Content uploaded by Piyush Yadav
Author content
All content in this area was uploaded by Piyush Yadav on Jul 08, 2020
Content may be subject to copyright.
Proceedings of the Academic Track, FOSS4G Europe 2020
Human Assisted Artificial Intelligence Based Technique
to Create Natural Features for OpenStreetMap
Piyush Yadav
Dipto Sarkar
Shailesh Deshpande
Edward Curry
NUI Galway
Ireland
Carleton University
Canada
TCS Research
India
NUI Galway
Ireland
Introduction
OpenStreetMap (OSM) is arguably the largest crowdsourced geographic databases with more
than one million contributors [1]. The ongoing contributions make OSM an ever-evolving
spatial dataset with the improving quantity, quality, and coverage of data across all types of
map features. While OSM has stood up to various tests of data quality and accuracy with
regards to a variety of manmade features (e.g. building, roads), data completeness of high-
level natural feature classes like vegetation (grass, forest), impervious surface (industrial
regions) and water (lakes, river, delta) is relatively sparse. This is partly because OSM data
creation tools (e.g. iD) provide an intuitive interface to digitize smaller features, such as
building and road segments, but is cumbersome for natural features (e.g. delta, scree,
geological regions) which are quite complex and are spread over larger regions.
Recently, Facebook presented RapiD editor
1
which uses Artificial Intelligence (AI) techniques
to generate features for faster map editing. RapiD autodetects roads features from high-
resolution satellite imagery which are then curated by users to create the final dataset. The
RapiD program has certain limitations: for example, discovering linear features like road is
easier as compared to natural elements (such as delta and forest) which are uneven [2].
Secondly, most of the high-resolution imagery is still not freely available which contradicts the
preamble of the open-source community. In this work, we propose an AI-based technique
using freely available satellite images like Landsat and Sentinel to create natural features over
OSM in congruence with human editors acting as initiators and validators. The method is
based on Interactive Machine Learning [3] technique where human inputs are coupled with
the machine to solve complex problems efficiently as compare to pure autonomous process.
We use a bottom-up approach where a machine learning (ML) pipeline in loop with editors is
used to extract classes using spectral signatures of images and later convert them to editable
features to create natural features.
Proposed Approach
Figure 1. High-level architecture for the proposed approach
As per Fig.1, our proposed method is divided into two-part:
1
https://mapwith.ai/
Raster Processing
Users/Editors
Training Data
Selection
Satellite
Imagery Machine Learning
Models
CPU GPU Classified
Image
Misclassified
Image
Update Training Samples
Train Models
Pixel to Point
Registration
Spectral Grouping
Based Boundary
Creation
Natural Features
(Lines/Polygons)
Edit Error-Prone
Boundaries
PostGIS Database
Vector Processing
O
S
M
Low-level Raster Processing: To start the process, the user labels training samples
belonging to different classes by selecting pixels from the imagery. It is assumed that the user
has some knowledge about the natural features in the area. The number of macro (e.g. water,
vegetation) or micro (e.g. lake and grass) classes can be varied depending on the user’s
expertise of the region. The ML models are then trained using the training data. The proposed
system consists of a suite of different ML models like Support Vector Machine (SVM), Nearest
Neighbour (NN) and Convolutional Neural Networks (CNN) running on GPU instances for
efficient processing. The models predict the class of every pixel and classify the image. The
misclassified regions can again be selected by the user to improve model accuracy by adding
new training samples iteratively. Finally, the image registration is performed over a classified
image so that each pixel spatial location is known.
Top-level vector processing: At top-level image primitives and complex objects are created
using the proposed grouping method.
Image Primitives: A spatial entity can be represented using geometry-based features like
points, lines, and polygons. The point is a zero-dimensional image primitive located in
cartesian coordinate while a line represents a one-dimensional object defined by a list of
points. Similarly, a polygon is a two-dimensional bounded region created from a list of lines.
Organization of Image Primitives: On a high-level abstraction, an image can be represented
as a collection of points representing each pixel in the image. The classified image pixels are
converted to points where each point represents a spatial location with a known class (e.g.
water). A grouping-based algorithm is devised based on class similarity and vicinity to combine
points to line and lines to polygons. The grouping occurs in the following order: 1) if adjacent
points have same class then group them, 2) the first level grouping is then converted to
boundary edges (line) and closed polygons.
The intermediate image primitives are stored in a PostGIS database where elements can be
fetched using the query. The editors can manually refine the error-prone automated
boundaries which can then be exported to OSM after validation.
Results
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure 2. Wax Lake Delta (a) Google Map View (b) OSM view (c) Landsat 8 Image (d) Pont representation of water
pixels (e) Waterbody Boundary (f) Polygon for different water streams (g) Waterbody medial skeleton
The proposed approach is implemented in Python 3. For vector processing, PostGIS database
is used to store the intermediate results using pyscopg2 adapter. The area of study is Wax
Lake River Delta in the United States to identify complex delta boundaries [4]. Fig 2 (a) and
(b) shows the comparison between Google Maps and OSM of the study area, respectively,
where OSM is missing the delta boundaries. The Landsat 8 image of the selected region (Fig.
2 (c)) is classified using SVM classifier where soil, water and vegetation pixels are classified
based on their spectral signatures. The misclassified region training samples were again
collected to improve classification accuracy. Fig. 2 (d) shows the water pixels as vector points
which are later connected to form boundary (Fig.2 (e)), polygons (Fig. 2 (f)) and lines (Fig. 2
(g)) using class-based grouping algorithms. QGIS is used as a front-end editing tool for
detected features as it consists of robust vector processing suites to edit the features. The
final curated features are then can be exported to OSM from bottom to top levels, i.e. point
and line and polygon.
Conclusion
In this work, we propose a human assisted-ML framework to contribute natural features to
OSM. This tool works in conjunction with the open-source GIS platform QGIS and provides
an intuitive interface to create natural features spread over a large spatial extent. The
contributions made by the users using this tool will help OSM move closer to being a complete
spatial data repository. We plan to release this tool as a plugin for QGIS in the coming months.
References
[1] 1 million map contributors! 2018.https://blog.openstreetmap.org/2018/03/18/1-million-
map-contributors/.
[2] Mooney, P., Corcoran, P. and Winstanley, A., 2010, September. A study of data
representation of natural features in openstreetmap. In Proceedings of GIScience (Vol.
150).
[3] Amershi, S., Cakmak, M., Knox, W.B. and Kulesza, T. 2014. Power to the People: The
Role of Humans in Interactive Machine Learning. AI Magazine. 35, 4 (2014).
[4] Deshpande, S., Sowmya, A., Yadav, P., Ladha, S., Verma, P., Vaiapury, K., Gubbi, J.
and Balamuralidhar, P., 2017, April. CogVis: attention-driven cognitive architecture for
visual change detection. In Proceedings of the Symposium on Applied Computing (pp.
151-154).
