Source-linked AI summary
Lightweight Multi-scale Hierarchical Anomaly Detection and Localization for Geospatial Big Data Applications at the Edge
Thomas Benton Townsend, Joshua Bean, Benjamin K Tkach, Narcisa Gabriela Pricope, Dimitrios Michael Manias
TL;DR
Critical geospatial applications need anomaly detection despite growing storage, processing, and communication constraints in centralized architectures. The paper uses H3-based hierarchical aggregation and multi-scale drill-down at the edge to detect and localize macro anomalies. It reports over 99.7% fewer terminal-resolution evaluations than a flat baseline while filtering flickering anomalies.
Problem
Growing geospatial data volume and velocity make centralized raw-data processing increasingly difficult, while existing approaches often overlook edge deployment and flickering-anomaly filtering.
Method
The framework aggregates geospatial observations with H3, compares statistics against historical baselines, and drills down from low-resolution anomalies to higher-resolution localization.
Results
Over 99.7% fewer terminal-resolution evaluations were required than the flat baseline: approximately 4,800 versus 2 million hexagon evaluations.
Takeaways & Limitations
The approach filters noise-induced flickering anomalies while focusing analysis on spatially rooted macro anomalies in an edge-oriented geospatial application.
Abstract
from arXiv · showhide
As an increasing number of critical applications, including environmental, emergency, meteorological, and agricultural, rely on real-time anomaly detection in geospatial data streams, challenges related to the storage, processing, and communication of this data arise. Traditionally, large volumes of data have been sent to centralized processing locations for insight extraction. Given the big data context of these applications, this approach becomes increasingly infeasible as data volume and velocity continue to increase. This paper proposes a lightweight edge-oriented approach for anomaly detection and localization for geospatial data streams. By leveraging the H3 discrete global grid system and a multi-scale drill-down logic, the proposed approach significantly reduces computational overhead, achieving a 99.7\% reduction in evaluations compared to traditional flat-scan methods. Furthermore, by filtering out noise-induced flickering anomalies at lower resolutions, spatially-persistent anomalous signals can be efficiently identified. The results demonstrate that the proposed framework effectively distills massive geospatial data into actionable insights.
I. INTRODUCTION
Geospatial applications face growing storage, communication, security, and anomaly-interpretation challenges as data volume and velocity increase. The paper proposes lightweight edge processing that detects and localizes spatially persistent macro anomalies through hierarchical drill-down.
- Geospatial data comes from diverse platforms and links measurable attributes to physical locations across environmental, atmospheric, and societal applications.
- Raw-data transmission is increasingly problematic because satellites have limited transmission windows, low-power sensors depend on gateways, and data volumes are expanding rapidly.
- Raw-data transmission also creates security risks for environmental monitoring, weather forecasting, disaster response, and agricultural intelligence.
- Distributed edge processing is presented as essential for developing communication-efficient and secure geospatial analytics.
- Geospatial anomaly detection must distinguish spatially rooted macro anomalies from higher-resolution flickering anomalies and account for their onset and duration.
- The proposed method aggregates observations into low-resolution spatial groups, thresholds them against historical observations, and drills down only into anomalous regions for localization.
II. RELATED WORK
Prior geospatial anomaly-detection work addresses multimodality, but edge-feasible deployment and filtering of noise-induced flickering anomalies remain underemphasized. This paper responds with a lightweight hierarchical framework for macro-anomaly detection and localization.
- Existing methods address multimodal anomaly detection, while edge-feasible deployment remains largely overlooked.
- Most existing work does not filter noise-induced flickering anomalies, potentially producing numerous false alarms and large alert volumes in large-scale deployments.
- The paper contributes a lightweight, resource-efficient framework for anomaly detection in geospatial data streams.
- Drill-down logic reduces unnecessary computations by expanding only spatial regions that warrant higher-resolution analysis.
- Hierarchical discrete global-grid partitioning localizes anomaly sources at higher spatial resolutions.
III. SYSTEM MODEL
The system separates edge sensing and analysis from core storage, visualization, and policy functions. This federated design keeps computation at the edge because raw-data transfer is undesirable under privacy and resource constraints.
- The edge region contains sensing, processing, and analysis, while the core provides historical storage, visualization, and policy functions.
- The architecture intentionally isolates centralized core functions from edge computation to avoid transferring raw data.
- The system is federated: computation occurs entirely at the edge, and insights are passed to the core.
- Big-data volume, velocity, variety, veracity, variability, and value shape the architecture, with 1km NDVI data representing approximately 813 million pixels.
B. Resource Efficiency Motivation
Resource constraints motivate federated, event-driven processing that aggregates data and communicates only relevant insights. This reduces transferred data and can improve information freshness for real-time decisions.
- B. Resource Efficiency Motivation: Preprocessing can group pixels, extract aggregate statistics, identify anomalous regions, and transmit areas of concern instead of all raw observations.
- B. Resource Efficiency Motivation: Sending only relevant insights saves communication resources and provides actionable information upon receipt.
- B. Resource Efficiency Motivation: Reduced data transfer lowers susceptibility to transmission errors.
- B. Resource Efficiency Motivation: Edge processing and small alerts can exploit short communication windows and reduce processing-induced staleness, improving Age of Information.
A. Algorithm Description
The algorithm aggregates geospatial observations with H3 and uses hierarchical drill-down to detect anomalies and localize their sources across resolutions.
- Algorithm Description: The drill-down evaluates cells from a starting resolution to a maximum resolution and examines child cells only when a parent is anomalous.This reduces computation by screening large areas at lower resolutions before finer localization.
- Algorithm Description: H3 partitions the world into hexagonal regions that support aggregate statistics and anomaly localization.Each hexagon has six equidistant neighbors, reducing distortion.
- Algorithm Description: Each cell evaluation retrieves points within its boundary, computes aggregate statistics, and compares them with historical values using thresholding.The cell boundary is obtained from its H3 identifier.
- Algorithm Description: The anomaly detector computes historical summary statistics and flags observations outside mean or median thresholds or established extrema.The threshold parameter α controls the mean and median threshold construction.
1 Retrieve points belonging to g at resolution res
The point-retrieval step stops without flagging a cell when its region contains no points or lacks sufficient data.
- 1 Retrieve points belonging to g at resolution res: No points cause the procedure to return the cell g with flag 0.
- 1 Retrieve points belonging to g at resolution res: Insufficient data also causes the procedure to return the cell g with flag 0.
- 1 Retrieve points belonging to g at resolution res: The retrieval step therefore passes forward only cells with usable observations.This follows from the stated return condition for empty or insufficient inputs.
3 Query annual statistics S for those points
The anomaly-detection step compares the current observation with historical thresholds and extrema, then returns the resulting anomaly flag for collection and reporting.
- 3 Query annual statistics S for those points: An anomaly is identified when the current observation exceeds mean or median thresholds or falls outside established minimum and maximum values.
- 3 Query annual statistics S for those points: Detected anomalies are passed to the hierarchical algorithm for collection and reporting.
- 3 Query annual statistics S for those points: The detector returns a binary anomaly flag for the H3 cell and year under analysis.
- 3 Query annual statistics S for those points: Insufficient historical statistics cause the detector to return flag 0.
B. Complexity Analysis
The hierarchical complexity depends on the number of H3 resolutions and branching factor, while its worst case evaluates every anomalous cell across levels.
- Complexity Analysis: 122 starting-resolution hexagons are evaluated when none of them is anomalous.
- Complexity Analysis: The worst-case evaluation count is approximated as 122×PRmax r=Rmin 7r when every hexagon at every resolution is anomalous.
- Complexity Analysis: Although this worst case can exceed a flat highest-resolution scan, it is considered unlikely because geospatial anomalies are spatially sparse.Non-anomalous cells do not require further computation.
C. Experiment Description
The experiment uses yearly 1 km NDVI data, selecting 2014 as a proof-of-concept target and building anomaly thresholds from 2002–2013 observations. The drill-down demonstration spans H3 resolutions 0 through 5.
- Dataset and configuration: The evaluation uses an NDVI dataset with 1 km spatial resolution and yearly observations collected from 2002 onward.The target year is 2014, while historical data from 2002–2013 provides aggregate statistics for threshold construction.
- Dataset and configuration: 2014 was selected as the target year for this initial proof-of-concept evaluation.The selection also allows future cross-referencing with known 2014 geophysical events, although that contextualized validation is outside this work.
- Dataset and configuration: The drill-down demonstration starts at H3 resolution 0 and proceeds to a maximum resolution of 5.At resolution 5, each hexagon represents an average surface area of 253 km2.
V. RESULTS AND ANALYSIS
The hierarchical drill-down approach evaluates anomalous regions progressively instead of scanning every hexagon at every resolution. It substantially reduces computation and reported anomalies while emphasizing spatially persistent macro anomalies over flickering detections.
- Comparison design: The comparison evaluates the proposed drill-down method against a flat approach that scans every hexagon at each resolution.The comparison targets computational efficiency and filtering of flickering anomalies while retaining spatially rooted macro anomalies.
- Computational evaluations: The flat approach grows exponentially across resolutions because each H3 hexagon contains approximately seven hexagons at the next resolution.Figure 2 uses a logarithmic y-axis, making this exponential growth appear linear.
- Computational evaluations: The drill-down method evaluates all 122 resolution 0 hexagons but expands only those exhibiting anomalies at higher resolutions.Non-anomalous low-resolution hexagons and their children therefore avoid further drill-down processing.
- Computational evaluations: Over 99.7% fewer terminal-resolution evaluations are required, with approximately 4,800 hexagon evaluations versus 2 million for the flat baseline.The reduction also directly decreases the time between data collection and anomaly reporting when considering the AoI metric.
- Anomaly filtering: The drill-down approach flags fewer anomalies because cross-scale persistence filters localized noise and flickering anomalies.This prioritizes spatially significant macro anomalies for reporting to the centralized agent.
VI. CONCLUSION
The paper presents a lightweight H3-based hierarchical method that localizes geospatial macro anomalies through multi-scale drill-down. An NDVI 1 km example demonstrates computational efficiency and filtering of noise and flickering anomalies, while broader validation remains future work.
- Conclusion: The proposed method is a lightweight multi-scale hierarchical algorithm for anomaly detection and localization in edge-oriented geospatial applications.It uses H3-based low-resolution grouping followed by drill-down toward anomalous regions.
- Conclusion: H3 partitions the world into hierarchical regions that are screened at low resolution and expanded to localize regional anomalies.The approach focuses subsequent processing on detected anomalous areas rather than the full spatial domain.
- Conclusion: The NDVI 1 km use case demonstrates computational efficiency and effective filtering of noise and flickering anomalies while focusing on spatially rooted macro anomalies.This conclusion is based on the illustrative dataset and does not establish performance across other geospatial sources.
- Future work: Future work will evaluate the approach on other geospatial data sources and varied spatial and temporal resolutions.Planned extensions also include adaptive thresholding and a multi-layer anomaly detection model.