Abstract
Accurate mapping of infrastructure networks like irrigation canals and roads from satellite imagery is often hindered by noisy or incomplete ground truth data. Standard semantic segmentation models lack global topological awareness, leading to disconnected network predictions. We present IGraSS (Iterative Graph-constrained Semantic Segmentation), a framework that integrates deep learning with graph-theoretic constraints to iteratively refine network connectivity. By enforcing global properties—such as reachability from a water source—IGraSS automatically repairs gaps in training data, significantly improving the connectivity and accuracy of the resulting infrastructure maps.
Introduction
Infrastructure networks are the backbone of resource management and disaster planning. While satellite imagery provides a wealth of data, manual mapping is labor-intensive and error-prone. Standard AI models for image segmentation operate pixel-by-pixel, often failing to capture the functional connectivity of a network. A canal segment is only useful if it connects to a water source, yet shadows or vegetation often create artificial gaps in both imagery and ground truth datasets. IGraSS addresses this by leveraging the inherent graph-level properties of infrastructure to supervise local pixel classification.
The IGraSS Framework
The IGraSS framework cycles between two primary components: a Learner and a Graph Refinement Module.
The Learner
The Learner is a semantic segmentation model (e.g., Swin Transformer or DeepLabV3+) that processes multi-modal satellite imagery. Beyond standard RGB channels, it incorporates the Normalized Difference Water Index (NDWI) and Digital Elevation Models (DEM). These additional modalities provide critical context: NDWI highlights water content, while DEM data helps the model learn that canals typically follow specific elevational logic, distinguishing them from roads or shadows.
Graph Refinement and Iterative Learning
The Graph Refinement Module treats identified canal pixels as a graph. It identifies "unreachable" segments—those disconnected from the main network—and uses a constraint solver based on Dijkstra’s shortest path algorithm to find the most likely connecting paths. These "pseudo-labels" for missing gaps are added back into the training set. Over multiple iterations, the ground truth is cleaned, and the model becomes increasingly adept at identifying the full, connected network.
Experimental Evaluation
The authors evaluated IGraSS on two primary datasets:
- Canal Network Dataset: 30,000 patches of 3m-resolution PlanetScope imagery from central Washington, incorporating RGB, NDWI, and DEM data.
- Road Network Dataset: 60cm/pixel imagery of New York City, used to test generalizability by repairing intentionally introduced gaps in OpenStreetMap data.
To better evaluate network structure, the authors proposed r-neighborhood metrics. Unlike traditional pixel-wise metrics that penalize slight spatial offsets, these metrics allow for a small buffer, providing a more realistic assessment of topological accuracy.
Key Results
- Connectivity Improvement: IGraSS reduced the percentage of unreachable canal segments from 18% to just 3%.
- Segmentation Performance: Refining the ground truth improved the F1-score of the Swin Transformer by 5.8% and boosted ResUnet performance by 10%.
- Generalizability: The framework successfully repaired road networks by minimizing total shortest path length, proving the versatility of the graph-constrained approach.
Conclusion
IGraSS demonstrates that global topological constraints can effectively supervise local segmentation tasks. By treating ground truth as a refinable asset rather than a static input, the framework overcomes the limitations of noisy data. Future work will explore integrating flow direction logic and adaptive thresholding to further enhance robustness in complex terrains.