Source-linked AI summary
Leveraging Crowdsourced GPS Data for Road Extraction from Aerial Imagery
Tao Sun, Zonglin Di, Pengyu Che, Chun Liu, Yin Wang
TL;DR
Aerial imagery can leave roads obscured or disconnected, while local verification is labor-intensive. This paper fuses crowdsourced GPS with aerial imagery using GPS rendering, augmentation, and a 1D transpose convolution. The authors report better road extraction than image-only or GPS-only models and improved robustness in new areas.
Problem
Aerial imagery can be occluded or insufficient for road extraction, while systematic local verification is labor-intensive.
Source: §1 · pp. 1–2Method
The paper fuses crowdsourced GPS and aerial imagery, using GPS rendering, data augmentation, and 1D transpose convolution.
Source: §1 · p. 2Results
Takeaways & Limitations
Aerial imagery and GPS complement each other: imagery works best for residential-road detection, while GPS can recover arterial roads under severe image occlusion.
Source: §6 · p. 8Map matching GPS traces to roads by topology could add confidence, but is beyond the scope of this paper.
Source: §5 · p. 7
Abstract
from arXiv · showhide
Deep learning is revolutionizing the mapping industry. Under lightweight human curation, computer has generated almost half of the roads in Thailand on OpenStreetMap (OSM) using high-resolution aerial imagery. Bing maps are displaying 125 million computer-generated building polygons in the U.S. While tremendously more efficient than manual mapping, one cannot map out everything from the air. Especially for roads, a small prediction gap by image occlusion renders the entire road useless for routing. Misconnections can be more dangerous. Therefore computer-based mapping often requires local verifications, which is still labor intensive. In this paper, we propose to leverage crowdsourced GPS data to improve and support road extraction from aerial imagery. Through novel data augmentation, GPS rendering, and 1D transpose convolution techniques, we show almost 5% improvements over previous competition winning models, and much better robustness when predicting new areas without any new training data or domain adaptation.
1. Introduction
Aerial imagery can miss or misclassify roads, while road gaps and incorrect connections make routing unreliable and local verification labor-intensive. The paper fuses crowdsourced GPS with imagery and reports improved road extraction and robustness in new areas.
1. Introduction
- Computer-generated mapping has reached large scale, including 370 thousand km of roads in Thailand and 125 million building polygons in U.S. Bing maps.The Thailand roads account for 46% of the country's total roads.Source: §1 · p. 1
- Aerial road mapping is challenged by model overfitting, occlusion, visually indistinguishable features, and the need for local knowledge.For roads, incorrect routes can lengthen travel, lead people into restricted areas, or cause fatal accidents; OSM therefore favors local mappers and requires local verification for large-scale edits.Source: §1 · p. 1
- The fused image-and-GPS model significantly outperforms models using images or GPS alone, and GPS augmentation improves performance in new areas.The introduction also reports that augmentation is especially effective against overfitting, despite differences in GPS quantity and resolution in new areas.Source: §1 · p. 2
- The paper proposes using crowdsourced GPS alongside aerial imagery to extract roads and help resolve cases where imagery is insufficient.The GPS data can help discover roads, confirm continuity, and avoid misconnection.Source: Figure 1 · p. 1, §1 · p. 2
2. Related Work
Prior work spans aerial-image segmentation, road-connectivity repair, and inference from GPS traces, but comparatively little work combines crowdsourced GPS with aerial imagery. This paper builds on that gap using larger, noisier GPS datasets and more recent segmentation models.
2. Related Work
- Aerial segmentation has a long research history and substantial real-world use, including computer-generated roads in Thailand and building polygons in U.S. maps.Earlier image-processing solutions were often limited in accuracy and difficult to generalize to real-world datasets.Source: §2 · p. 2
- Deep segmentation evolved from classifying center pixels using larger image patches to fully convolutional models that segment whole images with more context.Early patch-based approaches could not learn complicated street or building blocks because of limited input size; later challenge leaders used U-net [24] or Deeplab.Source: §2 · p. 2
- Road-connectivity methods repair gaps through post-processing, topology-aware losses, or iterative CNN-guided search; GPS can complement them by confirming connectivity or its absence despite image occlusion.The cited approaches include Conditional Random Fields, shortest-path search, Line Integral Convolution, topology-aware loss [20], and iterative search [5].Source: §2 · p. 2
- Prior road inference from GPS traces includes crowdsourced data, with Kernel Density Estimation commonly used for robustness to GPS noise and disparity.Earlier work often used dense samples from controlled experiments, while more recent work examined crowdsourced data under varying sampling intervals and noise levels.Source: §2 · p. 2
- Only limited prior work combines GPS data and aerial imagery, and this paper tests the approach with larger crowdsourced datasets and state-of-the-art segmentation models.Earlier combined-data work used an outdated segmentation model and GPS from a controlled experiment.Source: §2 · p. 2
3. Crowdsourced GPS Data
The Beijing taxi dataset spans about a week and the Shanghai taxi dataset about half a year, with large sample counts in both. Crowdsourced GPS varies in noise, spatial distribution, sampling interval, and measurement resolution, motivating models robust to differing data characteristics.
3. Crowdsourced GPS Data
- The Beijing dataset contains about 81 million samples from around 28 thousand taxis, while Shanghai contains 1.5 billion samples from around 16 thousand taxis.Beijing covers about one week; Shanghai spans about half a year.Source: §3 · p. 3
- Crowdsourced GPS can be noisy because of multipath errors, variable receiver quality, non-Gaussian filtering, or privacy-related resolution reduction.The paper notes urban canyons, tunnels, and elevated roads or bridges as settings where multipath errors occur.Source: §3 · p. 3
- GPS samples are unevenly distributed: highways and intersections may have orders of magnitude more data than other roads, while some residential roads receive none.Samples may also cluster in non-road areas, and data can be shifted or collected during different time periods.Source: §3 · p. 3
- Sampling intervals and measurement resolutions differ across vehicles and datasets, while 50 cm/pixel imagery creates a mosaic of empty and multiply sampled pixels.Beijing sampling intervals most commonly include 10, 60, 180, and 300 seconds; Shanghai's include 16 and 61 seconds. Beijing coordinates are roughly 1 m resolution, versus roughly 10 m for Shanghai.Source: §3 · p. 3
4. Method
The method combines crowdsourced GPS layers with aerial RGB imagery, augmenting GPS inputs and adapting the decoder to road geometry. GPS rendering and four-direction 1D transpose convolutions support road extraction across GPS conditions.
4. Method
- GPS layers are concatenated with RGB input, while GPS augmentation and a 1D decoder are designed to improve robustness and road-extraction accuracy.The approach is intended to work with existing CNN-based semantic segmentation networks.Source: §4 · p. 4
4.1. Architecture
- 1D transpose convolutions replace 3×3 decoder filters with four directional filter sets, matching parameter count and computation while helping reduce gaps on roads with sparse GPS samples.The method uses horizontal, vertical, and two diagonal directions; each 1D filter has nine parameters.Source: §4.1 · p. 4
4.2. Data Augmentation
- GPS augmentation counters overfitting and sensitivity to GPS quantity and quality by subsampling, lowering resolution, perturbing samples, and omitting random areas.Sub-resolution augmentation also trains the model to handle larger unseen GPS quantities.Source: §4.2 · pp. 4–5
4.3. Rendering
- Gaussian kernel smoothing renders GPS samples as image layers, with log scaling making infrequently traveled roads more prominent than linear scaling.Connecting consecutive samples can shortcut intersections and curves under low sampling frequency, and did not improve results in these experiments.Source: §4.3 · p. 5
5. Experiments
Across the experiments, combining aerial imagery with GPS and the proposed enhancements improved road extraction, including a 4.76% gain over the image-only baseline and smaller IoU loss in a new domain. The experiments also examine GPS-only and image-only baselines, GPS rendering and augmentation, and GPS-based verification.
5. Experiments
- The experiments used 350 Beijing and 50 Shanghai satellite images at 50 cm/pixel, with manually labeled road pixels and a 70/10/20 train-validation-test split.Except for the new-area experiment, training and testing used Beijing imagery and GPS data.Source: §5 · pp. 5–6
- For GPS-only extraction, deep neural nets outperformed KDE, LinkNet performed best, and the 1D decoder helped relatively shallow models.The 1D decoder gave about a 1% increase against more complex models; D-LinkNet's additional complexity led to more overfitting on the relatively simple GPS data.Source: §5 · p. 6
- With image and GPS input, D-LinkNet remained the top performer; other relatively strong image models gained about 2%, with the 1D decoder adding about 1%.The comparison concerns the models evaluated with both input types.Source: §5 · p. 6
- GPS augmentation improved performance with degraded GPS inputs and added about 0.5% over the top-of-the-line performance.Figure 7 compares augmentation techniques using subsets and reduced-resolution GPS inputs.Source: §5 · p. 6, Figure 7 · p. 7
- The chosen GPS inputs were kernel-size-three Gaussian rendering and a sampling-interval channel; adding sampling interval alone gave the best gain among tested extra layers.Figure 8 evaluates rendering kernel sizes and scales.Source: §5 · p. 7, Figure 8 · p. 7
- GPS input can flag disagreement with image-only predictions for verification, while the combined inputs produced relatively complete roads under dense tree canopy and sparse GPS sampling.Map matching GPS traces to roads by topology was beyond the paper's scope.Source: §5 · pp. 7–8
- 4.76% performance gain over the image-input-only baseline was achieved by combining the evaluated improvements.Table 3 reports the overall gain from the improvements over the image-only baseline.Source: Table 3 · p. 7, §5 · p. 7
- On Shanghai testing, IoU dropped 18.9% for the model trained on both datasets, versus 31.6% for the image-only model; GPS augmentation further enhanced the gain.Table 4 reports the new-area testing results.Source: Table 4 · p. 7, §5 · p. 7
6. Conclusion
Fusing crowdsourced GPS data with aerial imagery yields better road segmentation than either source alone. GPS augmentation improves performance in new test areas, while the two sources complement each other across road types and occlusion conditions.
6. Conclusion
- GPS-plus-image input gives much better road segmentation than using images or GPS data alone with deep neural net models.The conclusion states the comparison without giving a numerical margin here.Source: §6 · p. 8
- GPS data augmentation is very effective against overfitting, and the method performs much better in new testing areas than other models.Source: §6 · p. 8
- Aerial imagery works best for residential-road detection, while GPS can recover arterial roads under severe image occlusion, making the sources complementary.The conclusion notes residential roads are relatively simple, numerous, and infrequently traveled, while GPS handles complicated highway systems.Source: §6 · p. 8