Optimized Classification of Suspended Particles in Seawater by Dense Sampling of Polarized Light Pulses

Abstract

Suspended particles affect the state and vitality of the marine ecosystem. In situ probing and accurately classifying the suspended particles in seawater have an important impact on ecological research and environmental monitoring. Individual measurement of the optical polarization parameters scattered by the suspended particles has been proven to be a powerful tool to classify the particulate compositions in seawater. In previous works, the temporal polarized light pulses are sampled and averaged to evaluate the polarization parameters. In this paper, a method based on dense sampling of polarized light pulses is proposed and the experimental setup is built. The experimental results show that the dense sampling method optimizes the classification and increases the average accuracy by at least 16% than the average method. We demonstrate the feasibility of dense sampling method by classifying the multiple types of particles in mixed suspensions and show its excellent generalization ability by multi-classification of the particles. Additional analysis indicates that the dense sampling method basically takes advantage of the high-quality polarization parameters to optimize the classification performance. The above results suggest that the proposed dense sampling method has the potential to probe the suspended particles in seawater in red-tide early warning, as well as sediment and microplastics monitoring.

Keywords: classification; dense sampling; polarized light pulses; suspended particles.

Figure 1

Schematic diagram of the optical system of the experimental setup.

Figure 2

The dense sampling of polarized light pulses: (a) The temporal signals; (b) polarized light pulse processing algorithm.

Figure 3

The schematic diagram of backpropagation algorithm structure.

Figure 4

The confusion matrix of different polarized light pulses methods: (a) PLP-Ave model; (b) PLP-All model; (c) average accuracy of five models.

Figure 5

The confusion matrix of eight suspended particles classification: (a) PLP-Ave model; (b) PLP-All model.

Figure 6

(a) The accuracy curve of different processing methods; (b) the MSE of different processing methods.

Figure 7

The average accuracy of five polarized light pulses processing methods for signal particles.

Figure 8

Temporal values of a polarized light pulse processed by the PLP-All method and the dash line is the value given by the PLP-Ave method: (a) Intensity values (I); (b) polarization parameter (q); (c) polarization parameter (u); (d) polarization parameter (v).