Remote estimation of plastic distribution in California agriculture using Sentinel-2 imagery: a case study

Global application of agricultural plastic mulch is increasing. While these films offer benefits for crop production, they also are a large source of plastic waste and pollution in both terrestrial and aquatic systems. The current scale of global mulch application and its deleterious impacts are unclear. Current estimates of agricultural plastic use remain limited by field-specific growing patterns and data limitations. Satellite spectroscopy may offer a means to reduce this uncertainty and model agricultural plastic use on large spatial and temporal scales, though temporally variable crop canopy growth and spectrally diverse plastic targets complicate this approach. To build toward improved plasticulture estimates, this study presents an interpretable, spectral-based classification of bulk plastic use in a California agricultural site using Sentinel-2 imagery. In situ reflectance measurements of soil and white, green, and black plastic mulch endmembers were collected using a Spectra Vista Corporation HR-1024i Spectroradiometer from 350–2500 nm. A combination of field observations and high spatial resolution NearMap imagery (n = 416 spectra) were compiled and divided into 60% training and 40% testing data using a stratified sampling approach. Informed by field reflectance data, mixture analyses, and Sentinel-2 reflectance imagery, an empirical decision tree classification was parameterized to classify dark and bright plastic debris from soil and crop (accuracy = 0.982, kappa coefficient = 0.973, macro-averaged F1-score = 0.967). Modeled agricultural plastic distributions were then retrieved through a time series of Sentinel-2 imagery over Santa Maria, CA, from April 2023 to January 2025, elucidating patterns in plastic application and crop canopy growth without the need for prior knowledge of field growing and management patterns. These results are encouraging for the utility of optical remote sensing in modeling plastic use patterns and identifying contamination hotspots throughout California and, potentially, similarly mulched agroecological systems.

Agricultural systems extensively use synthetic plastic polymers in practices collectively referred to as plasticulture (FAO, 2021; Hofmann et al., 2023). Plastic films—primarily made of low-density polyethylene—are the most common plasticulture tool (FAO, 2021), replacing glass in greenhouses, employed to wrap silage, and overlaying soil as a ‘mulch’ to support weed suppression, temperature and moisture modulation, and disease control (Jiang et al., 2017; Sanchez-Hernandez, 2019; Zhang et al., 2022). While plastic mulch films offer important benefits for crop production, their single-season in-field lifespan, rapid weathering, and challenges for complete removal due to soil adhesion also make these films the largest source of macro- and microplastic agricultural soil pollution (Hofmann et al., 2023) and increase the risk for terrestrial to aquatic plastic particle transfer (De Souza Machado et al., 2018; Qi et al., 2020; Weber et al., 2022). These films also pose significant challenges to waste managers, as weathering and contamination can impede recyclability (Sarpong et al., 2024). The growing dependence on plastic film for agricultural production is thus an important and increasing environmental concern.

While numerous estimates of the magnitude of annual plastic mulch use and waste disposal have been put forth—estimating a global minimum of 2.5 million tons—crop- and region-specific use patterns, variation in data availability and reliability, and rapidly increasing utilization limits their accuracy (FAO, 2021). For example, most information regarding plastic mulch use patterns is derived from data collected through various voluntary surveys of growers (e.g., Goldberger et al., 2019) or industry reports (Wang et al., 2025). Further, the same agricultural land may host multiple crops and be utilized by multiple farmers in a single year. Such complexities can thus result in incomplete estimates of the historical and current application of plastic mulch based on grower surveys alone. Our limited understanding of the extensification (spread) and intensification (applications per year in a given area) of plastic mulch use challenges the ability of stakeholders and policymakers seeking to understand the current status of plastic mulch use and ultimately develop policies that reduce the externalities associated with this technology.

Satellite spectroscopy offers the potential to improve our understanding of plastic mulch use patterns and identify potential hotspots for contamination risk at high spatial and temporal scales. In many regions, the distribution of semi-permanent plastic-covered greenhouses has been readily retrieved using diverse satellite-mounted sensors (e.g., Aguilar et al., 2022; Aguilar et al., 2021; Ma et al., 2021; Sun et al., 2021; Veettil and Xuan, 2022). Beyond greenhouses, however, retrievals of plastic-mulched farmland are complicated by temporally variable obscuration by crop canopy growth (Lu et al., 2015), subpixel mixing with reflectance by soil and crops, and high land use overturning. Studies often address these uncertainties by integrating region-specific labeled data or phenological patterns (e.g., Lu et al., 2025; Xiong et al., 2019) often at the cost of transferability to other agro-ecological systems. Further, non-white plastic mulches (e.g., black, green, silver) are not commonly included in remote sensing classification due to spectral similarity with common agricultural targets (Veettil et al., 2023), yet these materials are being increasingly employed for both agroecological (Al-Karaghouli et al., 1990) and aesthetic reasons (Conrad, 2024). The optimal way to resolve the challenges of variable subpixel mixing and crop canopy growth, while also considering the diversity of plastic mulch colors and compositions, remains an important open question in plastic mulch mapping.

To improve understanding of agricultural plastic use and build toward improved remote plasticulture estimates, we present an interpretable, spectral-based classification of agricultural plastic material using Sentinel-2 imagery. The objective of this study is not to create the highest accuracy retrieval of plastic use, but rather to offer a scalable, physically meaningful classification approach that is suitable for California agriculture and requires minimal training data. To this end, we address the following questions using Santa Maria, California–a hotspot of agricultural plastic film use–as a case study: with what accuracy can: 1) plastic mulch be detected at various stages of crop phenology using multispectral satellite imagery? 2) dark plastic mulch be correctly distinguished, in addition to bright mulch and greenhouses? 3) the seasonality of plasticulture be mapped and monitored using this approach?

Sample spectroscopy

To inform algorithm development, laboratory reflectance spectra of plastic samples were measured using a Spectra Vista Corporation (SVC) HR-1024i Spectroradiometer over the spectral range 350–2500 nm. Band spacing is 1.5 nm from 350 to 1000 nm, 3.8 nm from 1000 to 1890 nm, and 2.5 nm from 1890 to 2500 nm (Spectra Vista Corporation, 2012). White, black, and green plastic samples were collected in spring 2023 from 5 central and southern California agricultural sites and transported to San Diego State University for laboratory measurement. Residual dust and debris were removed using a KimWipe™ prior to analysis. Samples were secured using an SVC LC-RP Pro Leaf Clip and Reflectance Probe. Prior to each measurement, radiance spectra of a white Spectralon ® diffuse reflectance standard were collected, followed by ten radiance measurements of each sample. Reflectance spectra were computed by dividing sample radiance by white panel radiance (Equation 1).

Reflectance spectra of soil samples from 5 fields hosting plastic mulches were also recorded. Soil samples were illuminated by a Tungsten-halogen lamp light source. To minimize self-shadowing, the operator held the 15° field-of-view fiber optic in a nadir-looking configuration with the source oriented 45° off-nadir. Prior to each measurement, radiance spectra of a Spectralon diffuse reflectance standard were collected, followed by ten radiance measurements of each sample. Final reflectance spectra were computed by dividing the average of the 10 sample radiance scans by white panel irradiance. Sample plastic and soil reflectance spectra were further convolved to the spectral response function of Sentinel-2a to evaluate the extent to which observed spectroscopic differentiability was robust to degradation of spectral resolution.

Spectral metrics

Broadly informed by sample spectral properties (Section 2.1), an empirical decision tree classification provided separation among land cover classes. Dark plastics, a grouped class of black and green LDPE plastic mulch (see Section 2.4), were first isolated using reflectance magnitudes in the shortwave infrared (SWIR). Bare soil pixels were next identified by their uniquely positive spectral slope from 800–1600 nm compared to remaining unclassified pixels (i.e., crop and white plastics) (Figures 1, 2). White plastics were lastly distinguished from crops via broad spectral shape differences as indicated by Apparent Visible Wavelength (AVW). These metrics are expanded upon below.

Initial classification of dark plastic pixels was derived from the sum of Sentinel-2 SWIR bands. Spectra of soil and dark plastic pixels do bear similar spectral shapes in the infrared (865–1610 nm), though the overall magnitude (albedo) of infrared reflectance is highly dissimilar. While erroneous atmospheric correction could introduce confounding bias in reflectance magnitudes in the SWIR, atmospheric scattering is less pernicious at SWIR wavelengths than shorter visible wavelengths, and the difference in observed reflectance is sufficient (on the order of 30%–40%, as shown in Figure 1) to consider this metric generally robust to typical uncertainties associated with most atmospheric corrections.

Following classification of dark plastic and soil pixels, remaining pixels were classified as crop or white plastic cover using AVW. Originally developed for aquatic systems, AVW is a weighted harmonic mean of remote sensing reflectance (R rs) across the electromagnetic spectrum, given as:

AVW reduces the entire spectrum to a one-dimensional metric of color that is correlated to spectral shape (Vandermeulen et al., 2020) (Equation 2). Here we use AVW both as a metric of spectral variability among different plastic samples (Figure 1) and as a classification index to separate land cover classes. AVW is traditionally calculated using only the visible region from 400–700 nm, but in this case, we apply AVW to the full visible-shortwave infrared (VSWIR) spectrum (350–2500 nm). This study represents a novel application of AVW on surface reflectance imagery for terrestrial applications.

Normalized Difference Vegetation Index (NDVI) is a common spectral metric employed to discriminate vegetation from non-vegetated targets (Rouse et al., 1973), though it does not uniquely separate all four target classes in this case. Dark green plastic mulch exhibited a sharp increase in reflectance at red and near-infrared (NIR) wavelengths overlapping with that of vegetation (Figures 1, 3), while the low red and NIR reflectance of black plastic can make the normalized ratio sensitive to small radiometric or atmospheric errors. Bare soil and white plastic may also occupy overlapping low-NDVI ranges. The combination of SWIR albedo, the 865/1610 nm reflectance ratio, and AVW provided greater separability among the four crop, soil, white, and dark plastic classes.

Mixture analysis

Computations of linear spectral mixtures of plastic, soil, and crop spectra were used to simulate the effects of substrate mixing and crop overgrowth on endmember separability. Convolved sample plastic spectra (Figure 1) were used to compute linear spectral mixtures of white, black, and green plastic mulch in combination with either soil or vegetation in deciles from 0% to 100% of each material. A previously published spectral library of healthy crop vegetation was used for the simulated crop mixtures; the specifications of this dataset are outlined in Ely et al. (2018). The spectral metrics used for classification (Figure 4; Table 1) were then computed for each set of mixtures to quantify the subpixel fractions at which plastics were separable from soil or crop using this spectral approach (Figures 3, 5).

Following a similar approach to Sousa and Small, 2017, linear endmember mixtures were computed with variable atmospheric scattering to assess the robustness of spectral metrics to atmospheric perturbations (Supplementary Figures S2, S3). A shade endmember, represented as a Rayleigh scattering function of shape r α λ −4, was computed for three increasingly opaque atmospheric scenarios with maximum magnitudes of 0% (no atmosphere), 1% (low atmosphere), and 10% reflectance (high atmosphere). For 300 Monte Carlo iterations, the actual magnitude of Rayleigh scattering was randomly selected from a uniform distribution between zero and the specific maximum value (0, 1, 10). Each Rayleigh function was then linearly mixed with all possible combinations of plastic/soil or plastic/crop in 1% increments. Monte Carlo iterations were repeated for each atmospheric scenario. For every mixed spectrum, decision tree spectral metrics were then computed, illustrating potential classification bias introduced by atmospheric perturbations.

Classification and validation

A time series of atmospherically corrected cloud-free Sentinel-2a and 2 b L2A surface reflectance imagery from April 2023 to January 2025 over Santa Maria, CA was retrieved for classification (Figure 6). Images were downloaded from the Copernicus Data Space web portal. All reflectance bands were resampled to 10 m resolution pixels using bilinear interpolation. Non-agricultural pixels were masked using the California Department of Water Resources (DWR) Statewide Crop Mapping GIS Shapefile (California DWR, 2024).

Plastic endmembers were categorized differently in Sentinel-2 image classification than during in situ spectroscopic analyses. While in situ “white plastic” spectra only included low-density polyethylene (LDPE) plastic mulch, the “white plastics” category in NearMap and Sentinel-2 imagery was expanded to include white and clear plastic mulch, green houses, hoop houses, and other semi-permanent “bright” plastic debris, representing a mixed-use category. Spectral features of in situ white plastic mulch reflectance spectra were generally retained with this mixed-use white plastics grouping (Figures 1, 2). The spectral separability of black and green LDPE plastic mulch also degraded when observed in Sentinel-2 imagery (Figure 2). These materials were therefore combined into a single “dark” plastic endmember for image classification.

A reference dataset was compiled using a combination of NearMap aerial imagery (n = 322), in situ observations (n = 8), Google Street View records (n = 13), and points extracted from Sentinel-2 imagery (n = 73). The majority of reference data was derived from a high spatial resolution NearMap image over Santa Maria, CA on 05 July 2021. The spatial resolution of this dataset allows plasticulture to be clearly and distinctly identified by eye. Polygons from the 2021 California Statewide Mapping GIS Shapefile were then classified as either predominantly crop, soil, or white plastic based on overlying NearMap imagery. Perimeters of classified polygons were eroded by 10 m to minimize edge effects. Of note, NearMap imagery over Santa Maria, CA were primarily available during the summer season, when vegetation overgrows underlying black and dark green plastic mulch, resulting in substantial spatial mixing with crop canopies (Figure 7; Supplementary Figure S7). Consequently, “pure” unmixed dark plastic mulch spectra were not readily observed in NearMap imagery. Dark plastic validation points were instead acquired from in situ observations of agricultural sites beyond Santa Maria, CA, Google Street View records, and Sentinel-2 imagery directly. Of the final dataset (n = 416), points for each class were divided into 60% training (n = 252) and 40% testing (n = 164) datasets for use in model development and validation respectively using a stratified sampling approach. Classification thresholds (Figure 4; Table 1) were empirically selected based on histogram separability among the training dataset (Supplementary Figures S4–S6). Once validated, the empirical decision tree classification of crop, soil, and white and dark plastic debris was applied over Santa Maria, CA from April 2023 – January 2025 (Figures 4, 8, and 9).

To compare the proposed spectral decision tree with conventional machine-learning approaches, random forest and radial-basis-function support vector machine classifiers were trained using the same stratified training dataset (n = 252) and evaluated using the same held-out test dataset (n = 164; Supplementary Table S2). The three spectral metrics used by the decision tree (mean SWIR albedo, the 865/1610-nm reflectance ratio, and AVW) were used as predictor variables for all classifiers. Machine learning hyperparameters were selected using five-fold stratified cross-validation confined to the training dataset. SVM predictors were standardized independently within each cross-validation fold using a StandardScaler pipeline. The selected random-forest model used 300 estimators, square-root feature selection, unrestricted tree depth, a minimum split size of two, and balanced class weights. The selected RBF-SVM used C = 10, gamma set to scale, and unweighted classes. Computational performance was evaluated using 1,000,000 valid agricultural pixels randomly sampled from a representative Sentinel-2 image acquired on 12 November 2023. Prediction time was measured over five repeated runs after predictor calculation, and all classifiers were evaluated using the same feature matrix and computing environment.

Field spectra

White, black, and green plastic mulch samples have distinctive spectral shapes relative to each other (Figure 1). As expected, white plastic reflectance spectra are the brightest of all samples, with maximum values observed near 30%–50% reflectance at visible wavelengths in the 500–600 nm range. White plastic spectra monotonically decrease in reflectance from 600–2500 nm. Black plastic reflectances are spectrally flat to within 1%–3% throughout the entire VSWIR range, with subtle but observable broad peaks centered around 600 nm and troughs centered around 2400 nm. These low reflectance magnitudes resemble those of shadow and mud, presenting a serious non-uniqueness challenge in spectroscopic remotely sensed imagery. In contrast, green plastic reflectance spectra show an absorption feature centered near 675 nm and a sharp increase in reflectance at short NIR wavelengths ∼700 nm, reminiscent of the “red edge” observed in reflectance spectra of photosynthetic vegetation.

Plastic reflectance spectra are often characterized by narrowband absorption features at infrared wavelengths, though these features are not observed in sample spectra of white or black plastic mulch and only weakly observed in spectra of green plastic mulch (Figure 1). These narrowband absorption features are often leveraged to model the distribution of plastic debris using hyperspectral airborne and satellite imagery [e.g., (Castagna et al., 2023; Kühn et al., 2004)] but are not observed in our plastic debris samples (Figure 1). As a result, dominant broadband spectral features observed in hyperspectral resolution are retained after convolution to the spectral response function of the multispectral Sentinel-2a instrument for all three plastic mulch colors (Figure 1). This material is thus well-posed for algorithmic development relying on finer spatial and temporal resolution offered by multispectral satellite constellations.

Sample soil reflectance spectra have high variability in magnitudes, though overall spectral shapes and absorption features are consistent among all five locations. Generally, all soil reflectance spectra have high slopes from 350–1400 nm. All soil reflectance spectra have three absorption features centered near 1415, 1915, and 2200 nm. The depths of these features vary between locations. These absorption features are directly proportional to the levels of clay, silt, and sand present in each soil sample (Sørensen and Dalsgaard, 2005). While overall spectral shape and rates of increase are consistent in both hyperspectral and multispectral resolution, these three soil infrared absorption features are degraded in multispectral resolution (Figure 1).

In both hyperspectral and multispectral resolution, the full plastic and soil libraries are entirely separable with AVW. Soil spectra exhibit less variability in AVW between locations than plastic samples. In hyperspectral resolution, the entire plastic mulch spectral library has an AVW less than 958 nm while the entire soil spectral library has an AVW greater than 1189 nm. (231 nm difference). High AVW separability is consistent in multispectral resolution as well (126 nm difference). Decreased spectral resolution consequently decreased AVW values for all spectra, particularly for soil samples. Of note, spatial heterogeneity, atmospheric correction, bi-directional reflectance distribution function (BRDF) differences, and other confounding variables will impact separability with AVW in remote sensing imagery, though idealized laboratory measurements of plastic and soil samples are entirely separable.

Mixture analysis

Increasing plastic content in endmember mixtures produces systematic, material-dependent changes in spectral shape (Figure 3). At high mixture amounts, white plastics increase in reflectance at visible wavelengths (∼500–600 nm) by > 0.35 and decrease in reflectance at SWIR wavelengths (2000–2500 nm) by > 0.20. Black plastics have low magnitude reflectances of approximately 0.03–0.06 through the full VSWIR range. Consequently, strong differences in spectral shape and magnitude are observable when mixed with comparatively bright vegetation and soil, even at high plastic decile mixture amounts. Mean green plastic mulch reflectances are spectrally flat at a magnitude of approximately 0.13–0.15 from 800–2500 nm. These spectra markedly decrease in spectral slope from 800–1600 nm when mixed with vegetation and increase in slope when mixed with soil, while additionally degrading the near-infrared “red edge” feature.

Linear mixture analysis demonstrates moderate endmember separability using mean SWIR albedos, infrared band ratios, and AVW (Figure 5). Black plastic mulch is separable from pure soil and crop spectra above a 60% subpixel fraction using SWIR bands. Similarly, dark green mulch is detectable at >60% subpixel fractions when mixed with crops, though when mixed with soil, green plastic mulch is only detectable at high amounts (>90%). While white plastics have similar SWIR reflectance magnitudes to those of black and dark green plastics in sample reflectance data (Figures 1,3,5,,), white plastic SWIR reflectance magnitudes are greater when observed in Sentinel-2 imagery (Figure 2) and would likely not be detected using this metric alone. Among remaining unclassified pixels, soil is entirely distinguishable from pure crop spectra and from >30% subpixel white plastic mixtures using a reflectance ratio at 865 and 1610 nm. From there, >40% white plastic mixtures are separable from pure crop using AVW. In sum, the spectral metrics presented can readily distinguish plastic endmembers from those of pure crop and soil at high plastic subpixel amounts, particularly during initial plastic application. It should be noted that these thresholds are representative of idealized spectral mixtures and are subject to change when scaled to spectroscopic imagery due to uncertainties in atmospheric correction, BRDF, and 3-dimensional geometry.

Endmember separability in idealized mixtures is retained under simulated atmospheric perturbations (Figure 5; Supplementary Figures S2, S3). SWIR albedos are the least robust metric to atmospheric noise and while black plastic mulch is still separable at >60% subpixel fractions, increasing shadow rapidly degrades separability of dark green plastic mulch at < 90% subpixel amounts (Supplementary Figure S2). This suggests that dark green plastic mulch may have a narrow window of detection in Sentinel-2 imagery following initial mulch application and under well-illuminated conditions. Soil retains high separability from crop and >50% white plastic subpixel fractions even under high atmospheric noise (Supplementary Figures S2, S3). White plastic similarly retains high separability from crop using AVW under high atmospheric perturbations at >30% subpixel amounts (Supplementary Figure S3). It should be noted, however, that the AVW threshold at which these two classes are discriminable is at a much lower wavelength (∼750 nm, Supplementary Figure S3) in endmember mixtures than in Sentinel-2 training data (∼815 nm, Supplementary Figure S6). This may be attributed either to the scarcity of unmixed crop reflectance spectra in the Sentinel-2 training dataset or to elevated SWIR reflectances for white plastic pixels in Sentinel-2 imagery (Figure 2) compared to sample spectra (Figure 1). Even so, crop, soil, white plastic, and black plastic endmembers all demonstrate separability in idealized mixtures and under simulated atmospheric perturbations, though the detectability of dark green plastic mulch may be limited to well-illuminated, high-density mulch applications.

Validation

Dominant spectral features among pure and mixed endmember reflectance spectra persist in Sentinel-2 imagery (Figure 2). White plastic reflectance spectra peak in the visible range near 500 nm and decrease in magnitude through the infrared. Soil spectra increase throughout the infrared and peak ∼1600 nm, consistent with that of convolved sample spectra (Figure 1). Dark plastics, a grouped category of black and green plastic mulch, are strongly absorbing in the visible range, with greater variability observed in the infrared. This SWIR variability may be attributed, in part, to variable minerogenic “dust” overlying the plastic mulch, increasing SWIR reflectance akin to a linear mixture of soil and dark plastic reflectance. The combination of black and dark green LDPE plastic mulch into one endmember category may also contribute to this variability in the infrared range. Crop spectra are consistent with that of global vegetation (e.g., Small and Sousa, 2022), characterized by high absorption in the blue (<480 nm) and the red (∼675 nm), a reflectance peak ∼550 nm, a strong “red edge” increase in NIR reflectance ∼700 nm, and a decrease in NIR reflectance at wavelengths greater than 800 nm.

A four-class confusion matrix demonstrates strong overall classification performance (Figure 7). The classification yielded an overall accuracy of 0.982 and a Cohen’s kappa coefficient of 0.973, indicating agreement between observed and retrieved classes beyond statistical chance. High class imbalance exists, however, due to a lack of pure unmixed vegetation and dark plastic mulch spectra in available summertime NearMap imagery. Consequently, crop spectra had the lowest precision (0.83) and F1 score (0.91) due to commission errors with ground truth soil and dark plastic pixels. Balanced accuracies for each class were found to be 0.994 for crops, 0.987 for soil, 0.984 for dark plastics, and 0.992 for white plastics. Per-class confusion matrix statistics are detailed in Table 2 including computed sensitivity, specificity, positive and negative prediction values, precision, recall, F1 scores, prevalence, detection rates, detection prevalence, and balanced accuracies. Reported accuracies should be interpreted as an upper bound given partial spatial and temporal overlap between training and validation datasets.

The spectral decision tree performed comparably to the conventional machine-learning classifiers evaluated using the same held-out test dataset (Supplementary Table S2). The spectral decision tree achieved an overall accuracy of 0.982, multiclass balanced accuracy of 0.984, Cohen’s kappa coefficient of 0.973, and macro-averaged F1 score of 0.967. The random forest and RBF support vector machine both achieved slightly higher overall accuracies of 0.988 and kappa coefficients of 0.982. The random forest had a slightly lower multiclass balanced accuracy than the spectral decision tree, 0.982 versus 0.984, but a higher macro-averaged F1 score of 0.980. The RBF support vector machine achieved the highest multiclass balanced accuracy and macro-averaged F1 score, with values of 0.992 and 0.990, respectively. Overall, differences in predictive performance among the three approaches were small. The spectral decision tree had substantially faster inference, requiring 0.0240 ± 0.0043 s per 1,000,000 pixels, compared with 2.0880 ± 0.0827 s for the random forest and 2.6512 ± 0.1337 s for the RBF support vector machine.

Plastic distribution and temporal variability

An example time series classification of plastic mulch distribution in Santa Maria, California from June 2023 – January 2025 reveals patterns in land use overturning (Figures 8, 9). Total single-image plastic mulch coverage (white and dark plastics) reached two relative maxima of 1682.56 ha in November 2023 and 1841.21 ha in December 2024, as dark plastic mulch is installed during the winter months (Figure 9; Supplementary Figure S7). As the growing season progresses into the summer months (April - June 2023), crops overgrow underlying dark plastic mulch, decreasing retrieved dark plastic mulch coverage observable in a single image (Figure 9; Supplementary Figure S7). White plastic mulch is applied in the summer months (Holmes, 2024) and white plastics reach relative maxima of 1548.37 ha in July 2023 and 1576.90 ha in July 2024 (Figure 9; Supplementary Table S1). A baseline 630 ha of white plastic is retrieved through the full time series (Figure 9). The white plastic category in this retrieval includes green houses, hoop houses, and semi-permanent plastic structures (see Section 2.4) which are likely to account for this baseline plastic use. Operating under an assumption of single application of dark plastic mulch in the fall and white plastic mulch in the summer months, this classification posits an estimated 2593 ha of agricultural plastic use for the Santa Maria study area from the May 2023 – April 2024 period.

While plastic mulch is an increasingly ubiquitous component of modern agriculture, accurately estimating the contribution of these films to the global plastic waste and pollution burden has proven challenging. Only in 2025 has the EU’s Soil Monitoring Directive incorporated monitoring soil plastic contamination and other novel contaminants into its regulatory framework (Directive (EU), 2025). However, data on agricultural plastic film use remains limited and relies on a mixture of farmer reports and industry data (Goldberger et al., 2019; Wang et al., 2025). There is a paucity of robust methods for assessing the current extent and temporal dynamics of plastic film use, challenging our ability to rapidly detect areas of intensification (number of applications per year), extensification (spread across landscapes), and proximity to sensitive habitats. Because this single-use product markedly contributes to landfill loads and environmental plastic contamination (Tiwari and Sistla, 2024; Wang et al., 2025), the ability to rapidly document agricultural plastic mulch use dynamics over large spatial scales via a remote sensing approach is increasingly recognized as a necessary step to develop best practices for plastic film use in agricultural systems (Veettil et al., 2023). While current regional approaches exist to retrieve plastic distributions with high accuracy, a globally consistent plasticulture algorithm has yet to be developed due to high variability in crop patterns, growing seasons, and plastic film color. To address this challenge, we developed a spectral-based classification of plastic material in multispectral Sentinel 2 imagery using Santa Maria, CA as a test region. Here, we consider the performance, limitations, and transferability of this spectral classification.

Classification performance

Classified plastic use matches trends in agricultural practices

The seasonal dynamics of retrieved plastic mulch coverage align with established regional growing practices along the California central coast. Supplementary Figure S7 depicts strawberry growth under dark green plastic mulch in Monterey County, CA through the November 2024 to September 2025 growing period. In this region, dark plastic mulch is laid in the late fall months (Supplementary Figure S7), providing thermal insulation for seedling crops during winter months. This is consistent with periods of maximum satellite retrieved dark plastic mulch coverage of approximately 1045 and 1117 ha in November 2023 and 2024 over the Santa Maria study area (Supplementary Figure S7). As the growing season progresses, crop canopy obscures underlying dark plastic material and dominates the spectral signal (Figure 9; Supplementary Figure S7). White plastic mulch is commonly applied during summer plantings to reflect sunlight and reduce soil temperatures (Holmes, 2024). This is exhibited in this study’s classification by two relative maxima in white plastic acreage of approximately 1548 ha during July 2023 and 1577 ha in July 2024 (Figure 9). Unlike dark plastic mulch which is only observed in the fall to early spring during initial application, white plastic coverage maintains above a baseline of 630 ha through the entire time series (Figure 9). The “white plastic” class presented here is a mixed-use category which includes green houses, hoop houses, and other semi-permanent white and clear plastic structures which remain in use year-round, likely accounting for this baseline plastic cover. White plastic mulch application and removal may then account for the seasonal summer increases above this baseline. When tuned to the spectral properties of endmember materials, this classification elucidates trends in agricultural plastic application, rather than relying on prior knowledge of land use practices.

Comparisons with conventional machine learning classifiers

Direct comparison with conventional classifiers indicates that the proposed spectral decision tree provides predictive performance comparable to random forest and RBF support vector machine models trained using the same reference data and predictor variables. The primary advantages of the spectral approach are therefore not necessarily higher classification accuracy, but explicit physical interpretation of each decision, computationally simple and rapid inference, and the absence of iterative statistical model fitting once thresholds have been established. In contrast, machine learning classifiers may offer modest performance gains when sufficiently representative labeled data are available. All three approaches will require evaluation and potential recalibration when transferred to regions with different soils, crop architectures, atmospheric conditions, or plastic materials.

Sources of uncertainty

While classified plastic use mirrors farmer growing practices along the California central coast, several uncertainties must be addressed should the model be transferred to other agroecological systems. These include regional differences in three-dimensional structure and endmember compositions, overestimation of plastic use due to subpixel mixing, and flagging of plastic application prior to crop canopy overgrowth. These sources of uncertainty are expanded upon below.

Model transferability

Using Santa Maria as a test case, this study presents a framework for plastic coverage retrievals along California’s central coast, though application to other agricultural systems will likely require threshold reparameterization. Central coast California agriculture is often characterized by highly organized, evenly spaced rows of plastic mulch or bare soil (Holmes, 2024). Agricultural systems with irregular row geometries or crops with differing densities or leaf area indices (LAI) may impact pixel-scale canopy reflectance through increased shade, altered BRDF, and other confounding spectral features (Liu et al., 2024). Regions with more arid climates may also exhibit decreased soil moisture content and, consequently, increased reflectance at SWIR wavelengths (Lobell and Asner, 2002). This will likely alter the thresholds at which a SWIR albedo metric can discriminate dark plastic mulch from soil. Further, LDPE is the most common type of plastic material employed in California agriculture (FAO, 2021). Regions which rely on plastics with different pigmentation and composition than those presented here will more than likely impact decision tree thresholds (i.e., AVW) and require additional spectral data collection and model reparameterization.

Subpixel mixing

Quantifying the degree of subpixel mixing among vegetation, plastic mulch, and soil in imagery remains an important next step toward improved retrievals of agricultural plastic use. This pixel-level classification is currently limited to the resampled 10 m spatial resolution of Sentinel-2 imagery. The entire 100 m 2 pixel is assumed to be the dominant spectral endmember (i.e., vegetation, soil, or plastic). In reality, plastic mulch is juxtaposed by rows of exposed soil and the entire 100 m 2 pixel area is not one homogenous target. Pixels along field perimeters are also mixed with roads and other spectral targets not characterized in this study. Plastic application inside of greenhouses or hoop houses are similarly unresolved with this approach. Future studies can move beyond a discrete, pixel-level classification toward subpixel estimates using a multiple endmember spectral mixture analysis (MESMA, Roberts et al., 1998) or deep learning approach (e.g., Lu et al., 2025). The pixel-level approach presented in this study may be useful to constrain the priors of these more advanced subpixel retrievals.

Crop canopy overgrowth obscures underlying plastic mulch and contributes to temporally variable subpixel mixing throughout the growing season. Following application, crops progressively overgrow the underlying mulch over approximately 1–2 months, increasingly dominating the observed spectral signal and shifting classifications from plastic-covered surfaces toward vegetation classes (Figures 3, 5, 8, 9; Supplementary Figure S7). As canopy density increases, the spectral contribution of the underlying plastic mulch becomes progressively reduced within the 10 m Sentinel-2 pixel. We defer to future studies to determine best practices in flagging plastic mulch application prior to crop overgrowth and retaining this flag through the satellite record for improved pixel and subpixel level plastic use estimates.

Implications for environmental management

While providing a range of agricultural benefits, plasticulture is also an emerging threat to human and environmental health due to a multitude of end-of-life management challenges, including degradation of residual plastic fragments left behind in the field (FAO, 2021). These materials release chemical additives, and large residual plastic fragments in fields disintegrate into finer particles, making these farmlands hotspots of micro (<5 mm) and nano (1–1000 nm) plastics that can feed into surrounding water draining these landscapes or be transported by wind to other systems (Tiwari and Sistla, 2024). Terrestrial plastic pollution represents at least 80% of marine pollution globally (Kumar, 2020) and agricultural runoff is the leading cause of riparian water quality impairment; this risk is amplified by increasingly erratic climate conditions. Shifts in precipitation intensity increase runoff and erosional risk in agricultural systems, and flood events can exponentially increase the transport of plastic debris into riverine systems, with upwards of 100-fold increase in plastic transport observed during flood events (Van Emmerik et al., 2023; Zhao et al., 2024). However, our knowledge of the extent and factors controlling the movement of plastic debris out of agricultural fields remains scarce (Naderi Beni et al., 2023). Most studies addressing the migration of plastic particles from terrestrial to aquatic systems have focused on urban runoff (Cowger et al., 2022; Zhang et al., 2023) and there is virtually no data on the contribution of windblown agricultural plastic fragments to the marine environment.

This knowledge gap partly reflects a lack of a standardized methodology for rapidly and accurately identifying the application, intensification, and extensification of plastic mulch in agricultural landscapes, especially near sensitive waterways and other critical habitats. The spectral diversity of plastic targets has been a historical limitation of remote sensing retrievals (Veettil et al., 2023) beyond white and clear plastic mulch (e.g., Ma et al., 2025) and hoop houses (e.g., Veettil and Xuan, 2022). California agriculture is characterized by spectrally diverse plastic targets, and the approach presented in this study can build toward statewide retrievals of plastic use by incorporating multiple colored mulch targets. Quantifying the spatial and temporal scales of plastic mulch deployment using the methodology presented here can thus provide fundamental data to inform stakeholders engaged in trying to mitigate the plastic debris burden driven by plasticulture and help to identify likely hotspots for targeted agriculturally derived plastic pollution interventions.

Global application of agricultural plastic materials is increasing, yet remote sensing estimations of plastic-mulched farmland are complicated by temporally variable subpixel mixing and diversity in plastic mulch colors and compositions. Informed by in situ reflectance spectroscopy and spectral mixture analysis, this study proposes an interpretable approach for classifying agricultural plastic debris in Sentinel-2 imagery using Santa Maria, CA as a case study. The approach accurately discriminates both bright and dark plastic materials, suitable for the diversity of plastic targets characteristic of modern agriculture in California and elsewhere. Mixture analysis further reveals the anticipated subpixel fractions at which white and dark LDPE plastic mulches are spectrally separable from soil or crops (white >40%, black >50%, dark green >90%, Figures 3, 5), including under simulated variation in atmospheric scattering (Supplementary Figures S2, S3). When applied to the cloud-free Sentinel 2 record from April 2023 to January 2025, the classification retrieved two relative maxima in single-image plastic coverage of 1682.56 ha in 12 November 2023 and 1841.21 ha in 21 November 2024. The classification was parameterized from “pure” unmixed training spectra without incorporating region-specific labeled data. Even so, the retrieval reveals seasonal application of mulch materials, in addition to quantifying aggregate plastic use. Threshold reparameterization may be anticipated when scaling to other agricultural sites given differences in crop canopy architecture, soil aridity, and plastic material colors not characterized in this study, though its spectral basis may enable this reparameterization with minimal training data. The presented approach may be suitable for quantifying agricultural plastic use budgets, identifying the proximity of plastic material to waterways, and targeting hotspots of plastic pollution for agro-environmental management.