Status : Verified
| Personal Name | Mendoza, Julius C. |
|---|---|
| Resource Title | Enhancing Regression-Based Calibration Pipelines for Low-Cost Air Sensors through Data Partitioning and Spatial Transferability |
| Date Issued | 15 June 2026 |
| Abstract | Low-cost sensor (LCS) networks o[er scalable air quality monitoring, but their accuracy is frequently hindered by physical artifacts such as humidity-driven hygroscopic aerosol growth and internal thermal inflation. While machine learning calibration against reference monitors improves data quality, the need for prolonged colocation and scarce reference monitors limits scalability in resource-constrained areas like the Philippines. This study addresses these operational barriers by optimizing colocation periods, evaluating data partitioning strategies, and testing model transferability across highly divergent urban micro-environments. An empirical evaluation established a parsimonious 3-feature input configuration (native PM2.5, temperature, and relative humidity) as the optimal architecture, successfully capturing diurnal thermodynamics while avoiding the detrimental multicollinearity introduced by explicit temporal covariates. Utilizing a random block partitioning strategy, continuous models like Gaussian Mixture Regression (GMR) and Support Vector Regression (SVR) outperformed tree-based ensembles. By mapping fundamental physical relationships rather than memorizing temporal noise, this approach preserved peak localized accuracy (e.g., S04 MAE = 3.29 μg/m3) while softening the mathematical divergence typically experienced during extrapolation. Temporal optimization revealed that maximizing training data volume actively degrades predictive accuracy due to environmental concept drift. Optimal algorithmic stability and resource efficiency were achieved by strictly bounding the colocation training window to approximately 31 days, demonstrating that multi-month colocations are operationally inefficient. To evaluate spatial generalizability, a cross-deployment validation matrix was constructed across a university campus network and an external Environmental Management Bureau (EMB) station in Makati. The empirical results definitively demonstrated the physical limit |
| Degree Course | MS Environmental Engineering |
| Language | English |
| Keyword | air quality monitoring, low-cost sensor, machine learning, calibration, colocation |
| Material Type | Thesis/Dissertation |
Preliminary Pages
567.30 Kb
Category : F - Regular work, i.e., it has no patentable invention or creation, the author does not wish for personal publication, there is no confidential information.
Access Permission : Open Access