GeoCluster Preprocessing Framework Achieves 88.87-Meter Accuracy in Sigfox Device Localization

A hybrid framework named GeoCluster achieves a mean geodesic localization error of 88.87 meters while shrinking the input feature space by over 80 percent in Sigfox low-power wide-area networks. The lightweight and scalable system demonstrates that careful structuring of input data alone yields substantial performance gains in RSSI localization without altering the core machine learning architecture.

Accurate device localization remains a central challenge in low-power wide-area networks such as Sigfox, where RSSI measurements are characteristically sparse, noisy, and highly variable across geographic space. While machine learning regressors, particularly Random Forest algorithms, have shown promise for RSSI localization, their practical accuracy is frequently degraded by unsegmented spatial variability, irrelevant features, and inadequate preprocessing.

To address these constraints, the researchers proposed GeoCluster to couple a Random Forest regressor with three complementary preprocessing steps. The pipeline integrates geographic segmentation (GSEG), gateway selection based on feature importance ranking (GRANK), and cluster-based feature reduction (CFTR) to filter and organize raw network signals.

By systematically structuring input data rather than modifying the core machine learning architecture, the framework achieved an 88.87-meter mean geodesic error alongside a feature space reduction of over 80 percent. The authors emphasize that these results demonstrate the critical role of preprocessing in signal localization and provide an efficient approach for resource-constrained Internet of Things deployments.

The three complementary preprocessing operations specifically target the primary factors that degrade conventional machine learning models. The framework incorporates geographic segmentation (GSEG), gateway selection via feature importance (GRANK), and cluster-based feature reduction (CFTR). Furthermore, cluster-based feature reduction, or CFTR, condenses the remaining dimensions, successfully contracting the overall feature space by over 80 percent.

Experimental evaluation of the hybrid framework confirmed a mean geodesic error of 88.87 meters across the network. By maintaining a lightweight and scalable profile, the approach illustrates that the effectiveness of machine learning models in received signal strength indicator localization depends significantly on careful data preparation.

Subject of Research: Sigfox RSSI device localization and data preprocessing

Article Title: GeoCluster: preprocessing-centric hybrid framework for precision sigfox RSSI localization

Article References: Mansourinejad, M., Moradbeikie, A., Rayani, S., & Keshavarz, A. (2026). GeoCluster: preprocessing-centric hybrid framework for precision sigfox RSSI localization. Cluster Computing, 29(16), Article 941. https://doi.org/10.1007/s10586-026-06740-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06740-3

Keywords: Sigfox, RSSI localization, GeoCluster, low-power wide-area networks, LPWAN, Random Forest regressor, geographic segmentation, GSEG, gateway selection, GRANK, cluster-based feature reduction, CFTR, feature importance, mean geodesic error, spatial variability, Internet of Things, resource-constrained IoT, data preprocessing