IPNL Intelligent Positioning & Navigation Lab Search EN 简 繁

PositionNet: CNN-based GNSS positioning in urban areas with residual maps

Xu, P., Zhang, G., Yang, B., Hsu, L. T.

Applied Soft Computing (2023)

journal Q1 Featured page
DOI
Hong Kong route maps showing PositionNet solutions against ground truth.
Figure 4 visual montage: Hong Kong route maps showing PositionNet solutions against ground truth.

Summary

PositionNet uses CNNs and residual maps to learn urban GNSS position corrections.

Figures

Before the CNN architecture, the paper defines how satellite range residuals become image-like maps around the least-squares solution.

Geometry of the range residual map with satellite direction, sampling location, and user location.
Figure 1: geometry of the residual map used as PositionNet input.

The SDRes map removes receiver clock bias by differencing with a master satellite, creating a clearer location pattern for the network.

Single-differenced residual map layers intersecting at the user location.
Figure 2: SDRes maps strengthen the location pattern when multiple LOS satellites agree.

The first result maps PositionNet outputs across urban districts to show validation and testing behavior beyond a single scene.

Validation and testing PositionNet results over Tsim Sha Tsui, Whampoa, Mong Kok, and Kowloon Bay maps.
Figure 4: validation and testing results displayed on urban maps.

The case study then zooms into one large least-squares error to show how the output heat map points to a corrected location.

Case 1 PositionNet solution map and output heat map.
Figure 5: Case 1 positioning solution and output heat map.

The saliency maps explain that case by showing which satellite residual, SDRes, and C/N0 layers influenced the network.

Case 1 saliency map grid showing residual maps, SDRes maps, C/N0 maps, and saliency responses.
Figure 6: Case 1 saliency map explaining which satellite residual patterns influenced the network.

The aggregate saliency plot ties the case-level explanation back to measurement quality across validation and testing data.

Scatter plots of saliency map weighting versus pseudorange error for validation and testing data.
Figure 11: saliency map weighting versus pseudorange error for validation and testing data.

Key idea. PositionNet learns to correct urban GNSS errors with a CNN that reads residual maps — turning the spatial pattern of pseudorange residuals into a learned position correction.

Impact. Shows that data-driven methods can complement model-based 3D-mapping-aided GNSS, an early thread in the lab’s AI-for-positioning direction.