Deciphering Region-Level Signatures from Latency Measurements in LEO Satellite Internet

A hierarchical analytical framework for cross-region latency characterization using Starlink RTT measurements from the LENS dataset

Xiang Shi; Yifei Zhang; Peng Hu· 2026· DOI 10.48550/arXiv.2606.29324

The core problem

Low-Earth orbit (LEO) satellite Internet has become an indispensable infrastructure providing growing coverage for global users. Despite extensive measurement efforts, the principles underlying region-level performance characteristics remain insufficiently understood, limiting the ability to identify region-specific latency signatures under dynamic network conditions. This gap is critical as LEO constellations like Starlink expand rapidly, yet operators and researchers lack systematic methods to compare performance across diverse geographic deployments. The authors address this by formulating the problem of region-level latency characterization using Starlink round-trip time (RTT) measurements from the public LENS dataset. They propose a hierarchical analytical framework that transforms raw RTT sequences into multi-scale statistical features for cross-region comparison. The study focuses on five geographically representative regions, aiming to uncover how deployment factors such as infrastructure availability and dish-to-Point-of-Presence (PoP) distance influence latency. By bridging measurement data with statistical learning, the work seeks to enable more informed network planning

Innovation

Using data from five geographically representative regions, the authors demonstrate that latency differences are strongly associated with deployment factors, particularly infrastructure availability and Starlink dish-to-Point-of-Presence (PoP) distance. Mutual information analysis identifies minimum RTT as the most discriminative feature, which is further supported by XGBoost-based feature importance. This suggests that the lowest observed RTT in a region is a robust indicator of its unique latency signature, likely reflecting the best-case path and local infrastructure quality. The proposed model achieves 83% accuracy on short-term data, indicating that the multi-scale statistical features effectively capture region-level distinctions within a limited temporal scope. However, the performance degrades over longer periods, revealing limited temporal generalization. This degradation implies that latency signatures are not static; they evolve due to dynamic network conditions, satellite handovers, and varying load. The results highlight that while short-term characterization is feasible, long-term prediction requires adaptive models that can account for temporal drift. The study does

Low-Earth orbit (LEO) satellite Internet has become an indispensable infrastructure providing growing coverage for global users. Despite extensive measurement efforts, the principles underlying region-level performance characteristics remain insufficiently understood, limiting the ability to identify region-specific latency signatures under dynamic network conditions. This gap is critical as LEO constellations like Starlink expand rapidly, yet operators and researchers lack systematic methods to compare performance across diverse geographic deployments. The authors address this by formulating the problem of region-level latency characterization using Starlink round-trip time (RTT) measurements from the public LENS dataset. They propose a hierarchical analytical framework that transforms raw RTT sequences into multi-scale statistical features for cross-region comparison. The study focuses on five geographically representative regions, aiming to uncover how deployment factors such as infrastructure availability and dish-to-Point-of-Presence (PoP) distance influence latency. By bridging measurement data with statistical learning, the work seeks to enable more informed network planning and adaptive model design for long-term performance.

The authors leverage the public LENS dataset, which contains Starlink RTT measurements from diverse global locations. They select five geographically representative regions to capture a range of deployment conditions. The core methodology is a hierarchical analytical framework that processes raw RTT sequences into multi-scale statistical features. These features likely include metrics such as minimum, maximum, mean, median, and percentiles of RTT, as well as variability measures (e.g., standard deviation, interquartile range) computed over different time windows. The framework enables cross-region comparison by normalizing and aggregating these features. To identify the most discriminative latency signatures, the authors employ mutual information analysis and XGBoost-based feature importance. Mutual information quantifies the dependency between each feature and the region label, while XGBoost provides a model-based ranking. The combination allows robust feature selection. The resulting model is evaluated for short-term accuracy and tested for temporal generalization over longer periods. The methodology is summarized in the following Mermaid diagram:

Why it matters

The findings underscore that region-level latency in LEO satellite Internet is not merely a function of distance but is significantly shaped by deployment factors. Infrastructure availability—such as the number and quality of ground stations and PoPs—and the distance from user dishes to the nearest PoP emerge as strong correlates of latency differences. This aligns with the physical intuition that longer terrestrial backhaul segments add to RTT. The prominence of minimum RTT as the most discriminative feature suggests that it captures the baseline performance achievable in a region, relatively free from transient congestion. However, the limited temporal generalization of the model is a critical limitation. The degradation over longer periods indicates that static feature representations and models trained on short-term data cannot capture evolving network dynamics, such as satellite constellation changes, traffic patterns, and weather effects. The authors motivate the need for adaptive models and feature representations for long-term performance. Future work could explore online learning, recurrent neural networks, or time-series forecasting to address temporal drift. Additionally, expanding the number of regions and incorporating more deployment metadata (e.g., exact PoP locations, ground station capacities) could improve the framework's explanatory power. The study contributes a reproducible methodology and highlights minimum RTT as a key signature, but also reveals that region-level latency characterization remains an open challenge for long-term reliability.

Who should read this

CS practitioners and researchers

Opening member content…