Identifying optimal microwave frequencies to detect floating macroplastic litter using machine learning

Abstract

Microwaves (MWs) have emerged as a promising sensing technology to complement optical methods for monitoring floating plastic litter. This study uses machine learning (ML) to identify optimal MW frequencies for detecting floating macroplastics (>5 cm) across S, C, and X-bands. Data were obtained from dedicated wideband backscattering radio measurements conducted in a controlled indoor scenario that mimics deep-sea conditions. The paper presents new strategies to directly analyze the frequency domain signals using ML algorithms, instead of generating an image from those signals and analyzing the image. We propose two ML workflows, one unsupervised, to characterize the difference in feature importance across the measured MW spectrum, and the other supervised, based on multilayer perceptron, to study the detection accuracy in unseen data. For the tested conditions, the backscatter response of the plastic litter is optimal at X-band frequencies, achieving accuracies up to 90% and 80% for lower and higher water wave heights, respectively. Multiclass classification is also investigated to distinguish between different types of plastic targets. ML results are interpreted in terms of the physical phenomena obtained through numerical analysis, and quantified through an energy-based metric.