🤖 AI

IGraSS: Learning to Identify Infrastructure Networks from Satellite Imagery by Iterative Graph-constrained Semantic Segmentation

This paper introduces IGraSS, an iterative framework that combines multi-modal semantic segmentation with graph-constrained ground-truth refinement to significantly improve the accuracy of mapping infrastructure networks like canals and roads from satellite imagery by leveraging topological properties such as connectivity and reachability.

Original authors: Oishee Bintey Hoque, Abhijin Adiga, Aniruddha Adiga, Siddharth Chaudhary, Madhav V. Marathe, S. S. Ravi, Kirti Rajagopalan, Amanda Wilson, Samarth Swarup

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Oishee Bintey Hoque, Abhijin Adiga, Aniruddha Adiga, Siddharth Chaudhary, Madhav V. Marathe, S. S. Ravi, Kirti Rajagopalan, Amanda Wilson, Samarth Swarup

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a map of every single irrigation canal in a massive farming region using satellite photos. You want to know exactly where the water flows so farmers can manage it better.

The problem is that the "ground truth" maps we have (the ones experts drew by hand) are messy. They are like a puzzle with missing pieces and broken lines. Some canals look like they just stop in the middle of a field, even though water must be flowing there. If you try to teach a computer to recognize canals using these broken maps, the computer learns the mistakes and produces a bad map.

IGraSS is a new "smart teacher" system designed to fix this broken puzzle and teach the computer to draw a perfect map. Here is how it works, using simple analogies:

1. The Two-Step Dance: The Artist and The Editor

IGraSS works like a collaboration between an Artist and an Editor who take turns improving the map.

The Artist (The Learner): This is a computer program trained to look at satellite photos and guess, "Is this pixel a canal?" It uses different "eyes" to see the world: standard color photos (RGB), a special water-detecting filter (NDWI), and a 3D height map (DEM).
The Editor (The Graph Constraint): This is the smart part. The Editor knows a fundamental rule of physics: Water doesn't just appear out of nowhere. Every canal must be connected to a water source (like a river or a reservoir). If the Artist draws a canal that floats in the middle of nowhere, the Editor says, "That's impossible. Fix it."

2. The Iterative Loop (The "Fix-It" Cycle)

The system doesn't just run once; it runs in a loop, getting better every time:

Round 1: The Artist looks at the messy, broken maps and the satellite photos and draws a canal network.
The Check: The Editor looks at this new drawing. It finds "dead ends"—canal segments that aren't connected to any water source.
The Fix: The Editor uses a "connect-the-dots" strategy. It looks at the gaps and says, "If this canal is here, and that water source is there, the water must flow through this gap." It fills in the missing lines and updates the "ground truth" map.
Round 2: The Artist is shown the new, fixed map and tries again. Because the map is now less broken, the Artist learns better.
Repeat: They do this over and over. With every round, the map becomes cleaner, and the "dead ends" disappear.

3. The Results: From Broken to Connected

The paper tested this on canals in Washington state.

Before IGraSS: About 18% of the canal segments were "unreachable" (floating in space, disconnected from water).
After IGraSS: They reduced this to just 3%.

Think of it like fixing a leaky pipe system. Initially, 18% of the pipes were disconnected. After the Editor ran its logic, almost all the pipes were reconnected to the main water supply.

4. A New Way to Measure Success

The authors also realized that standard ways of measuring accuracy (like counting pixel-perfect matches) are too strict for thin lines like canals. If a computer draws a canal that is mostly right but shifted by one pixel, standard math says it's a failure.

The team invented a "neighborhood metric." Imagine you are checking a drawing, and you give the artist a little wiggle room. If the artist draws the canal within a small circle around the real canal, you count it as a success. This gives a more realistic score of how well the computer understands the structure of the network, not just the exact pixels.

5. It Works for Roads Too

To prove this idea is powerful, they tried it on road networks. Roads, like canals, must be connected. If a road ends in the middle of a field, it's a broken map. IGraSS successfully used the same logic to connect broken road segments, showing that this "Artist and Editor" team-up works for any infrastructure that needs to be connected.

In summary: IGraSS is a system that doesn't just let a computer guess; it forces the computer to follow the rules of physics (connectivity). By constantly checking its own work against these rules and fixing the mistakes, it turns a messy, incomplete map into a reliable, connected network.

Technical Summary: IGraSS – Learning to Identify Infrastructure Networks from Satellite Imagery by Iterative Graph-constrained Semantic Segmentation

1. Problem Statement

Accurate mapping of infrastructure networks, specifically irrigation canals, is critical for water management, irrigation planning, and infrastructure maintenance. However, existing manual mapping processes are labor-intensive, slow, and often result in noisy, incomplete, or outdated ground truth data. While deep learning-based semantic segmentation has advanced significantly for road networks, its application to irrigation canals is hindered by a lack of high-quality, well-annotated datasets.

A core challenge identified is that standard semantic segmentation models operate on local image patches and fail to capture global topological properties inherent to infrastructure networks, such as reachability (e.g., every canal segment must be connected to a water source) or connectivity. Consequently, models trained on incomplete ground truth often produce fragmented outputs, failing to identify disconnected canal segments or incorrectly labeling non-canal areas.

2. Methodology: The IGraSS Framework

The authors propose IGraSS (Iterative Graph-constrained Semantic Segmentation), a novel framework that combines a semantic segmentation module with a graph-based ground-truth refinement module. The framework operates iteratively to refine noisy ground truth and improve model performance.

Core Components

Semantic Segmentation Module (The Learner):
- Processes satellite imagery patches to predict canal pixels.
- Utilizes multi-modal input data: RGB imagery, NDWI (Normalized Difference Water Index), and DEM (Digital Elevation Model).
- Employs state-of-the-art backbone architectures, including DeepLabV3+, ResNet50, ResUNet, and Swin Transformer.
Graph-based Ground-Truth Refinement Module (The Constraint Solver):
- Treats the entire infrastructure network as a graph $G$ where nodes are grid cells (pixels).
- Identifies unreachable segments: Canal pixels that are disconnected from known water sources.
- Network Completion: For each unreachable terminal node $t$ , the algorithm identifies a local graph $H_t$ within a radius $\rho$ . It then computes a minimum weighted shortest path from $t$ to the nearest reachable source nodes ( $S_t$ ) using Dijkstra's algorithm.
- Pseudo-labeling: Pixels along the computed shortest paths are assigned positive pseudo-labels (changing from non-canal to canal), effectively "filling the gaps" in the ground truth.
Iterative Process:
- Iteration 0: The model is trained on the initial, potentially noisy ground truth.
- Iteration $i \ge 1$ :
  1. The learner predicts the network.
  2. The refinement module analyzes the output to identify unreachable terminals.
  3. Network completion instances are solved to connect these terminals to sources.
  4. The ground truth is updated with these new connections (pseudo-labels).
  5. The model is re-trained on the refined ground truth for the next iteration.

Evaluation Metrics

Recognizing that conventional pixel-level metrics (Precision, Recall, IoU) may penalize models for minor spatial misalignments in thin structures (like canals represented by line segments), the authors introduce r-neighborhood metrics. These metrics allow for small spatial deviations ( $r$ ) between predicted and ground truth pixels, providing a more nuanced evaluation of structural accuracy.

3. Key Contributions

The paper outlines five primary contributions:

Hybrid Framework: A novel approach combining learning with optimization/constraint satisfaction via pseudo-labels to enforce global graph properties (e.g., reachability).
Refined Metrics: The introduction of r-neighborhood metrics (r-Precision, r-Recall, r-IoU, r-F1) to better evaluate semantic segmentation of thin structures where pixel-perfect alignment is less critical than structural continuity.
Canal Network Identification: A demonstration of improved performance in identifying irrigation canals by leveraging reachability constraints to correct fragmented ground truth.
Generalizability: A proof of concept applying the framework to road networks, using pairwise distance constraints to complete road graphs, demonstrating the method's adaptability to different infrastructure types.
Open Resources: The release of refined, less noisy canal data for Washington State and the accompanying codebase to facilitate further research.

4. Experimental Results

The framework was evaluated on a dataset of 30,000 patches from central Washington (RGB, NDWI, DEM) and a road network dataset from New York City.

Performance on Canal Networks

Impact of Refinement: Training models with the refined ground truth generated by IGraSS significantly outperformed baseline models trained on original data.
- For the Swin Transformer model, using IGraSS improved Precision from 0.775 to 0.820 (+5.8%), Recall from 0.765 to 0.810 (+5.2%), and F1-score from 0.770 to 0.815 (+5.8%) on Test Set 1.
- Similar improvements (approx. 5-10%) were observed across DeepLabV3+ and ResUNet models.
Network Completion: The iterative process successfully reduced the percentage of unreachable canal segments from 18% to 3%.
Multi-modal Input: The inclusion of NDWI and DEM alongside RGB yielded the best results. Adding NDWI improved performance by 35-40%, while adding DEM contributed an additional 10-20%. The combination of all three modalities provided the highest boost (45-50% over RGB-only baselines).

Ablation and Parameter Sensitivity

Parameters: The framework's performance is sensitive to the confidence threshold ( $\alpha$ ) and radius ( $\rho$ ). A lower $\alpha$ (e.g., 0.01) with fewer epochs reduced unreachable canals but introduced noise, whereas a moderate $\alpha$ (0.1) with more training epochs yielded cleaner results.
Generalizability: The approach was successfully adapted to road networks, demonstrating that graph-theoretic constraints can effectively complete infrastructure maps across different domains.

5. Significance and Claims

The authors position IGraSS as a robust framework for addressing the specific challenges of infrastructure mapping where ground truth is often incomplete or noisy.

Addressing Data Scarcity: The paper claims that IGraSS effectively leverages graph-level properties (reachability, connectivity) to improve existing, imperfect ground truth, thereby reducing the reliance on perfectly annotated datasets.
Sustainable Development: By enabling automated, accurate canal mapping, the work supports water conservation and infrastructure modernization, aligning with UN Sustainable Development Goal 12 (Responsible Consumption and Production).
Methodological Advancement: The paper argues that integrating global constraints into the training loop via iterative refinement is a superior approach to standard supervised learning for topological problems, outperforming methods that rely solely on local patch-based decisions or post-processing heuristics.

The authors maintain a modest stance, noting that while the framework significantly improves connectivity, some unreachable segments remain due to the absence of visible canals or water sources within the defined radius in the imagery, highlighting the limits of remote sensing in certain conditions.