Satellite Imagery for Map Data Updating: Key Elements, Semantic Segmentation Techniques, and Future Challenges

For map services, the accuracy and coverage of map data are critical to service quality. Map data updates rely on multiple information sources such as trajectory heatmaps, ground‑level imagery, and satellite imagery. In recent years, the proliferation of remote‑sensing satellites and the emergence of high‑resolution spectral cameras have made satellite imagery an increasingly important source for map data updates because of its wide coverage, favorable viewing angle, and rich information.

Gaode (Amap) has evolved its use of satellite imagery from front‑end user display to manual data‑processing reference and now to proactive discovery of map updates. This article introduces the exploration and practice of upgrading satellite imagery from a passive reference to an active discovery tool.

Key Elements of Satellite Imagery

According to geometric structure, image elements are divided into three major categories: road elements, region (land‑cover) elements, and building elements.

Road Elements : include ordinary roads, fine‑grained roads (main/auxiliary roads, bike lanes, early‑right‑turn roads), and connection points (through‑roads, entrances/exits, U‑turns, intersections, etc.).

Region Elements : include building zones, demolition zones, water bodies, farmland, mountainous areas, forest land, greenhouses, and other land‑cover types.

Building Elements : refer to building blocks.

Advantages of Satellite Imagery in Data Updating

The road network is the foundation of map data; all road attributes, dynamic events, and POI guidance depend on accurate road‑network information. Satellite imagery, with its “bird’s‑eye” view, provides comprehensive information for assessing road connectivity, complex intersections, and over‑pass relationships. Moreover, its wide coverage and low cost make it ideal for supplementing road‑network data in areas with sparse heatmaps or where vehicle‑based collection is difficult.

POI (Point of Interest) coordinate accuracy is crucial for navigation. Statistics from Gaode’s POI database (Top 10 million) show that 70 % of POIs need to be bound to building blocks, indicating a strong dependency between POIs and the surrounding buildings.

Satellite Imagery Recognition Technology Exploration

Fine Semantic Segmentation

To improve algorithm precision, the focus is on integrating contextual information using architectures such as U‑Net, ASPP, Non‑local, and Attention mechanisms. Attention modules enhance the network’s focus on salient image regions, further boosting segmentation performance.

U‑Net Structure : An encoder‑decoder network with four down‑sampling and four up‑sampling stages, employing skip connections to fuse low‑level and high‑level features, enabling multi‑scale prediction and deep supervision.

ASPP : Utilizes dilated convolutions with varying rates to capture multi‑scale features and combine global and local information.

Attention : Merges shallow and deep layer features, generates attention parameters, and applies them to deep features to highlight important regions.

Non‑local : Encodes long‑range pixel relationships to inject global context into the feature maps, overcoming the limited receptive field of local convolutions.

Image Block Instance Segmentation

Two mainstream approaches exist: (1) Proposal‑based, which first detects objects and then performs semantic segmentation within the detected boxes; (2) Proposal‑free, which clusters pixels directly on the segmentation map. For building blocks, the proposal‑based method yields better results due to the diverse and “short‑wide” nature of building structures.

Because the base shape of most building blocks matches their roof shape, a multi‑task learning scheme is adopted: first segment the roof, then predict an offset vector from roof to base, reconstructing the base’s shape and position accurately.

Multi‑Element Recognition Results

Multiple recognition models have been designed for different satellite‑image elements, including ordinary road detection, fine road network detection, region classification, and building‑block detection, supporting various data‑update scenarios in Gaode.

Future Outlook & Challenges

Accurate & Rapid Road‑Network Updates : Users encounter navigation issues caused by outdated road data (e.g., new roads not reflected, blocked roads still suggested, missing turn‑around points). The goal is to accelerate detection of road‑network changes through visual‑algorithm optimization and multi‑source fusion.

Building Blocks & AOI in Digital Cities : Accurate building‑block and Area‑of‑Interest (AOI) data are essential for precise navigation and can support pandemic‑control measures. Leveraging satellite‑imagery discovery capabilities aims to enrich digital‑city data, bridging the virtual and real worlds.

Gaode’s visual‑intelligence team is actively recruiting for positions in computer vision, image recognition, point‑cloud processing, 3D, AR, and video algorithms. Interested candidates can send resumes to [email protected], indicating “Gaode Tech WeChat” as the source.