Optimizing Instrumental Odour Monitoring Systems in Drones by Feature Selection for Odour Detection and Odour Concentration Estimation

Abstract

Fugitive odour emissions from wastewater treatment plants (WWTPs) present ongoing analytical and environmental challenges. Drone-mounted Instrumental Odour Monitoring Systems (IOMS) enable real-time, spatially resolved chemical sensing; however, large sensor arrays increase calibration complexity and cost. To address this, IOMS optimization is formulated as a machine-learning feature-selection problem. A two-stage selection strategy is introduced, combining Sequential Forward Selection (SFS) and Interval Partial Least Squares (iPLS) regression to identify minimal, information-rich sensor subsets and optimal temporal measurement windows. The approach is evaluated using data from a hexacopter-borne IOMS equipped with 21 sensors operating over an active WWTP. Sensors are ranked according to their incremental contribution to odour-concentration prediction error reduction, followed by refinement of measurement intervals to capture relevant temporal dynamics. Validation on independent flight data demonstrates that a configuration comprising only three sensors with optimized time windows retains or improves predictive performance relative to the full array. For quantification, the Bland–Altman limits of agreement improve from ±7 to ±5.3 dBod, and the Pearson correlation increases from 0.80 to 0.89. For odour-detection task, a single sensor achieves an AUC of 0.95, slightly outperforming the full sensor set (AUC = 0.93). Bootstrap analysis reveals variability in feature selection, though consistent trends are observed: ammonia sensors dominate quantitative models, whereas low-temperature MOX sensors are preferred in detection. The findings demonstrate the effectiveness of feature-selection strategies in simplifying IOMS hardware while preserving chemometric performance.