Identification Accuracy of Cost-Effective Visible (VIS) to Near-Infrared (NIR) Sensor AS7265x on Plastic Waste in the Philippines

Proponent/s Name/s (Last Name, First Name, Middle Initial)

Ma. Tricia Jasmin M. Ocho, De La Salle University, ManilaFollow
Chloe Lizanne R. Bravo, De La Salle University, ManilaFollow
Jennika Love R. Loyola, De La Salle University, ManilaFollow
Kierra Jamielle Velandria Taroma, De La Salle University, ManilaFollow
Leianne Nina Arabelle E. Lampayan, De La Salle University, ManilaFollow

Document Type

Paper Presentation

Research Advisor (Last Name, First Name, Middle Initial)

Cu, Gregory G.

Abstract/Executive Summary

In the Philippines, plastic waste persists as a significant issue caused by the widespread use of single-use plastics and the lack of recycling systems. The lack of proper systems made waste segregation and identification inefficient and disorganized. Past studies explain the use of VIS/NIR in plastic segregation and the potential of spectral frequencies. However, there is limited investigation into the consistency of the AS7265x sensor to address reliability in real-world conditions, where contaminants, variable lighting, and material degradation significantly impact performance. Thus, the study evaluates the effectiveness of the AS7265x triad spectroscopy sensor, incorporating machine learning algorithms. An experimental setup was created using the AS7265x sensor and the Arduino IDE to capture wavelength patterns from six different plastic types within a controlled, light-proof bin. Collected data were then processed using WEKA software with machine learning classifiers Naive Bayes, Multilayer Perceptron, sequential minimal optimization (SMO), IBk (k-Nearest Neighbor), J48, and Random Forest to identify plastics according to their wavelength patterns. The Random Forest model yielded the highest accuracy of 92.06%, while other models showed lower performance. The findings suggest that the Random Forest model can analyze intricate and high-dimensional spectral inputs by ensemble learning, enabling it to categorize plastic types despite their subtle differences. Other lower-performing models, however, have been shown to struggle with non-linear connections and feature correlations. The machine learning-integrated sensor performed at varying levels; therefore, there is a need to explore and select the best-fitting model in real-world settings to achieve optimal accuracy in plastic identification.

Keywords

plastic waste identification; machine learning; AS7265x sensor; WEKA Workbench; Random Forest