PaperlayerView exact revision on arXiv

Source-linked AI summary

OpenStreetMap: Challenges and Opportunities in Machine Learning and Remote Sensing

John Vargas, Shivangi Srivastava, Devis Tuia, Alexandre Falcao

arXiv:2007.06277v1cs.CVeess.IV

TL;DR

OSM provides broadly used open geographic data, but volunteer editing creates heterogeneous completeness and quality. The paper reviews machine-learning methods that improve OSM layers or use OSM as reference or feature data in other applications. Its supported conclusion is that combining automatic and human-interactive approaches could improve OSM accuracy while using annotator labor efficiently.

  • Problem

    Volunteer editing produces heterogeneous OSM completeness and quality, while OSM’s value for geoscience and remote sensing motivates methods to improve and use these data.

  • Method

    The paper reviews machine-learning methods for improving OSM objects and for using OSM as reference or feature data in applications such as land-use classification and population estimation.

  • Results

    The review identifies methods spanning geometric objects and applications, including aerial-image correction of building annotations and OSM-supported land-use classification.

  • Takeaways & Limitations

    Combining automatic and human-interactive approaches could produce accurate OSM data while using annotators’ labor efficiently.

Abstract

from arXiv · show

OpenStreetMap (OSM) is a community-based, freely available, editable map service that was created as an alternative to authoritative ones. Given that it is edited mainly by volunteers with different mapping skills, the completeness and quality of its annotations are heterogeneous across different geographical locations. Despite that, OSM has been widely used in several applications in {Geosciences}, Earth Observation and environmental sciences. In this work, we present a review of recent methods based on machine learning to improve and use OSM data. Such methods aim either 1) at improving the coverage and quality of OSM layers, typically using GIS and remote sensing technologies, or 2) at using the existing OSM layers to train models based on image data to serve applications like navigation or {land use} classification. We believe that OSM (as well as other sources of open land maps) can change the way we interpret remote sensing data and that the synergy with machine learning can scale participatory map making and its quality to the level needed to serve global and up-to-date land mapping.

I. INTRODUCTION

OSM is a freely available volunteer-generated alternative to authoritative maps, but annotation quality and coverage vary geographically and by contributor skill. The paper reviews machine-learning methods that improve OSM data or use it with Earth Observation for applications across geoscience and related domains.

  • Motivation: OSM is a successful volunteered geographic information project offering freely available mapping data as an alternative to costly authoritative maps.Authoritative maps may contain temporal, spatial, and completeness inaccuracies because budgets limit regular updates.
  • Motivation: Volunteer numbers and mapping skills produce heterogeneous OSM completeness and quality, with urban areas generally updated more regularly than rural areas.Reported issues include missing roads, inaccurate road tags, and building-footprint positional errors.
  • Applications: OSM supports applications including land-cover validation, land-use classification, navigation, routing, aerial-image object detection, 3D city modelling, indoor mapping, and location-based services.These uses persist despite known completeness and quality issues.
  • Machine-learning synergy: Machine learning and Earth Observation are used both to complete or verify OSM layers and to train image-recognition models using OSM as label information.The latter use also raises concerns about label quality, standardization, and completeness.
  • Review scope: The paper reviews machine-learning methods for improving and using OSM, focusing on applications involving machine-learning techniques rather than an exhaustive catalog of OSM uses.Its organization covers OSM improvement, downstream applications, discussion, future research, and conclusions.

II. IMPROVING OSM DATA WITH MACHINE LEARNING

Machine learning methods improve OSM building data by detecting mismatches, vandalism, and annotation omissions or errors. Approaches combine OSM geometry with aerial imagery and neural or probabilistic models to assess, correct, remove, or create building footprints.

  • Detecting geometric mismatches: An autoencoder can learn geometric mismatch features and use reconstruction error as a regional score for OSM building-annotation quality.The method was evaluated against official governmental data in Toronto.
  • Detecting vandalism behavior: A clustering-based method detects possible vandalized building footprints by identifying outliers in geometric-feature space.Features include perimeter, elongation, convexity, and compacity; vandalism refers to intentional edits with wrong geometries.
  • Building-footprint errors: OSM building-footprint problems include incomplete annotations, geometric mismatches, nonexistent-building annotations, and missing buildings, especially in rural areas.Examples include incomplete annotations in São Paulo and Amsterdam and three rural error types illustrated over aerial imagery.
  • Correcting and creating annotations: Aerial-image methods use Markov Random Fields to align or remove problematic annotations and CNNs to predict missing building footprints.The rural-building approach addresses misalignment, annotations without matching buildings, and omitted buildings; another method uses contours from UAV oblique imagery.

B. Street Network

Machine learning methods improve OSM street networks by identifying missing roads, extracting roads from imagery or GPS, and predicting road types and attributes. These methods combine geometric, topological, temporal, and trajectory information with classifiers and neural models.

  • Motivation: OSM street-network quality affects route planning and map queries, while road networks exhibit heterogeneous quality and completeness errors even in urban areas.Street types, road completeness, and names are especially relevant to routing and querying.
  • Road-network completeness: A machine-learning method identifies candidate missing roads using OSM features such as connectivity, street type, and node degree.The approach uses candidate node pairs and finds empirical evidence that shortest-path distance correlates with straight-line distance.
  • Extracting roads from imagery: CNN-based approaches extract road pixels and centerlines from aerial imagery, with additional models generating or selecting candidates to repair network gaps.Other work combines aerial imagery and ground-based pictures to extract roads, parking spots, and sidewalks for possible OSM integration.
  • Road-type attributes: OSM geometry and topology support street-type classification, including distinctions among secondary, tertiary, and residential streets.Residential segments tend to be shorter, while tertiary roads connect with several residential streets; one approach classifies 21 street categories using a Random Forest.
  • GPS-based enrichment: GPS trajectories can detect OSM errors, recommend missing objects and attributes, and reconstruct road geometries, but sparse boundary data and GPS positional errors constrain performance.KNN performed best for geometry classification, Random Forest for object attributes, and multiple trajectories can mitigate 5–20 meter GPS errors.

C. Semantic tags

OSM semantic tags can be inconsistent because volunteers distinguish similar classes using varying knowledge and can create nonrecommended tag values. Machine-learning classifiers and tag-similarity methods support consistency checking, correction, and standardization.

  • Semantic-tag quality: Volunteer knowledge and interpretation can produce incorrect semantic tags because OSM lacks a rigorous object-classification system.The paper gives lake-versus-pond distinctions as an example of potentially inconsistent labeling.
  • Tag standardization: An unsupervised method identifies recommended OSM tags equivalent to annotator-created values using tag-usage statistics, geometry types, and similarity measures.The method addresses reduced standardization caused by tags outside OSM’s recommended value set.
  • Checking strategies: Classifiers can support contribution checking during editing, manual checking of selected existing objects, or automatic checking across OSM data.These strategies differ in when the classifier is applied and how users respond to inconsistent predictions.
  • Green-area labels: A KNN classifier uses geometric, topological, and contextual features to detect mislabeled green-area objects for manual verification or correction.The reviewed experiment examined four types of green-area objects and reported effective detection of potentially erroneous labels.

D. Points of Interest

Machine-learning methods support quality control and enrichment of OSM Points of Interest by evaluating new additions and predicting missing tags from names.

  • Quality control: POI quality control must assess spatial relationships that basic editing rules do not capture.Existing checks include basic topology rules such as avoiding duplicate elements, but they do not evaluate spatial relationships.
  • Quality control: A recommendation tool scores new POIs using spatial co-existence patterns and category-pair frequencies.It computes similarity with neighboring POIs and sums these scores into a plausibility score.
  • Quality control: The POI plausibility model should use data from the same city because different cities exhibit different spatial patterns.
  • Quality control: A Paris case study found a new ATM much more plausible downtown than beside a river and bridge, in a park, or outside the city.
  • Tag enrichment: Another method predicts missing POI tags from names using k-gram features and a Random Forest classifier.The approach can infer labels such as restaurant for a POI named “Chicken Palace.”

III. USING OSM DATA WITH MACHINE LEARNING ALGORITHMS

This section reviews machine-learning applications that use OSM data for land-use and land-cover mapping, building detection, navigation, traffic estimation, and fine-scale population estimation.

  • Machine-learning applications use OSM data for land-use and land-cover classification, building detection, navigation, traffic estimation, and fine-scale population estimation.

A. Land use and land cover

OSM supports land-use and land-cover mapping as reference data and as a source of labels for models using satellite, aerial, and ground-based imagery. Its value varies by class and benefits from combining complementary views.

  • Motivation: Land-use maps differ in quality across locations, while frequent updating remains important for keeping them useful.Governmental, commercial, and crowdsourced sources each have different update and quality characteristics.
  • OSM quality: Some OSM classes—forest, water bodies, and agricultural areas—show good quality and could support land-use planning.The comparison covered thematic accuracy and completeness against GMESUA datasets in seven European metropolitan regions.
  • Remote-sensing models: Supervised classifiers use OSM boundaries and labels with time-series Landsat or aerial imagery to predict land-cover classes.
  • Multimodal imagery: Ground-based imagery and pretrained CNN features enable prediction of 13 land-use types for OSM objects.
  • Multimodal imagery: Combining aerial and ground-based imagery improves land-use prediction because each view can reveal features the other obscures.Ground views help identify churches, whereas aerial imagery can better reveal sports facilities behind fences.

B. Building detection and segmentation

OSM building annotations provide training references for CNN-based segmentation and vectorization, while rural-building methods address sparse, lower-quality annotations through iterative detection and active learning.

  • Urban buildings: OSM building annotations are widely used as reference data for CNN-based building segmentation.
  • Vectorization: Active Contour Models and mesh-based approximation convert learned or binary building maps into vectorial footprints suitable for GIS software.
  • Rural buildings: Rural building annotations are less frequent and lower quality than urban annotations because rural buildings are sparsely distributed.
  • Rural buildings: A CNN-based rural-building detector iteratively selects new annotation samples through active learning and retrains the model.

C. Navigation

OSM supports navigation and traffic-related applications, including routing, localization, and traffic prediction. Machine-learning methods use OSM’s geographic and network information alongside other data to estimate traffic conditions and speed.

  • Navigation: OSM building and road-network data have similar quality to commercial map services in several urban areas, supporting navigation and routing applications.Reported applications include real-time shortest-path computation and pedestrian routing in open spaces.
  • Navigation: Visual odometry combined with OSM data achieves better vehicle-localization results than visual odometry alone.
  • Traffic estimation: Traffic prediction models combine OSM geographic objects with points of interest, weather, temperature, and housing prices to classify congestion.With class weighting, accuracy is high for the good-traffic class but poor for the other classes.
  • Traffic estimation: The Cluster-based Transfer Model for Prediction uses OSM road features and source-area traffic data to predict target-area traffic speed.CTMP outperforms Neural Networks and Support Vector Regression baselines.

E. Fine-scale population estimation

OSM building footprints and points of interest support fine-scale population estimation through areal interpolation and machine-learning models. These approaches operate at building or grid-cell scales using urban features derived from OSM.

  • Building-level estimation: OSM building footprints and points of interest can support building-level population estimation through areal interpolation without requiring 3D building models.
  • Grid-level estimation: A CNN-based method estimates population density for 200 × 200 meter grid cells from OSM-derived urban features.Features include building area, number of buildings, and number of points of interest.

IV. DISCUSSION AND FUTURE WORKS

The discussion identifies future directions for combining OSM with machine learning, multimodal data, user interaction, and richer search interfaces. It emphasizes improving data quality and coverage while extending OSM-based applications.

  • Multimodal data: Multimodal systems combine remote-sensing imagery, ground-based pictures, governmental data, tracking data, or social-media text for OSM-related classification and updating.Ground-based pictures may provide more current, freely available information despite less complete coverage than GSV.
  • Multimodal data: Social-media-based land-use classification offers massive data volumes but uses fewer land-use classes and requires workflows for denoising and filtering geolocated posts.
  • Supporting users: Interactive machine-learning workflows can reduce annotation effort by asking users to verify or correct automatically extracted roads.The reviewed road-network study reports greater efficiency than traditional manual annotation; active learning selects a small set of informative samples for labeling.
  • Supporting contributors: OSM’s low contributor activity motivates gamification and other efforts to encourage more frequent volunteer editing.From January to May 2019, active monthly contributors were fewer than 1% of registered users.
  • Supporting contributors: Contributor experience, area familiarity, annotation statistics, geometry, and temporal information can support estimation of annotation quality and trust.
  • Search and interaction: Natural-language interfaces such as NLmaps extend OSM search beyond place-name lookup by answering textual geographic questions and displaying matching objects on a map.NLmaps translates natural language into a machine-readable formula and structured Overpass query, but cannot answer some complex questions.

V. CONCLUSION

The review covers machine-learning efforts to improve OSM data and to use OSM in applications across other domains. It highlights complementary roles for OSM, machine learning, remote sensing, and human interaction.

  • The review covers machine-learning methods that improve OSM points, lines, and polygons, using both traditional and deep learning techniques.Only a few reviewed methods have been implemented in OSM editors.
  • A second group of studies uses OSM either as reference data for training models or as features for applications such as land-use classification and population estimation.
  • Combining automatic and human-interactive approaches could improve OSM accuracy while using annotators’ labor efficiently.
  • The increasing availability of up-to-date remote sensing data creates opportunities at the interface of machine learning, OSM, and other disciplines.
Loading 2007.06277v1…