On both soils, the spectra of the films are divided into three groups (Fig. 3): black (Black – type 1, Black – type 2, Black – biodegradable, Black&White – small (black side), Black&White – large (black side)), white (Black&White – small (white side), Black&White – large (white side)), and transparent (Transparent – with holes, Transparent – small, Transparent – large). White and black film reflectance is not particularly influenced by the background soil (Fig. 3). In contrast, transparent film reflectance changes according to the soil, with thinner films being closer to the soil spectra than thicker films (Fig. 3). However, the distance between the plastic film and the soil spectra changes across the spectral region. For instance, Transparent – with holes is closer to the spectra of silt loam in the visible, while Transparent – small gets closer in the NIR and SWIR (Fig. 3). On sandy loam, Transparent – with holes and Transparent – small spectra are very close in the visible, while Transparent – large gets closer to the soil spectra in the NIR and SWIR (Fig. 3). These relationships could be due to different additives in the plastic films, targeting different uses and influencing the transmittance of the films. Transmittance measurements could further elucidate how plastic film transmittance changes across the spectral region, and could be a relevant topic for crop production (Jones et al., 2021).

Overall, the reflectance of pristine plastic films showed a high coefficient of variation of 40 % on average, while both soils had a similar coefficient of variation of around 4 %. The reflectance of black films is low across the entire spectrum, and the coefficient of variation is 71 % on average, ranging from 10 % for the Black – type 2 on silt loam to 135 % for the Black&White – small (black side) on silt loam. White films have the highest mean reflectance and the lowest coefficient of variation, ranging from 6 % to 17 %. The two different white films present a similar spectral shape, being extremely reflective in the visible spectra, and decreasingly reflective through the SWIR, while the thicker film had higher reflectance across the whole spectra on both soils (Fig. 3). The transparent films generally follow the spectrum of the soil background, and transparent films on a less reflective soil result in lower reflectance of plastic films, and vice versa. The characteristic soil-shaped spectrum is interrupted by absorption features in the SWIR, which are shown by white films at the same wavelengths (Fig. 3). Specifically, the absorption features were found around the following wavelengths: 1215, 1730, 1765, 2312, and 2352 nm. As a result, HI_1215 and HI_1732 allow distinguishing all non-black films, in both pristine and treated conditions, from soils and mixtures, and from crop residues (Fig. 4a and b). While both indices guarantee a complete separability between plastic and non-plastic, HI_1732 values are one order of magnitude greater than HI_1215. ND_1715 has values similar to HI_1732, but the broader spectral interval results in outliers for soils and mixtures, represented by rare minerals (e.g., Erionite, Xenotime) overlaying with ND_1715 values of plastic (Fig. 4c). For both transparent and white films, the indexes generally have higher values for films with a higher thickness (Fig. 4).

Compared to pristine films, crumpled and dirty films decreased the reflectance variability between replicates of hyperspectral measurements. The average coefficient of variation was reduced to 30 % for crumpled films, 9 % for crumpled and dirty films, and 9 % for dirty films. Overall, crumpling did not influence plastic film reflectance, as the changes in reflectance are within the range of the coefficient of variation for all the films (Figs. 5 and S2–S4). On the contrary, the presence of soil on the film surface influenced plastic film reflectance (Figs. 5 and S2–S4). This is particularly evident in the case of white films, where the reflectance of pristine films is substantially higher compared to dirty films in the visible spectra, while the differences between the spectra are reduced for higher wavelengths (Figs. 5 and S2–S4). Black film reflectance also changes when covered by soil, but the negative values of reflectance difference in Fig. 5 indicate that the presence of soil increased the reflectance of the plastic films. The differences between spectra increase for higher wavelengths, as soil reflectance increases (Fig. 5). Transparent films are the ones affected the least by treatments, with some changes induced by the presence of soil in the case of sandy loam only (Fig. 5). All these differences were consistently observed for the multispectral measurements in proximal sensing, despite generally higher coefficients of variation (e.g., 45 % for pristine films and 15 % for soils, on average).

The spectral separability between films and soil observed in proximal sensing is confirmed by the multispectral acquisitions from the UAV (Fig. 6). White films have the highest spectral separability from non-plastic classes, especially in the bands of the visible spectra, while the reflectance slightly overlaps with other non-plastic elements when approaching the infrared (Fig. 6). Transparent film reflectance largely overlaps with the reflectance of soil and other non-plastic elements, while black films are quite distinct from soils, especially in the infrared bands, where soil reflectance increases (Fig. 6). Overall, all the 0.5 cm datasets had similar performances in detecting the plastic films placed on the field (Fig. 7a), while the multispectral 1.2 cm dataset had higher omission errors. The difference between the 1.2 cm dataset and the 0.5 cm datasets mainly consists in the missed detection of transparent films (Fig. 7a). White films placed on the field were detected with all four datasets in almost every fold, while black films had a generally higher and more fluctuating omission error (Fig. 7a).

Despite the good performances in detecting the plastic films placed on the field, all the datasets overestimated the presence of plastic on the field (Fig. 7b). Overall, plastic overestimation ranges from an average of 21 times for multispectral 1.2 cm dataset to 44 times for RGB 0.5 cm dataset. Black films encountered the highest overestimation compared to the other films, with RGB datasets having a higher overestimation compared to multispectral datasets (Fig. 7b). After black films, transparent films showed the second highest overestimation, finally followed by white films (Fig. 7b). The overestimation was generally related to the confusion of black films with shadows, white films with highly reflective pixels, and transparent films with soil (Fig. 8).

A spectral library of white, black, and transparent LDPE films was built, nearly covering all plastic film colours used in agriculture as crop covers. This work introduces an open-access spectral library for agricultural plastic films and explores non-visible regions of the electromagnetic spectrum for identifying plastic films as residues in agricultural soils. Within white and transparent films, differences related to thickness and transparency of the films mainly affected the mean reflectance, while the resulting spectral shapes were highly comparable (Fig. 3). Transparent film spectra are highly influenced by soil spectra (Fig. 3), and our experimental setup allowed us to acquire the resulting reflectance of soil and plastic films. An experimental setup such as the one used by Jones et al. (2021) would allow a complete characterisation of spectral radiative properties – reflectivity, transmissivity, and absorptivity – and provide insights into spectral changes of transparent films with different backgrounds. Differences between black films can hardly be observed, as spectra are flat and the reflectance is low, producing a low signal-to-noise ratio and high coefficient of variations (Fig. S5).

The mean reflectance of white and transparent films showed much higher variability across replicate measurements compared to soils, resulting in higher coefficients of variation. This is related to both the experimental design and the optical properties of the films. In fact, the measurements were acquired around 4 hours before and 4 h after solar noon, with solar zenith angle ranging from 32 to 59°, and solar azimuth angle ranging from 103 to 262° (SunCalc.Org., 2025). Moreover, the plastic films have a smooth surface, which leads away from the assumption of Lambertian reflector – typically adopted in remote sensing – and causes specular reflection to take over diffuse reflection phenomena (Goddijn-Murphy and Dufaur, 2018; Goddijn-Murphy et al., 2017). This is particularly evident in the case of pristine films, which had the highest coefficient of variation, while the presence of soil on the film surface increases the roughness of the films, favouring diffuse reflection and decreasing the variability of the reflectance between replicates. The implications for plastic residue detection are mostly positive, as particle deposition is expected on land, while specular reflection may be a limiting factor in other applications, such as the detection of clean plastic films with satellite data.

Despite a decrease in the coefficient of variation, we did not find any relevant influence of crumpling on film reflectance, while soil was the main driver of changes in reflectance between pristine and treated films (Fig. 5). Plastic films retained soil particles differently after the dirty treatment, and the comparability between films may be limited in our experimental setup. However, black films appeared to have higher retention of soil particles compared to white films (Fig. 1), while changes in plastic film reflectance were particularly evident for white films in the visible spectra, where the distance between the spectra of plastic films and soil is the highest (Figs. 3 and 5). The presence of soil did not have a particular influence on plastic absorption features, enabling plastic identification through spectral indexes, also on soil-covered plastic films (Fig. 4). However, as the amount of soil on the film increases, a progressive deterioration of plastic spectra is expected, until reaching a soil cover at which plastic films will not be detectable anymore. Identifying this threshold of soil cover will be helpful in delineating the boundaries of remote sensing use for detecting plastic film residues on cropland. Our experimental setup did not allow for controlling the amount of soil placed on the films, but aimed at representing field conditions, where soil particle retention of the film surface varies based on film properties. Moreover, indoor spectroscopy would allow exploiting a higher signal-to-noise ratio and better control over the amount of soil placed on the films compared to outdoor measurements, where wind can remove soil particles from the film surface. We also advise against using a contact probe in indoor spectroscopy, as we experienced plastic film surface melting in contact with the light source.

Film spectra acquired from UAV presented similar spectral shapes to the proximal sensing acquisitions (Figs. 3 and 6). White films had very high reflectance for shorter wavelengths and decreasing reflectance for longer wavelengths, transparent films had reflectance similar to soil, and black films had a very low reflectance. However, it must be considered that the reflectance of white films in the UAV acquisitions was highly influenced by assigning the maximum reflectance of 1 to no data values, corresponding to saturated pixels. This particularly influenced the blue and green bands, where the number of saturated pixels was more than half, while the red and red edge bands had around 20 % and 10 % of saturated pixels, respectively, and the NIR had no saturated pixels.

Good performances were achieved in detecting true positives of plastic films (Fig. 7a). Despite the presence of other plastic residues on the field, plastic films were overestimated compared to the expected number of plastic-covered pixels (Fig. 7b). The results of our study might have been influenced by structural flaws of the survey, such as the absence of ground control points or the low overlap of images during flight, which can fall below the recommended threshold of 70 % when flying at 7 m above ground with the camera used. This could have resulted, for example, in band misalignment and halo artefacts that are visible in the RGB composite (Fig. 2a). However, general considerations can be made on how detection accuracy changes depending on the film colour and on the region of the spectrum used.

Specifically, black films had the highest overestimation, which was due to the presence of shadows on the field (Fig. 8a). In our study area, recent tillage practices induced soil clod formation and emphasised the impact of shadows (Fig. 2). However, surface microtopography is always expected on arable land, and we identify shadows as the main limiting factor in black film detection, similarly to previous works dealing with black plastic detection (Iordache et al., 2022; Shan et al., 2018). White films were slightly overestimated in correspondence with highly reflective soil regions (Fig. 8a). As with shadows, the topography of our study area might have emphasised the issue, creating soil surface angles that result in specular reflection and produce high reflectance values, comparable to those of white films. While techniques like brightness thresholding could improve the results of the classification (Iordache et al., 2022), their implementation and effectiveness may be site-specific, and a few measures could be adopted during data collection to address the challenges related to the illumination geometry. Possible solutions are multi-temporal flights, high-resolution DSM (Digital Surface Model), or flying with a homogenous cloud cover and the highest solar elevation angle possible.

While transparent films had a lower overestimation compared to black films (Fig. 7b), the source of error is related to a strong spectral overlay with soil in the visible and near infrared (Figs. 3 and 6), rather than a structural flaw of the survey. In this case, the only solution to substantially improve plastic detection may be limited to having sensors with bands located in the SWIR. The additional bands on the multispectral camera did not show a substantial improvement in transparent film detection, especially when compared with RGB+indexes 0.5 cm dataset (Fig. 7). For black films, the datasets showed similar producer accuracies, but the multispectral datasets showed a lower overestimation compared to the RGB datasets (Fig. 7). However, the difference in spatial resolution must be considered, as coarser pixel size negatively affects both the areal estimates and the minimum mapping unit. At the same time, for equal pixel size, RGB cameras allow for increasing flight height, consequently reducing flight time and increasing area coverage. Overall, the results of our study suggest that the increased spatial resolution of an RGB camera should be favoured over the higher spectral resolution of multispectral cameras when aiming at the detection of macroplastic residues of around 5 cm.

Regarding the algorithm used and the experimental setup, it needs to be highlighted that using more sophisticated algorithms like deep learning and image segmentation techniques would likely increase detection accuracy. However, our experimental setup might have introduced bias when using textural features (e.g., fixed size and shape of plastic films, placement of plastic films under soil clods) and finding the best image classification algorithm goes beyond the scope of our study. Instead, a random forest allowed us to relate the results to spectral features, and to define which sensors and survey workflow are best suited for plastic films detection on soil. Lastly, a high surface occupation of plastic film residues in agricultural fields is rarely documented, and our experiment was designed to replicate a realistic scenario. Future studies could consider increasing the size of the observations related to plastic films to have more robust results.

We have shown potential workflows for detecting plastic film residues with RGB or multispectral cameras. However, the accuracy of these approaches will be site and film-specific. Machine learning approaches could be limited by the reduced size of training samples. At the same time, the absence of distinct spectral features on current sensor technology hampers the application of traditional methodologies such as maximum likelihood or threshold selection based on spectral indexes. A coarse-textured, clear soil may represent the optimal site for detecting black films, providing a good contrast between the film and the background, and reducing the complexity of the topography by limiting the formation of soil clods. Such soils are also more likely to induce sensor saturation and a lower contrast with clear films, making the detection of white films harder. On the contrary, a fine-textured dark soil may provide a good environment for white film detection, while black film detection would be limited by a decreased contrast and eventually hindered by the presence of soil clods and direct sunlight. Our study investigated a fraction of all possible soil–plastic film combinations, and further research is needed to comprehensively assess the capabilities of UAV-based detection with current technologies. However, plastic films do not share any absorption feature in the visible and near infrared (Fig. 3), and their identification with currently available multispectral broadband sensors is mostly colour-driven and influenced by spectral shape ambiguities (Garaba, 2025).

While the use of plastic absorption bands in the aquatic environment may be limited by water absorption features (Knaeps et al., 2021; Moshtaghi et al., 2021; Garaba and Dierssen, 2018), their use is extremely promising for soils. Plastic indexes enable unambiguous identification of plastic film residues on soil (Fig. 4), allowing their detection even when spectral features overlap with other regions of the spectrum, such as the visible range. We showcased the use of plastic indexes available in the literature, but we provided open access to the spectral library used in this study to promote the development of new spectral indexes and detection algorithms (Garaba, 2025). Traditional detection algorithms, such as threshold selection on spectral indexes, potentially enable overcoming the limitations of machine-learning-based methodologies related to the reduced size of training datasets. Within currently available indexes, even the broadband index ND_1715 resulted very effective in identifying plastic films from other elements that could possibly be found on cropland – except for some rare minerals that will not be found on arable land (Fig. 4c) – demonstrating the effectiveness of broad bands located at plastic absorption features and encouraging the development of multispectral cameras tailored to plastic detection.

Plastic residues rarely exceed 100 cm² on agricultural land (Piehl et al., 2018; Stefano and Pleissner, 2022; Wang et al., 2022), placing high requirements on spatial resolution, which is traded off with spectral resolution and areal coverage in sensor technology (Shaw and Burke, 2003). This strongly limits the use of technologies with large areal coverage, such as satellite data. In addition, the use of technologies with high spectral resolution on airborne platforms, such as hyperspectral sensors, would require low flight altitudes, resulting in increased flight time and associated costs, as well as the collection, storage, and processing of large volumes of data. The availability of broad bands related to SWIR absorption features on miniaturised sensors, together with high-resolution visible bands, would reduce costs, data volume, and complexity of current technologies, representing a milestone for monitoring any non-black plastic residue on land.

Additionally, according to Castagna et al. (2023), remote sensing offers the potential to distinguish at least three plastic types, exploiting the different locations of absorption features in the SWIR. PE, PP, and PVC – the most produced plastic types in Europe and the most used polymers in agriculture (Plastics Europe AISBL, 2024) – have similar absorption features, and their distinction might be challenging (Castagna et al., 2023). PET instead – the fourth most produced plastic type in Europe (Plastics Europe AISBL, 2024) – has different absorption features, and it can be distinguished from PE, PP, and PVC by using narrow-band indexes (Castagna et al., 2023). On agricultural land, this could mean distinguishing plastic residues generated by agricultural management from plastic residues generated by littering (e.g., PET bottles). Building spectral libraries and characterising the absorption features of the most common plastic residues is key to direct future efforts in developing new sensors, which we identify as a necessary technological development to monitor soil plastic contamination. The introduction of legal limits on plastic concentration in soil may not be far off, as initial regulatory examples begin to emerge (Meixner et al., 2020). This will likely drive demand for faster and standardised monitoring and increase the appeal of a sensor for on-soil plastic detection. Moreover, as macroplastic residues can fragment into microplastic (Yang et al., 2022; Song et al., 2017; Julienne et al., 2019), reliable macroplastic monitoring will also support better microplastic assessments.