Beta Testing an AI-Based Physical Analysis Technology for Microplastic Quantification and Characterization

Abstract

Microplastic pollution is accumulating at alarming rates in the natural environment. New and innovative technologies are needed to help understand the gravity of the global microplastic pollution. In this study, a portable artificial intelligence system using image capture and analysis technology was beta tested to determine its suitability for microplastic quantification and characterization. Many factors were examined, including quantity, colour, shape and appearance (i.e., fragment, pellet, and film), and environmentally simulated (i.e., weathered and humic acid soaked). These were all factors considered. The beta prototype showed a pronounced aptitude for microplastic detection with a clean microplastic detection accuracy of 89% and an environmentally simulated microplastic detection accuracy of 77%. The beta prototype was compact, easy to use, and provided extensive information about the samples through its machine learning algorithm. The beta prototype would be well-suited for both scientific research and citizen science and is ideal for larger (≥0.5 mm) and lighter-coloured microplastic characterization.

1. Introduction

In the modern world, plastic plays a delicate role: it enables untold benefits to society through safe food delivery, life-saving medical practices, and energy- and fuel-efficient transportation and infrastructure, yet its unsustainable manufacturing, usage, and disposal methods are poisoning vital resources—water, soil, and air [1,2,3,4,5]. The global plastic production rate (excluding plastics produced from recycled goods) was 367 Mt in 2020, up from 335 Mt in 2016, and it is expected to continue to increase despite the mounting demands that producers and users be held accountable for their role in global pollution and climate change [6,7]. Plastic pollution is a multifaceted issue that requires the alignment of local, national, and global social, economic, and environmental agencies, meaning fast and coordinated system changes are challenging to enact [7]. A larger emphasis has been placed on citizen science in the last decade—where accessible microplastic (MP) sampling methods and identification and quantification technologies are critical for success [8]. Due to the omnipresence and complexities of solving plastic pollution, citizen science is a key resource that has yet to be fully utilized.

A major proponent of plastic pollution is MP pollution. MPs are defined as any plastic particle ranging from 0.1 μm to 5.0 mm in size [9,10]. MP contamination is ubiquitous, yet the lack of standardized methods for MP analysis makes it difficult to understand the severity of the problem. Quantification and characterization of MPs is critical to monitor the evolution of plastic pollution in the environment, but current MP characterization methods require complex sample collection and pre-treatment with manual sorting using stereomicroscopes, dissecting microscopes, and compound microscopes, which is labour-intensive and time-consuming [9,11,12]. Additionally, these techniques require expensive and cumbersome equipment, which cannot be used in the field, and may not be readily available. Furthermore, they are limited in the number of MPs that can be characterized at one time due to the small area of analysis. Counting and classifying MPs becomes increasingly difficult the smaller they become (0.1 μm–1 mm), but colour, size, density, and appearance (i.e., fragment, bead/pellet, film, fibre, etc.) are commonly used parameters for characterizing larger MPs (1.0–5.0 mm) [9,11,13,14,15].

The beta prototype tested in this study was created with the objective to help quantify, characterize, and ultimately mitigate MP pollution in aquatic and terrestrial environments. Ocean Diagnostics developed a portable AI-based imaging system for rapid physical analysis of MP pollution. At this time, the technology is the only known portable imaging system with a dedicated AI-based web application to analyse MP samples in real-time, but there are many analytical methods paired with AI-based applications and machine learning algorithms to enhance research methods. The simplest form is mobile phone cameras paired with deep learning algorithms, which was the company’s original prototype. Other studies also employed this method where researchers were able to capture images of MPs (1.0–5.0 mm) with digital or phone cameras using 16 megapixels (or higher) and they fed that data into open-source neural networks to automatically classify MPs into different categories (such as size and shape) [12]. The difficulty with this method is accurately training and utilizing the neural networks, which takes significant computational skill, time, and data, and may not be available to the average person. Additionally, many issues arise when utilizing phone cameras—each digital or phone camera is different, and the image capture settings are variable and highly dependent on the user. Training neural networks on such variable data does not lead to accurate representation since the AI may struggle with accurately identifying the same MPs from one image to another. Ocean Diagnostics’ technology differs from other image capture approaches in that the camera and lighting are mounted on a tripod, ensuring predefined settings and stability and preventing model and user variability.

Other studies have demonstrated more advanced but complex approaches, such as combining Fourier transform infrared spectroscopy (FTIR) and Raman spectroscopy with AI to more efficiently catalogue and identify various spectra, as well as critical regions of spectra [16]. Combining infrared analytical methods with AI-based neural networks identifies the chemical composition of the MP as opposed to its physical characteristics; however, this is outside the scope of the beta prototype.

Machine learning models and neural networks, which are a subset of AI, are at the forefront of many research fields due to their automation and continuous availability, as well as their low-cost operations [12,17,18,19], making them highly valuable for MP research. The ability to process large quantities of information and perform repetitive tasks with a high degree of accuracy makes AI an incredible tool. AI allows for an improved understanding of large data sets and automation of complex but well-understood analyses and calculations. The combination of a simplistic image capture technology with machine learning can bridge the gap between citizen science and scientific research while providing important plastic pollution information to the world at large.

The objective of this study was to evaluate the beta prototype of a portable imaging system with a dedicated AI-based web application and test its ability to identify and characterize a wide range of MP samples under simulated, challenging environmental conditions with an emphasis on its usability for basic research and citizen science. Three factors were taken into consideration for the study: (1) quantity, (2) appearance (i.e., fragment, pellet, and mixture of fragments, pellets, and films), and (3) untreated vs. treated (i.e., weathered and soaked in humic acid). Most image capture or visual identification technologies require samples to be clean of organic matter and any other constituents before they can be accurately identified, but this study aimed to test the limits of the beta prototype since sample pre-treatment can be damaging to the sample, as well as energy-intensive and time-consuming.

2. Materials and Methods

2.1. Fragments, Pellets, and Films

Plastic from take-out and produce containers, dish soap bottles, cat toys, nail glitter, and other items from household recycling bins were cut and blended to create MP fragments (0.5–5.0 mm). Spherical, cylindrical, and domed plastic nail art and jewels, as well as plastic craft beads, were used as MP pellets (1.0–5.0 mm). Cling wrap, plastics from wrapped produce, produce bags, and other film-like packaging and wrapping plastics were cut to form MP films (1.0–5.0 mm).

2.2. Humic Acid Solution

A humic acid solution was created from humic and fulvic acid powder (OHD Canada) and water to create a concentrated form. The solution was made by mixing 65 g per litre to make a 6% concentrate, which is frequently used to fertilize agriculture, lawns, and gardens. Humic acids are naturally occurring organic compounds found in soil, water, peat, and coal, and they have significant physical, chemical, and biological implications in aquatic and terrestrial environments [20]. For this reason, humic acid was applied to help simulate environmental MP samples.

2.3. Beta Prototype

The beta prototype is a simplistic design consisting of five components (Figure 1) and a standalone AI web application. The AI web application automatically controls and adjusts the camera and light once the tripod is connected to a Windows computer to ensure optimal image capture and analysis. After a sample is imaged and uploaded to the web application, a surplus of information is provided by the machine learning algorithm. However, the beta prototype only identifies MPs based on their image (i.e., shape and colour); it does not have the capability of detecting the chemical composition of the plastics being imaged. Should detection of chemical composition be required, a secondary more advanced analytical method, such as FTIR or Raman spectroscopy, is necessary.

The AI model for performing categorical characterization of MPs was developed in Python using the SciKit-Learn package. Input data was generated from the beta prototypes’ imaging system, including max width (mm), surface area (mm²), perimeter (mm), convex perimeter (mm), aspect ratio, circularity, convexity, solidity, and colour data and compiled for particles within the categories of Pellet, Fragment, and Line (n = 625). This data was used as training and validation data for the model at a percent ratio of 80:20 using SciKit’s train_test_split function. A 6-layer back propagation neural network was trained on this underlying data and tested on a subset (20%) of the original data for validation. The model showed a 91.3% accuracy rate on the validation dataset, outperforming alternative models such as linear regression, random forest, and k-nearest neighbors by greater than 5%. The neural network was saved for future use and a function was implemented into the main analysis software that could run individual categorical predictions on each new particle detected by the software. The categorical prediction, alongside the confidence percentage, was then appended to data rows in the main analysis files. In the future, a deep learning approach will be implemented for physical image data which should show improved performance over the model trained exclusively on physical size data.

The AI particle categorical prediction, or prediction of MP shape, consisting of fragment, pellet, and line, was also an approach proposed by Hartman [21], where it was suggested that most MPs in the 1.0–5.0 mm range would fall into one of the three specified categories. Fragments are MPs derived from macroplastics, pellets are small beads and primary industrial plastic pellets, and lines are MPs derived from fishing lines and nets, meaning thicker fibres or thin fragments with a high length-to-width ratio.

2.4. Image Analysis

Image analysis is a critical component in determining the accuracy of the beta prototype; an image of the sample was taken prior to the use of the technology to compare the AI-generated image from the beta prototype. The under- and over-counts determine the accuracy and differentiate between the types of errors encountered by the AI.

Step 1: Two images of the MP samples were captured, first with an iPhone XR camera, followed by the beta prototype (Figure 2a,b).

Step 2: Visual comparison was used to determine over-counts (circled in red) and under-counts (circled in green) (Figure 2c,d).

2.5. Experimental Design

Two experiments were conducted: untreated and treated. Both experiments consisted of MPs made from a variety of colours (including clear, white, and black), but were separated by MP appearance and quantity (

Table 1. Experimental design consisting of microplastic (MP) shape and appearance (fragment, pellet, mixture of fragments, pellets, and films), quantity (10 and 50) and classification (untreated and treated) with the corresponding sample number.

Sample	MP Shape/Appearance	MP Quantity
Untreated Experiment		10
1	Fragment
2	Pellet
3	Fragment, Pellet, Film Mixture
4	Fragment
5	Pellet
6	Fragment, Pellet, Film Mixture
Treated Experiment		50
7	Fragment
8	Fragment, Pellet, Film Mixture

). MP appearances consisted of three categories: fragment, pellet, and a mixture of fragments, pellets, and films. The untreated experiment consisted of clean, untouched MPs. The treated experiment consisted of weathered and humic acid-soaked MPs. For the treated experiment, half of the produced fragments and pellets were roughened using sandpaper and nail files to create weathered and matte MPs, which were then soaked in a 6% humic acid solution. Half of the produced MP films were mechanically stretched and crumpled, as well as sanded to create deformed, roughened, and matte MP films, which were then soaked in a 6% humic acid solution. The mixture of roughened MPs was soaked for 24 h in the humic acid solution before being removed and left to dry for 24 h. After the 48-h period, they were analysed.

Detecting fibres was outside the scope of the beta prototype, and they were therefore excluded. Bristles were included as part of the fragment samples. Despite the AI predictions, pellets were very easily detected due to their distinct curvatures and circular shapes, so they were omitted from the treated experiment. Two MP quantities, 10 and 50, were utilized to test the beta prototype’s accuracy between small and large samples.

3. Results and Discussion

The beta prototype showed a pronounced aptitude for MP detection with untreated detection accuracies of 90% and 88% and treated detection accuracy of 77% (

Table 2. Detection accuracy (%) of the beta technology when evaluating multicoloured microplastic (MP) fragments, pellets, and a mix of fragments, pellets, and films, at MP quantities of 10 and 50, and artificial intelligence (AI) MP shape (fragment, pellet, line) prediction accuracy (%) from the beta technology’s neural network machine learning model for each of the samples.

Sample	Detected	Actual *	Accuracy (%)	MP Shape AI e Prediction Accuracy (%)
MP Quantity: 10		Untreated MPs
1	10	-	100.0 a	90 a
2	9	-	90.0 b	78 b
3	8	-	80.0 c	75 c
		Average	90.0 ± 8.2
MP Quantity: 50		Untreated MPs
4	48	38	78.0 a	90 a
5	47	-	94.0 b	40 b
6	50	46	92.0 c	72 c
		Average	88.0 ± 7.1
MP Quantity: 50		Treated MPs
7	43	41	82.0 ad	100 ad
8	52	36	72.0 cd	97 cd
		Average	77.0 ± 5.0
Untreated Fragment Average			86.7 ± 9.6
Untreated Pellet Average			92.0 ± 2.0
Untreated Mixture Average			81.3 ± 8.2
Fragment Average				93.3 ± 4.7
Pellet Average				59.0 ± 19.0
Mixture Average				81.3 ± 11.1

Notes: * Detection was corrected to account for double and triple counts on single MPs, as well as missing/not visible MPs, in generated image. ^a Fragment samples. ^b Pellet samples. ^c Mix (fragment, pellet, and film) samples. ^d Humic acid-soaked samples. ^e Artificial intelligence (AI) utilizing a trained neural network machine learning model on 1000 MPs to determine MP shape. The trained model has three possible MP shape outputs: fragment (includes films), pellet (includes beads and other shapes with a semi-uniform curvature) and line (thin fragment or thick fibre with uniform narrow width (approximately 0.5 mm) and long length (approximately 3×–10× the width)).

). The dirtier MPs were harder to identify based on overall detection accuracy, but sample 4 (50 untreated MP fragments) had a detection accuracy of 78%, whereas sample 7 (50 treated MP fragments) had a detection accuracy of 82%. The similarity in detection accuracy could suggest that dirty MP fragments from environmental samples may be just as easily detected as untreated MP fragments, but further experimentation is required.

The beta prototype was very well-suited for pellet and fragment detection, with overall untreated detection accuracies of 92% and 87%, respectively. The mixed samples had the lowest average detection accuracy (81.3%), which was due to the presence of MP films. Films were more difficult to detect accurately due to their thinness, transparency/translucency, and crumpled or lined surface appearance. The beta prototype camera utilized a bright flash to detect MPs, so any matte, shadowed, lined, or dark portion of an MP did not register or resulted in one MP being seen as many. Roughened or irregular edges, as seen in the films and some of the fragments, were also a deterrent to the detection accuracy. Based on these findings, clean-cut 2-D shapes with distinct circular or rectangular shapes were easiest to detect.

The AI MP shape prediction showed promise with a high degree of accuracy for fragments (93%) but was lower for pellets (59%). The pellets that were misidentified were categorized as fragments, which may be due to the types of MP pellets used in this study (i.e., craft and nail art jewels and beads), which are not typical industrial pellets, but could also be an indication that further machine learning is required to perfect the prediction accuracy of larger quantities of pellets. Lorenzo-Navarro et al. [11], who utilized a similar image capture and analysis technology that incorporated machine learning, also described pellets being misidentified as fragments, as well as fragments being misidentified as pellets, with the reasoning that it was difficult for the AI beta-version software to differentiate between curved or rounded fragments and irregular pellets. Conversely, the study only demonstrated a few instances of AI prediction inaccuracy, deeming the incongruities as unnoteworthy overall. Masserelli et al. [12] also identified erroneous identification of pellets when fragments were rolled or placed on their short end, resembling more spherical morphology to the machine learning algorithm; however, despite this, the AI was able to identify all shapes of MPs to a 90% accuracy.

The beta prototype’s AI prediction accuracy was 100% and 97% for the treated MP fragments and treated MP mixture, respectively, demonstrating a possible higher aptitude for AI detection when the samples are not clean. The only caveat being that a lower detection accuracy may be more typical with treated MPs, which minimizes the influence of the AI MP shape predictions.

MP fragments showed a significant decrease in detection accuracy from an MP quantity of 10 (sample 1) to 50 (sample 4), which can be explained by the addition of more blue fragments made from the same macroplastic source (a royal blue polyethylene terephthalate (PET) bottle). These blue PET fragments repeatedly confounded the beta prototype. Only one of six blue PET fragments was detected in sample 4, which was the same singular blue PET fragment placed and detected in sample 1 (Figure 3a,b, MP 10) and 3 (Figure 5a,b, MP 8). However, the (approximate) 1.5 mm by 2 mm blue PET fragment that was detected was inaccurately categorized as a line by the beta prototype as opposed to a fragment. Similarly, two of the three fragments that were detected in sample 6 were also incorrectly categorized for shape/size. All other coloured PET MPs (excluding black PET MPs) and similarly blue-coloured polypropylene (PP) fragments and pellets were easily detected by the beta prototype. It is possible that there may be an additive or other compound in the blue PET plastic that interferes with the beta prototype, and more investigation is required.

The MP quantity increase from 10 to 50 did not have a significant impact on the overall detection accuracy, as there was no significant difference (p = 0.8) between the detection accuracies of the two untreated experiments. This indicates that large sample sizes that remain within the camera view do not impact the beta prototype’s ability to detect MPs. The MP mixture showed a significant increase in detection accuracy from an MP quantity of 10 to 50, which was not expected. This discrepancy may be explained due to the variability of the mixture samples, which contained a greater variety of shapes and types of MPs for the beta prototype to identify. These samples were randomly picked prior to the use of the beta prototype, meaning that sample 3 happened to have a higher ratio of difficult-to-detect MPs.

Ten randomly chosen, untreated, multicoloured MP fragments were evaluated using the beta prototype, of which all ten MPs were detected (Figure 3). Lorenzo-Navarro [11] proposed that a minimum of 16 megapixels were required to visualize small particles (1 mm), but the beta prototype was able to identify blue bristles (MPs 8 and 9 on Figure 3b), which have surface areas between 0 to 0.5 mm² (9) and 0.5 to 1 mm² (8). Indicating that the prototype can accurately target smaller MPs, such as thinner fragments and thicker fibres (i.e., bristles or fishing nets/lines), despite only having an 8-megapixel camera. However, MP 10, which may be confused as a bristle given the reported size and shape, is a significantly larger (approximately 1.5 mm by 2 mm) blue PET fragment (discussed previously in Table 2) that was inaccurately captured by the beta prototype.

The colours of the fragments were not as accurately detected; however, as MP 2 consisted of two distinct colours and sections, translucent white and yellow, wherein the translucent portion was 1/3 of the MP, but only the yellow component was identified. Additionally, MPs 4 and 6, which make up approximately 20% of the total colour in the image, are green, but the beta prototype identified 50% of the MPs to be green. MP 3, which could be misconstrued as being green from the beta prototype’s image, is a reflective silver MP, and MP 7, which could be misconstrued as brown, is bubble gum pink in colour.

Ten randomly chosen, untreated, multicoloured MP pellets were analysed using the beta prototype (Figure 4). Three darker-coloured pellets were included, which were the most difficult for the beta prototype to detect (MP 6, deep blue and MP 7, deep purple)—one of which was not detected (MP 10, matte black). Black, deep purple, and deep blue, especially those MPs that are matte finish, are nearly, if not fully, undetectable for the beta prototype due to the black background. An alternative would be to provide a background mat with dual colours, such as white and black, ensuring that all variations of darker colours (including black) could be detected.

MPs 8, 9 were identical, gold styrene spheres (approximately) 1 mm in diameter, but due to their spherical curvature, were identified as much smaller dots (Figure 4a,b). Spheres are difficult for any camera with a pinpoint light source or hard directional light (such as the LED light on the tripod) to see since that light cannot reflect around the sphere; it can only detect the portion of the sphere where the light is directed. A larger, softer light source would be required for spherical objects to be fully observed. The remainder of the pellets that were utilized were half domes, providing a larger surface area parallel to the light source, and thus, were more visible.

A combination of ten randomly chosen, untreated, multicoloured MP fragments, pellets, and films were investigated using the beta prototype (Figure 5). Two of the ten MPs were not identified: a deep purple pellet (diameter 2.5 mm) and an iridescent, black pellet (diameter 4.5 mm). Like Figure 4, the darker-coloured MPs could not be detected; however, in this case, the significant size and iridescence of the black pellet made it highly noticeable to the naked eye. Regardless of visibility, all black MPs have the potential to be overlooked by the beta prototype due to the obscurity of the black background pad. However, if the colour of the background pad is changed, black and other very dark-coloured MPs can be detected.

The beta prototype’s AI analysis determined green had the highest occurrence (37.5%), followed by white (25%), and yellow, grey, and blue (12.5% each). The actual colours of the sampled MPs were different from the detected results: MP 1 was yellow, MPs 2 and 5 were white, MP 3 was green, MP 4 was pink, MPs 6 and 8 were blue, and MP 7 was gold. The AI-determined colour distribution did not perform adequately, particularly for the transient colours, as it was unable to detect pink and gold and inaccurately detected grey and the distribution of the remaining colours.

Fifty randomly selected, untreated, multicoloured MP fragments were evaluated using the beta prototype (Figure 6). The size limitations of the beta prototype were tested by using translucent blue and white toothbrush bristles, with impressive results: eight of nine bristles (MPs 37 to 41 and 43 to 45) were detected. The beta prototype may have a high aptitude for fishing line and net MP detection, which would be highly beneficial to coastal and freshwater sampling surveys.

AI analysis issues were more apparent in this sample—two blank areas within the generated image were counted as MPs by the AI technology (MPs 47 and 48). Additionally, the generated image for MP 46 (Figure 5a,b) was not a detected MP but the number “8” from the numbering applied by the AI technology to quantify the MPs within the sample image. MPs 46, 47, and 48 could not be included as viable MP detections since they did not provide reliable information to the user.

Shinier or more reflective MPs interfered with the colour detection accuracy of the beta prototype. The utilization of a camera flash automatically made shiny MPs appear as grey (27.5%) or white (25%), even though grey and white only accounted for 28% of the MP colour in the sample. The reflectivity of the MPs also resulted in individual MPs being identified as multiple MPs by the beta prototype; for example, MPs 19, 36, and 48, MPs 3 and 42, and MPs 30 and 33 were only three MPs in the actual sample, not seven. Curved and irregularly shaped MPs, such as those with multiple planes meeting along lines or points, deceived the AI because the light was reflected differently based on the curvature and angle of the planes in reference to the position of the camera light. The shadowed areas on MPs were seen as blank spaces by the beta prototype, typically resulting in the reflected areas of the MP being identified as individual MPs.

Interestingly, matte transparent/translucent fragments were penetrated by the camera light in such a way that it created shadowed and highlighted areas within the fragment, also altering the AI’s perception of MP size and shape. An example of this is MP 12, which was like MPs 2 and 3 in size and shape but differed in colour (i.e., it was transparent), which caused the penetrating light to be directed out the side of the fragment, ultimately only highlighting the height (thickness) of the MP and not the surface perpendicular to the camera light.

Fifty randomly selected, untreated, multicoloured MP pellets were evaluated using the beta prototype (Figure 7). Due to the size and shape uniformity of the pellets, they were easier for the AI to identify; however, like previous analyses, dark colours interfered with the detection accuracy. The larger sample size magnified the dark colour issue, skewing the generated data; MPs 32 to 47 were not correctly pictured, resulting in significant differences between the actual size and reported size of the MPs (Figure 7a,b). Additionally, MP 38, which was a black, cylindrical pellet on its side, was inaccurately detected as a line. The positioning of the MPs in reference to the camera light will impact their visibility and, therefore, the detection accuracy (especially if the MPs are darker coloured). However, in real MP pellets, there is significantly less uniformity, so it would be unlikely for a pellet to be as perfectly cylindrical as MP 38 and, thus, unlikely for a pellet to be seen as a line.

A combination of fifty randomly selected, untreated, multicoloured MP fragments, pellets, and films were analysed using the beta prototype (Figure 8). The MP variability (i.e., type, size, and shape) within the sample was expected to cause more difficulties for the AI technology but had little impact on its detection accuracy. There was one notable occurrence: MPs 33, 39, and 42 were the same black and silver, S-shaped fragment—again, demonstrating how the curvature or irregularity of an MP may impede the beta prototype’s detection accuracy. Other than overcounting one MP, only two MPs were not identified within the sample, resulting in a detection accuracy of 96%.

Fifty randomly chosen, treated, multicoloured MP fragments were evaluated using the beta prototype (Figure 9). The addition of humic acid resulted in a decrease in colour vibrancy with grey (41.9%) being the most prominent colour detected (Figure 9c), despite the MPs being covered in matte, dark brown humic acid. Most of the humic acid was perceived as grey by the AI analysis, which could mean a higher resolution camera may be needed for the best colour accuracy, or there may be room for improvement with the beta prototype’s colour detection algorithm. However, any dirty samples will interfere with the beta prototype’s colour detection accuracy and must be taken into consideration when analysing untreated MP samples.

Despite the addition of dark brown humic acid to the MPs, the detection accuracy was substantial at 82%. If MP colour was not significant to the user, it could be possible to forego sample pre-treatment (i.e., removal of organic compounds, such as humic acid, from the MPs) while still maintaining a high detection accuracy and accelerating sample analyses.

Fifty randomly selected, treated, multicoloured MP fragments, pellets, and films were evaluated using the beta prototype (Figure 10). The effect of the humic acid was more pronounced in this sample, resulting in a total detection accuracy of 72% and a greater emphasis on the colour grey with an occurrence of 63.5% (Figure 8c). Overall, the vibrancy of the MP colours became muted or undetectable when analysed by the beta prototype.

4. Limitations and Future Applications

One limitation is the beta prototype’s inability to detect fibres, which are one of the most prominent MPs in the natural environment [7]. Fibres, smaller MPs (≤1 mm), and thin MPs got stuck in the mat and blended in with the fibres from the background pad. Bright red fibres that were visible to the naked eye were not distinguished by the current version of the beta prototype, but the advancements in (cost-effective) cameras and the future development of machine learning software may be fast to adapt to these conditions. It should also be noted that the detection of microfibres in environmental samples is a challenge, even for advanced and expensive analytical laboratory instruments.

Additionally, any wind or air movement (including breathing) has the potential to influence the detection of MPs, especially the lighter-density MPs, on the background pad. It is necessary to control air movement when situating the MPs for analysis to ensure accurate image capture and analysis by the beta prototype. At the current time, the beta prototype is compatible with the Windows operating system, which is typical for scientific instruments, and expanding it to the Mac OS would increase its usability and uptake for citizen science projects.

The future of this technology is promising due to its easy user interface, yet advanced technical analysis. A definitive advantage is the simplicity of the design: it is compact and easy to carry, set up, and use. The light-blocking shroud is very effective at blocking light pollution, and the AI analysis provides remarkably accurate information that is continually improving due to the nature of machine learning algorithms. It is accessible not only to the scientific and research communities but to the general population as well. This study’s iterative approach to testing and technology development contributed to a commercially available model of the prototype called the Saturna Imaging System.

Citizen science is a critical factor in the fight against plastic pollution, and this technology could find many opportunities in local communities, municipalities, schools, and academic institutions. Elementary, secondary, and undergraduate schools could generate extensive information on plastic pollution by utilizing technologies like this in their everyday science and laboratory classes. Data collection is critical in understanding the severity of the plastic pollution problem, and the data generated with this technology could contribute immensely to MP information databases, minimizing the knowledge gaps in the field. Experimentation done with the beta prototype can help determine the occurrence and quantities of MPs in many areas and regions that have not yet been researched or provide further intel on areas of focus. Characterizing the pollution (i.e., the MPs) in the environment is the first step to mitigating MP pollution since the transport, fate, and, ultimately, its source may be gleaned from the analysis.

There is a huge knowledge gap in the occurrence and quantification of MP pollution in local environments, and simple, accurate, and inexpensive instruments are needed to fill this knowledge gap with the support of scientists and the broader public.

5. Conclusions

The beta prototype tested in this study showed excellent potential in MP analysis, particularly for those that are larger than 0.5 mm and with lighter colours. It is simple to use and provides advanced information for the user. Its machine learning algorithm and high detection accuracies of 90% and 88% (untreated MPs) and 77% (treated MPs) make this technology very valuable to the scientific community and general population alike. Its potential contribution to MP information databases, specifically relating to occurrence and quantification, could be significant to minimizing the knowledge gaps in the field and creating a global roadmap of MP pollution sinks and, ultimately, MP sources.