Convolutional neural network for soil microplastic contamination screening using infrared spectroscopy

https://doi.org/10.1016/j.scitotenv.2019.134723Get rights and content

Highlights

  • No standardized method exists to measure the extent of microplastic in soil.

  • Microplastic polymers within soil can be detected with vis-NIR spectroscopy.

  • Microplastic contaminated soil can be screened with CNN model and vis-NIR spectra.

  • It is a rapid and promising method to categorize the extent of microplastic in soil.

Abstract

Microplastics are emerging pollutants that exist in our environment. Microplastics are synthetic polymers that have particles size smaller than 5 mm. Rapid screening of microplastics contamination in the soil could assist in identifying anomalous concentrations of microplastics in the terrestrial environment. Because there is no rule on the maximum concentration limit on how much microplastics can exist within the soil, the concentration of microplastics collected from industrial areas around metropolitan Sydney was used as a baseline. Spectra obtained from the visible-near-infrared (vis-NIR) spectra has been shown to be feasible in predicting microplastics in the soil. Instead of creating a regression model predicting the concentration of microplastic, a classification model for screening was proposed. A convolutional neural network (CNN) model was trained to classify the soil sample into various degrees of contamination based on concentration. We also delved into the CNN model to understand how the CNN model classifies the spectral data input. The model performance was first tested on two levels of classification (contaminated vs. non-contaminated). The model was able to classify the uncontaminated samples into the appropriate class more accurately than the contaminated samples. When the number of classes were gradually increased, the classification accuracy for the higher level of contaminated samples improved. Transfer learning CNN model further improved the classification prediction only on the extremes, but not the intermediate classes.

Introduction

Plastic materials are an emerging contaminant concern in our environment. Since 1907, the first development of modern plastic ‘Bakelite’ by Leo Baekeland using formaldehyde and phenol, mass production of plastics have been enhanced (Thompson et al., 2009). Over the last 40 years, the use of plastic had increased by about 25-fold (Machado et al., 2018). In the middle to high income countries, the share of plastics in municipal waste increased from less than 1% in 1960 to more than 10% by 2005 (Jambeck et al., 2015, Geyer et al., 2017). This abundant plastic debris is highly durable and potentially lasts for a long time in the environment.

Plastic is used extensively in our daily life because they are light, moisture resistant and relatively cheap. It can be extruded, moulded, cast, spun and applied as a coating. The use of this synthetic polymer ranges from cups to components for cars or aeroplanes, with the most substantial use today as disposable packaging (Thompson et al., 2009). Plastic comes in all different shapes and sizes, and depending on the material, could be damaging to human health. The plastic issue originally arises as an aesthetic problem, and further, it was found that it caused choking and entanglement of wildlife (Maso et al., 2003, Barnes et al., 2009, Thompson et al., 2009). Significant scientific evidence has pointed out the environmental impacts of plastics on the marine environment, such as transport of persistent organic pollutants, non-indigenous species to new locations and microalgae associated with red tides (Mato et al., 2001, Barnes, 2002, Maso et al., 2003, Rios et al., 2007). The plastic debris that ends up in the marine environment comes from dumping by recreational fishing and boats, fishing fleet, merchant ships, and litter carried by rivers and municipal drainage systems (Derraik, 2002).

Most of the plastic produced probably ends up in the landfill and becomes a pollutant in the soil, ocean, and waterways. Compared to plastic pollution in the marine environment, pollution in the soil ecosystem has not received as much attention, which can potentially be 4- to 23-fold larger than the contaminants that exist in ocean basin (Machado et al., 2018). On land, most of the plastic inputs come from packaging (Derraik, 2002, Rios et al., 2007, Jambeck et al., 2015). In agroecosystems, the primary source of plastic comes from plastic mulch films and indirect source from the application of biosolids (Ng et al., 2018). Plastic mulch is used extensively to increase crop production by increasing soil temperature, conserving moisture, reducing evapotranspiration, weed control and blemish control of product (Kasirajan and Ngouajio, 2012, Ng et al., 2018). Biosolids are organic-rich materials resulting from the wastewater treatment facility that may contain microplastics. The concentration of microplastic in the agroecosystems are expected to be high because both treated wastewater and biosolids are used in agriculture for irrigation and fertilizer respectively (Mohapatra et al., 2016, Nizzetto et al., 2016).

Various types of plastic polymers along with its common uses are summarized in Table 1. The durability and increase usage of plastics becomes waste management problem and potentially lasting for a long time in our environment. Even plastics which are designed to be biodegradable may not fully decompose since they are affected by other factors such as exposure to light, oxygen, and temperature (Swift and Wiles, 2004).

Currently, there is no standardized method to measure plastics contamination in soil and the environment. The inconsistencies in sampling method result in various sampling units, such as abundance per surface area, abundance to depth, volume in cubic meters and weight ratios (Hanvey et al., 2017). Different extraction techniques (physical separation, density separation, and filtration) also affects the measurement of plastic contaminants collected. The identification and quantification of plastic polymers are typically based on visual sorting with the aid of an optical microscope or infrared spectroscopy. Nonetheless, the optical method is unreliable because the recovery rates are dependent on the colour vibrancy and the size of the plastics (Hanvey et al., 2017).

This work will be mainly focusing on screening microplastics contamination in soil. Microplastics are plastic polymers with particle sizes smaller than 5 mm (NOAA, 2018). These microplastics can be further categorized into two types – primary and secondary. Primary microplastics are microplastics that are intentionally manufactured in small sizes, while secondary microplastics are formed from the fragmentation of macroplastics (>25 mm).

Each plastic polymers exhibit unique infrared spectral signatures (see Fig. 1). Visible-near-infrared (vis-NIR) spectra has been successfully used to predict various physical and chemical properties of soil samples and was found to be useful in quantifying microplastics in soil (Corradini et al., 2019, Paul et al., 2019). Corradini et al. (2019) spiked known concentrations of plastic polymers (LDPE, PET, and PVC) ranging between 1 and 100 g/kg on to the soil and predicted the concentration of plastic polymers based on the vis-NIR spectra data. However, we believe this high concentration (~10%) is too high for real-world scenarios. In a case study by Fuller and Gautam (2016), the median of microplastic concentration found within industrial areas near Sydney was approximately 0.24% with a value ranging from 0.03 to 6.7% (3 to 67 g/kg). This study is different from Corradini et al. (2019) who used the plastics between 0.5 and 1 mm diameter and Paul et al. (2019) who used plastic diameters of <125 µm. The concentration used by Corradini et al. (2019) is quite high (up to 100 g/kg). Furthermore, we also explore the effect of soil types and compare the different spectra absorbances which may affect the prediction quality of the models.

The development of a screening model (contamination classes) instead of a regression model (plastic concentration) is proposed because of the physical characteristics of plastics (solid and light weight particles which do not spread and mix with the soil evenly in comparison with liquid contaminants) and the minimal surface contact area with the vis-NIR probe. In a previous study (Ng et al., 2019), application of convolutional neural network model (CNN) as a regression model to predict various soil properties using spectral data had been demonstrated. Aside from regression models, CNN has been used successfully used to classify 1.2 million images (Krizhevsky et al., 2012). We also explored the use of transfer learning on classifying the degree of contamination in the samples. Transfer learning is an approach in which the parameters of a pre-trained network is utilized and fine-tuned on the target dataset (Pan and Yang, 2010). This method requires less computational time and power. It is also used when the new dataset is relatively small.

The objective of this work is to compare and evaluate the use of train from scratch CNN and transfer learning CNN as a classification model to categorize extent of microplastics (LDPE and PET) contaminated soil samples ranging from 0 to 5%. We first delved inside the CNN layers to have a sneak peek on what the model learns and evaluate the model performance as the complexity increases (more classes).

Section snippets

Sample preparation

Twelve different soil samples were collected from various locations across New South Wales (NSW), Australia. The soil samples range from sandy to clayey with clay content ranging from 7.7 to 62%. Soil samples were first air-dried, then ground and sieved to pass through a 2-mm sieve. The soils were spiked with two types of plastics (PET and LDPE) between the concentration of 0 and 1000 mg microplastics per 20 g of dry soil (0–5% by mass, 0–50 g/kg). For a particular type of soil, the amount of

The effect of plastic on the soil

The vis-NIR spectra of the microplastic spiked soil are shown in Fig. 3. However, it was hard to visualize the degree of contamination from the raw spectra. The continuum removal method was applied to visualize the absorption features of plastic contamination in our samples. Continuum removal is a method of normalizing spectra to allow comparison of individual spectra from a common baseline (Clark and Roush, 1984). Samples of the continuum-removed spectra of the soil samples spiked with both

Discussions

Monitoring the pollution level within the soil is an important step in monitoring soil condition and health. Microplastic contaminants are emerging pollutants exist within the soil. The potential health risks related to exposure to microplastics are not yet understood, and more research is yet to be done. Currently, there were only a few of chemometrics approaches using infrared spectroscopy for microplastics analysis in soil. Spectra acquisition process took less than a minute per sample. Paul

Conclusions

In this study, CNN is used to categorize microplastic contamination in the soil. To our best knowledge, this is the first time that CNN model has been trained to categorize contaminants in the soil. The prediction from the CNN model looks promising for the screening of microplastic contamination in soil.

We found that although it could distinguish the non-contaminated and low contaminated sample from the medium and highly contaminated sample, it could not differentiate the degree of

Declaration of Competing Interest

The authors declare that there is no conflict of interest.

Acknowledgment

This work is funded by the ARC Linkage Project LP150100566, Optimized field delineation of contaminated soils.

References (31)

  • S. Fuller et al.

    A procedure for measuring microplastics using pressurized fluid extraction

    Environ. Sci. Technol.

    (2016)
  • R. Geyer et al.

    Production, use, and fate of all plastics ever made

    Sci. Adv.

    (2017)
  • J.S. Hanvey et al.

    A review of analytical techniques for quantifying microplastics in sediments

    Anal. Methods

    (2017)
  • K.M. He et al.

    Delving deep into rectifiers: surpassing human-level performance on ImageNet classification

    IEEE Int. Conf. Comput. Vision (Iccv)

    (2015)
  • J.R. Jambeck et al.

    Plastic waste inputs from land into the ocean

    Science

    (2015)
  • Cited by (85)

    View all citing articles on Scopus
    View full text