pixels. This enables the training
of a deep learning model for binary segmentation to detect
swimming pools in aerial images. At the same time, the
coordinates of the corner points of the image are saved to
preserve the geo-reference.
As a second step, a U-Net (Ronneberger et al., 2015) was trained
using binary cross entropy loss and SGD optimizer (Robbins and
Monro, 1951). The U-Net architecture consists of a contracting
path (encoder) and an expanding path (decoder) with skip
connections between them. The model was trained to perform
binary segmentation on the before-mentioned dataset, where
each pixel in the image is classified as either swimming pool or
background. The binary cross entropy loss function measures the
difference distance between the prediction and the ground truth,
whereas the SGD optimizer minimizes the loss function during
training by adjusting the weights of the U-Net model.
Figure 7: a) original image, b) ground truth mask, c) prediction -
Preliminary segmentation results of the trained U-Net on new
aerial images.
Lastly, the trained model can be used for binary segmentation of
swimming pools in new aerial images. Figure 7 shows two
examples of preliminary results. The first example demonstrates
the effectiveness of the trained U-Net model for identifying
swimming pools in aerial images as it successfully segmented the
swimming pool (highlighted in yellow) from the surrounding are
with high accuracy. However, the second example highlights the
difficulty of the task as the model misclassified a pond as a
swimming pool. This misclassification could be due to the similar
shape and color of the pond and swimming pools, as well as the
presence of another swimming pool and residential buildings
near the pond.
The segmentation results can finally be aligned with the geo-
referenced original images which enables the comparison with
publicly available information from OpenStreetMap. In future
work, we plan to significantly improve the segmentation results
through incorporating more training data and refining the training
process to thereafter identify missing swimming pools in
OpenStreetMap. By doing so, we might be able to contribute to
the improvement publicly available geospatial data in general.
4. EMERGING RESEARCH QUESTIONS
DERIVED FROM THE USE CASES
The use cases have shown that there is a big need to improve the
workflows during data preparation, data quality checking, data
and model integration, and data analysis in remote sensing
scenarios. We now derive general research questions triggered by
the use cases.
4.1 Improving data preparation and data quality checking
To improve data preparation and data quality checking of big
data, first geospatial data cleansing should be applied. Adjusting
the workflow of Nelder and Wedderburn (1972) to remote
sensing scenarios and following Mazroob et al. (2020), geospatial
data cleansing should retain the following rules:
- Remove unwanted observations as irrelevant data: Outliers
can negatively distort data models, in particular linear
regression models in comparison with decision trees.
Therefore, removing outliers will help model performance.
Irrelevant data usually includes duplicate records, missing
or incorrect information and poorly formatted data sets.
- Predict missing values - categorical or numerical, because
data analysis algorithms mainly do not accept missing data:
To manage missing data for categorical features, a class is
added and this handles the case of no missing values. As for
missing numeric data, the observation should be indicated
and replaced with a “0” to satisfy the model’s algorithm
requirement of no missing values enabling it to predict the
best estimate for missing values rather than just the mean
(Lee and Nelder, 2002).
- Remove unwanted data including duplicate, redundant and
irrelevant data.
The correctness of acquired remote sensing data has to be
measured by different data quality metrics. Depending on the
application of the data, various quality measures can be defined.
However, most important is accuracy of the remote sensing data.
Accuracy defines how close the measured values are to the true
values. Applied to remote sensing data, accuracy can be
measured in geometric, radiometric or temporal context.
The remote sensing data lifecycle has strong relation to data
quality dimensions and their adequate metrics. The definition of
accuracy dimensions related to data preparation are as follows:
- Geometric precision: instability of the observation.
- Spatial precision: correctness of the spatial representation
of the feature.
- Radiometric precision/ stability: correctness of the
quantization.
- Spectral precision: correctness of the boundaries of the
spectral bands
- Temporal precision: goodness of the data capture date and
time.
- Spatial accuracy: accuracy of position of features in
relation to Earth.
- Radiometric accuracy: correctness of the intensity values
(radiance uncertainty).
- Spectral accuracy: correctness of the sensor’s imaging
capability in the given channel.
- Temporal accuracy/validity: quality of the remote sensing
product in time (how long does it store good information).
Besides accuracy, completeness, redundancy, readability,
accessibility, and consistency are notable data quality
dimensions in remote sensing.
Data quality requirements are based on data user specification.
Inspection is carried out to evaluate whether data meets
specification. Certain ISO standards exist for data quality check
to examine fit for their intended purpose: Smart city operations
are considered as a major consumer of satellite data and small-
scale imagery, that is beside using high-resolution images to
detail all cadastral level information and perform data
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
45

management. In fact, those smart city operations tend to insure
the completeness of city smartness, e.g. governance; safety
security; mobility; sustainability; livability; resilience, etc.
Similar to tabular and vector-based data, raster-based remotely
sensed data and its processed information including the
formulated knowledge resulting from analysis and solution
scenarios, that are acquired via satellite sensors/images play a
major role in feeding those smart city operations. It is hence
certain that raster or vector (derived from raster) data
preparation, data quality checks, model integration, and data
analysis directly or indirectly affect the smartness of cities, e.g.
producing cloud-free and haze-free satellite images leads to more
proper image interpretations and classifications allowing by that
to feed operations such as managing forests in a smarter and safer
way. Thus, automating actions such as accelerating the proper
selection of the most effective image is an important data
preparation and quality checking-based activity that in future can
be performed at the level of database operations, reducing by that
the expenditure in skilled operator time.
In fact, it would fare to mention that international open satellite
imagery providers, as well as commercial ones, have been
working extensively to ease access and encourage the usage of
provided data and information, by partially performing some of
the data preparation and data quality checking that require skilled
labour. In other words, they not only provide corrected images
based on user selection criteria but also produce the so-called
Analysis Ready Data (ARD) which overcome many initial steps
of data preparation and quality checking. The preparation acts for
the satellite data providers can include the following among
others:
- Radiometric corrections can include calibration
insuring the consistency between sensors and
overtime, and also atmospheric corrections of images.
- Geometric corrections such a geo-referencing, ortho-
rectifications, colour balancing, mosaicking, etc.
- Metadata – including usable data masks (UDM)
- Time series of the same region.
Further, perhaps some unsupervised pre-classifications and
initial segmentation can also be categorized under preparation,
the parameters of which can be predicted based on provided
usage and application area information. Organizations such as
Copernicus (open access Hub) and Nasa, also other platforms
such as ARCGIS online from ESRI, Google Earth Engine,
Sen2Cube from Austria, Earth Observation browser, Earth
Observation Compass, USGS, CEOS, and many others even take
it to a further level, where some provide Analysis ready Data
(ARD). In fact, some organizations even construct the so-called
data cubes, providing users ready prepared blocks that can be
combined with pre-defined tools to prepare many ready maps that
are based on basic analysis functionalities, workflows and
formulas. ARD maps cover atmospheric, marine and land
different themes.
All produced products shall be subject to uncertainty, which
entitle the necessity of quality checking. Other than checking the
internal structure of the data such as duplications and missing
values, error values etc., the data has to be validated and assured
to reflect the respective existing reality, and that can be done by
ground truth or by comparing the data to more trusted solid data
sets.
4.2 Improving data and model management for CNN-based
data analysis
Hitherto data analysts have to pass through a long and time-
consuming process chain across several software systems to
spatially or/and temporally select particular regions and time
intervals out of big satellite image data. To automate and shorten
the process of data selection, a geospatial database should support
this process providing spatial, temporal, and spatio-temporal
operations on satellite image data such as (see also Mazroob et
al., 2020):
- Automatically checking geometric, topological, and
temporal constraints on satellite image data to detect data
errors.
- Seamlessly selecting arbitrary, a-priori not defined tiles
from one given scene (defining one time step) of a satellite
image.
- Selecting the same tile at different scenes (versions of the
tile).
- Detecting the differences of values in pixel attributes of two
scenes (change detection).
- Overlaying data from different sources and domains for the
same region (e.g. SENTINEL and weather data).
“Seamlessly selecting a tile” means that the data have to be
selected spatially independent of à priori fixed partitions.
Furthermore, the temporal selection of the same region within a
time interval has to be supported by a spatio-temporal database
operation. The same is true to compute the differences between
two images of the same region generated at two different time
steps. Note that the overlay of two images from different data
sources has to be applied carefully: the generation of “integrated
models” is a sophisticated task and has to consider a variety of
geometric, topological, temporal, and semantical constraints.
Picking up the example of our first use case (section 3.1), the
automatic checking of phase errors in interferometric synthetic
aperture (InSAR) radar data could be executed by setting data
constraints such as “the phase must not be greater than 2π” up to
complex algorithms implementing unwrapping operations. The
advantage of this database-supported approach is that the data has
no more to be loaded from a data platform and later laboriously
spatially or/and temporally selected as wanted, but the selection
of the data is immediately done in the geospatial data
management system.
Besides satellite image data, also the models used for data
analysis such as Convolutional Neural Networks (CNNs) should
be managed in a geospatial database management system. The
persistent storage and retrieval of these models will significantly
improve the reproducibility of image analysis such as CNN-
based image analysis. Therefore, not only the “original models”
should be stored in the database, but also optimized versions of
the (learning) models. To provide the models by a database is
even more important, if data from different sources are involved.
4.3 Improving data integration
The rapid growth in the number of ground, airborne, and satellite
sensors has led to an unprecedented increase in the amount,
variety, and rate of collection of remote sensing data. The process
of integrating huge volumes of heterogeneous geospatial data
using traditional data models on the geographic grid is lossy,
computationally expensive, and time-consuming (OGC, 2017).
Recently, Discrete Global Grid Systems (DGGS) have been
proposed as an alternative geospatial reference framework that
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
46

can facilitate the fusion of multi-source remote sensing data
(Goodchild, 2018; OGC, 2017; Gibb, 2021).
A DGGS applies a partitioning approach to divide the Earth’s
surface into a group of uniform cells at various levels of
resolutions. Its hierarchical structure can effectively support the
sampling, storage, modeling, processing, analysis, integration,
and visualization of voluminous and heterogeneous remote
sensing data.
The adoption of a DGGS in modelling and management of
remote sensing data has two objectives. The first objective is to
facilitate the integration of heterogeneous remote sensing
datasets for a study area using a DGGS. Figure 8 summarizes the
DGGS parameters, which are chosen to provide an optimal
hierarchical tessellation of Canada’s landmass.
Figure 8: Example DGGS configuration parameters for Canada’s
landmass.
The integration of remote sensing datasets includes the following
three tasks:
1. Data preparation: remote sensing data will be converted into
a common horizontal and vertical datum.
2. Data quantization: remote sensing data will be assigned to
DGGS cells at various resolutions based on their horizontal
and vertical accuracy.
3. Quality Control: Ground control points will be used to
validate the quantization of the remote sensing dataset by
calculating and comparing the post-DGGS and pre-DGGS
statistical errors.
The second objective is to support the geo-processing of the
‘analysis-ready’ remote sensing data modelled into a DGGS.
Various generic and application specific analytical operations
need to be developed to allow the extraction of consistent derived
data from the DGGS. These operations need to be extensively
validated through experimental testing for various study areas
based on acquired knowledge from past works carried out by both
domain experts and data analysts to guarantee an efficient
toolbox for reliable decision-making processes.
5. CONCLUSIONS AND OUTLOOK
In this paper we presented use cases showing the emerging need
of improving data quality and the management of data and
models in a geospatial database management system before and
during data analysis. Different new processes of data preparation
and quality checking, integration of data across different scales
and references systems as well as efficient data and model
management showed that data analysis in satellite remote sensing
scenarios has to be supported by various data quality and data
management techniques. Based on the use cases we discussed
emerging research questions concerning data preparation and
data quality checking, data and model management, and data
integration, cf. Figure 9.
Figure 9: Overview of a typical remote sensing data processing
workflow.
Data cleansing during data preparation, accuracy dimensions, the
applicability on smart city operations, spatio-temporal database
operations, the integration of Discrete Global Grid Systems and
learning models in a geospatial database have been presented as
possible solutions for an improved data analysis in remote
sensing scenarios. In our future work we will embed integrated
data management and data analysis architectures into existing
remote sensing and geospatial applications. Finally, we intend to
apply some of the introduced data preparation and data
management techniques to support earth observation scenarios in
the United Arab Emirates, with special emphasis on applications
such as smart city management and intelligent traffic solutions.
ACKNOWLEDGEMENTS
We thank the State Office for Geoinformation and Rural
Development of Baden-Württemberg, Germany for their support.
REFERENCES
Amirian P., Basiri A., Winstanley A., 2014. Evaluation of Data
Management Systems for Geospatial Big Data. In: Murgante B.
et al. (eds) Computational Science and Its Applications. ICCSA
2014. Lecture Notes in Computer Science, vol 8583. Springer.
Babu, G.P., 1997. Self-organizing neural networks for spatial
data. Patt. Recogn. Lett. 18, pp. 133–142.
Baumann, P., 2010. The OGC Web Coverage Processing Service
(WCPS) Standard, Geoinformatica, 14(4)2010, pp. 447-479.
Baumann, P., Furtado, P., Ritsch, R., Widmann, N., 1997. The
RasDaMan Approach to Multidimensional Database
Management. Proc. 12th Annual Symposium on Applied
Computing (SAC'97), San Jose/USA, February 28 - March 2,
1997.
Baumann, P., Mazzetti, P., Ungar, J., Barbera, R., Barboni, D.,
Beccati, A., Bigagli, L., Boldrini, E., Bruno, R., Calanducci, A.,
Campalani, P., Clement, O., Dumitru, A., Grant, M., Herzig, P.,
Kakaletris, G., Laxton, J., Koltsida, P., Lipskoch, K., Mahdiraji,
A.R., Mantovani, S., Merticariu, V., Messina, A., Misev, D.,
Natali, S., Nativi, S., Oosthoek, J., Passmore, J., Pappalardo, M.,
Rossi, A.P., Rundo, F., Sen, M., Sorbera, V., Sullivan, D.,
Torrisi, M., Trovato, L., Veratelli, M.G., Wagner, S., 2016. Big
Data Analytics for Earth Sciences: the EarthServer Approach.
International Journal of Digital Earth, 9(1), 2016, pp. 3 – 29.
Baumann, P., Misev, D., Merticariu, V., Pham Huu, B., 2018.
Datacubes: Towards Space/Time Analysis-Ready Data.. In: J.
Doellner, M. Jobst, P. Schmitz (eds.): Service Oriented Mapping
- Changing Paradigm in Map Production and Geoinformation
Management, Springer Lecture Notes in Geoinformation and
Cartography. pp. 269-299.
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
47

Bishop, C., 1995. Neural Networks for Pattern Recognition,
Oxford University Press, Inc. New York, NY, USA, 482p.
Breiman, L., 2001. Random forests. Machine learning, 45(1), pp.
5-32.
Breunig, M., Kuper, P.V., Butwilowski, E., Thomsen, A., Jahn,
M., Dittrich, A., Al-Doori, M., Golovko, D., Menninghaus, M.,
2016. The story of DB4GeO – A service-based geo-database
architecture to support multi-dimensional data analysis and
visualization. ISPRS Journal of Photogrammetry and Remote
Sensing, 117, pp. 187-205.
Chen, M., Mao, S., & Liu, Y., 2014. Big data: A survey. Mobile
networks and applications, 19(2), pp. 171-209. Copernicus. Open
Access Hub. https://scihub.copernicus.eu/ (20 May 2019).
Cheng T., Haworth J., Anbaroglu B., Tanaksaranond G., Wang
J., 2014. Spatiotemporal Data Mining. In: Fischer M., Nijkamp
P. (eds) Handbook of Regional Science. Springer, Berlin,
Heidelberg, pp. 1173-1193.
Danielson, J. J. and Gesch, D. B., 2011. Global multi-resolution
terrain elevation data 2010 (GMTED2010). Reston, Virginia,
USA: Earth Resources Observation and Science (EROS) Center,
U.S. Geological Survey, Open-File Report 2011-1073. Available
from: https://pubs.usgs.gov/of/2011/1073/pdf/of2011-1073.pdf
[Accessed 1 May 2020].
Dean, J. and Ghemawat, S., 2008. MapReduce: simplified data
processing on large clusters. Communications of the
ACM, 51(1), pp.107-113.
ESA. SENTINEL-1. https://sentinel.esa.int/web/sentinel/-
missions/sentinel-1 (02 March 2023).
Gibb, R. (Ed.) 2021. Topic 21 - Discrete Global Grid Systems -
Part 1 Core Reference system and Operations and Equal Area
Earth Reference System. Open Geospatial Consortium.
https://docs.ogc.org/as/20-040r3/20-040r3.html
Goodchild, M. F. (2018) Reimagining the history of GIS. Annals
of GIS, 24(1), 1–8.
Hahn, K., Reiner, B., Höfling, G., Baumann, P., 2002. Parallel
Query Support for Multidimensional Data: Inter-object
Parallelism. 13th International Conference on Database and
Expert Systems Applications (DEXA), September 2-6, 2002, Aix
en Provence, France.
Heck, A. PhD Thesis, Geodetic Institute, Faculty of Civil
Engineering, Environmental and Geo-Sciences, Karlsruhe
Institute of Technology (KIT), 2019.
Hooper, A., 2010. A statistical-cost approach to unwrapping the
phase of InSAR time series. In Proceedings of the International
Workshop on ERS SAR Interferometry, Frascati, Italy (Vol. 30).
Hooper, A., Segall, P., & Zebker, H., 2007. Persistent scatterer
interferometric synthetic aperture radar for crustal deformation
analysis, with application to Volcán Alcedo, Galápagos. Journal
of Geophysical Research: Solid Earth, 112(B7).
Hooper, A., Bekaert, D., Spaans, K., and Arıkan, M., 2012.
Recent advances in SAR interferometry time series analysis for
measuring crustal deformation. Tectonophysics, 514-517:1–13,
Jan. 2012. ISSN 00401951. doi: 10.1016/j.tecto.2011.10.013.
Hu, F., Xu, M., Yang, J., Liang, Y., Cui, K., Little, M.M., Lynnes,
C.S., Duffy, D.Q. and Yang, C., 2018. Evaluating the open source
data containers for handling big geospatial raster data. ISPRS
International Journal of Geo-Information, 7(4), p.144.
Kanellopoulos, I.; Wilkinson, G.G., 1997. Strategies and best
practice for neural network image classification. Int. J. Remote
Sens., 18, pp. 711–725.
Klein, L.J., Marianno, F.J., Albrecht, C.M., Freitag, M., Lu, S.,
Hinds, N., Shao, X., Rodriguez, S.B. and Hamann, H.F., 2015,
October. PAIRS: A scalable geo-spatial data analytics platform.
In 2015 IEEE Internat. Conference on Big Data, pp. 1290-1298.
Laney, D., 2001. 3D data management: Controlling data volume,
velocity and variety. META Group Research Note, 6(70).
Lary, D.J., 2010. Artificial Intelligence in Geoscience and
Remote Sensing, Geoscience and Remote Sensing, IntechOpen,
24p., doi: 10.5772/9104.
Lary, D.J., Alavi, A.H., Gandomi, A.H., Walker, A.L., 2016.
Machine learning in geosciences and remote sensing, Geoscience
Frontiers, 7(1), pp. 3-10, doi.org/10.1016/j.gsf.2015.07.003.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning, Nature
Volume 521, pp. 436–444, doi.org/10.1038/nature14539.
Lee, J.-G., Kang, M., 2015. Geospatial Big Data: Challenges and
Opportunities, Big Data Research 2 (2015), pp. 74-81.
Lenka, R. K., Barik, R. K., Gupta, N., Ali, S. M. ; Rath, A.,
Dubey, H., 2017. Comparative Analysis of SpatialHadoop and
GeoSpark for Geospatial Big Data Analytics, 6p., Arxiv ID:
1612.07433.
Leskovec, J. and Sosič, R., 2016. SNAP: A general-purpose
network analysis and graph-mining library. ACM Transactions
on Intelligent Systems and Technology (TIST), 8(1):1, 2016.
Li, S., Dragicevic, S., Castro, F. A., Sester, M., Winter, S.,
Coltekin, A., Pettit, C., Jiang, B., Haworth, J., Stein, A., Cheng,
T., 2016. Geospatial big data handling theory and methods: A
review and research challenges, ISPRS Journal of
Photogrammetry and Remote Sensing, May 2016, Vol.115,
pp.119-133.
Li., M., McGrath, H., and Stefanakis, E., 2021 Integration of
Heterogeneous Terrain Data into Discrete Global Grid
Systems. Cartography and Geographic Information Science
(CaGIS). Taylor& Francis. https://doi.org/10.1080/15230406.
2021.1966648
Lombacher, J., Hahn, M., Dickmann, J., Wöhler, C., 2016.
Potential of radar for static object classification using deep
learning methods, IEEE MTT-S International Conference on
Microwaves for Intelligent Mobility (ICMIM), San Diego, CA,
2016, pp. 1-4, doi: 10.1109/ICMIM.2016.7533931.
Mathieu, P.-P., Aubrecht, C., 2018. Earth Observation Open
Science and Innovation, ISSI Scientific Report Series, ISBN:
9783319389677.
Mazroob, N., Breunig, M., Al-Doori, M., Heck, A., Kuper, P.,
Kutterer, H., 2020. Towards Intelligent Geodatabase Support for
Earth System Observation: Improving the Preparation and
Analysis of Big Spatio-Temporal Raster Data. Int. Arch.
Photogramm. Remote Sensing. Spatial Inf. Sci., XLIII-B4-2020,
Commission IV, WG IV/7, pp. 485-492,
https://doi.org/10.5194/isprs-archives-XLIII-B4-2020-485-
2020.
Nelder,, J.A., Wedderburn, R.W.M., 1972. Generalized Linear
Models. Journal of the Royal Statistical Society, Series A, 19(3),
pp. 92-100.
NRCan, 2013. Canadian Digital Elevation Model product
specifications [online]. Available from:
https://ftp.maps.canada.ca/pub/nrcan_rncan/elevation/cdem_mn
ec/doc/CDEM_product_specs.pdf [Accessed 1 March 2021].
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
48

NRCan, 2019. High Resolution Digital Elevation Model
(HRDEM) – CanElevation Series – product specifications
[online]. Available from: https://ftp.maps.canada.ca/pub/
elevation/dem_mne/highresolution_hauteresolution/HRDEM_P
roduct_Specification.pdf [Accessed 1 March 2021].
OGC, 2017. Topic 21: discrete global grid system abstract
specification [online]. Available from:
http://www.opengis.net/doc/AS/dggs/1.0 [Accessed 15
November 2019].
Ouyang, L., Huang, J., Wu, X. and Yu, B., 2013. Parallel access
optimization technique for geographic raster data. Geo-
Informatics in Resource Management and Sustainable
Ecosystem, pp. 533-542.
Qin, Q., Gillies, R.R., Lu, R., Chen, S., 2004. An Integration of
Wavelet Analysis and Neural Networks in Synthetic Aperture
Radar Image Classification, Int. Arch. Photogramm. Remote
Sens. Spatial Inf. Sci., XXXV-B2, pp. 181-186.
Robbins, H., Monro, S., 1951. A stochastic approximation
method. The annals of mathematical statistics (1951): pp. 400-
407.
Ronneberger, O., Fischer, P., Brox, Th., 2015. U-net:
Convolutional networks for biomedical image
segmentation. Medical Image Computing and Computer-
Assisted Intervention–MICCAI 2015: 18th International
Conference, Munich, Germany, October 5-9, 2015, Proceedings,
Part III 18. Springer International Publishing, 2015.
Rumelhart, D.E.; Hinton, G.E.; Williams, R.J., 1986. Parallel
Distributed Processing. MIT Press: Cambridge, MA, USA.
Spatial Hadoop, 2023. SpatialHadoop – a MapReduce
Framework for Spatial Data, http://spatialhadoop.cs.umn.edu/
(last accessed: 02.03.2023).
Taylor, R. C., 2010. An overview of the
Hadoop/MapReduce/HBase framework and its current
applications in bioinformatics. In BMC bioinformatics, BioMed
Central.
Werner, M., Chiang, Y.-Y., Handbook of Big Geospatial Data,
Springer, 2021, https://doi.org/10.1007/978-3-030-55462-0
Xiaoqiang, Y., & Yuejin, D., 2010. Exploration of cloud
computing technologies for geographic information services. In
Geoinformatics, 18th International Conference, pp. 1-5.
Yang, X., 2022. Data Management and Model Integration for
CNN-based Image Analysis for Remote Sensing Data, Master
Thesis, Geodetic Institute, Faculty for Civil Engineering, Geo-
and Environmental Sciences, KIT, 73pp.
Yu, H., Lan, Y., Yuan, Z., Xu, J., and Lee, H., 2019. Phase
Unwrapping in InSAR : A Review. IEEE Geoscience and
Remote Sensing Magazine, 7(1):40–58, Mar. 2019. ISSN 2168-
6831, 2473-2397, 2373-7468. doi:
10.1109/MGRS.2018.2873644
Zhong, Y., Sun, S., Liao, H., Zhao, Y. and Fang, J., 2011. A novel
method to manage very large raster data on distributed key-value
storage system. 19th International Conference on
Geoinformatics (2011), pp. 1-6.
Zhu X., Tuia D., Mou L., Xia G., Zhang L., Xu F., Fraundorfer
F., 2017. Deep Learning in Remote Sensing: A Comprehensive
Review and List of Resources, IEEE Geoscience and Remote
Sensing Magazine, 5(4), 8-36
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume X-1/W1-2023
ISPRS Geospatial Week 2023, 2–7 September 2023, Cairo, Egypt
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.
https://doi.org/10.5194/isprs-annals-X-1-W1-2023-41-2023 | © Author(s) 2023. CC BY 4.0 License.
49