Hyperspectral Images of Vine Leaves Treated with Antifungal Products

Abstract

Hyperspectral imagery provides detailed insights for vineyard vegetation assessment, enabling improved pesticide management within precision agriculture. For this reason, the dataset presented here includes hyperspectral images acquired from grapevine leaves treated with two copper-based formulations: ZZ Cuprocol (containing 70% w/v copper oxychloride) and Cuprantol Duo (composed of 14% w/w copper oxychloride and 14% w/w copper hydroxide). In addition, a commonly used contact pesticide in both intensive and traditional viticulture, Folpet—free of copper but containing sulfur and chlorine—was also evaluated in its commercial formulation Vitipec Azul (Cimoxanil 6% w/w, Folpet 37.5% w/w, Ascenza, Portugal). For each product, six different dilution levels were prepared along with a distilled water control. Leaf samples were collected and analyzed during the 2023 growing season from three shoot locations (basal, middle, and apical) and from both orientations of the vine canopy: east and west. Following pesticide treatment, leaf hyperspectral images were captured using a 300-band Pika L camera (Resonon, Bozeman, MT, USA), mounted on a mechanical scanning platform synchronized with the imaging system.

1. Summary

Fungal diseases, and in particular downy mildew, are one of the biggest concerns of winegrowers throughout the vineyards worldwide [1]. Copper-based antimicrobial compounds have been widely used to control phytopathogens for nearly fourteen decades [2], and the use of copper is still the most effective way to control downy mildews [3].

Improper fungicide use, whether due to under-application or over-application, as well as spray drift that prevents the product from reaching the vine canopy and instead deposits it onto the soil, leads not only to financial losses but also to the progressive contamination of agricultural soils [4,5].

The most extended method to determine this drift is based on the use of Water Sensitive Papers (WSP). It has shown as inoperative to representatively characterize the pesticide deposition along the whole vineyard. There is a present demand for improving spray-based vine treatments in terms of quantifying product depositions: boosting efficiency of deposition, reducing drift and increasing sprayer output. Hyperspectral images have shown to be an effective and non-destructive tool in viticulture. It has been applied to a wide range of applications, such as classifying grapevine varieties [6,7], grape ripeness estimation [8,9], quality assessment of wine grapes [10,11], and monitoring in real time the vineyard canopy [4]. Although hyperspectral imagery has already been used to detect pesticides in grapes [12,13], measuring their deposition on leaves is still an open challenge that requires further research. To bridge this gap, the present work proposes a dataset containing hyperspectral images of vine leaves that could be used to increase the efficiency of treatments in vineyards. Although many successful projects in agroforestry and related fields have used low-cost passive imaging sensors like RGB and NIR, many applications need higher spectral accuracy that only multispectral and hyperspectral sensors can offer [14,15,16].

Both hyperspectral and multispectral methods involve capturing images in which, for each spatial element within the image, a spectrum of the energy reaching the sensor is measured. The key difference between them lies in the number of bands (or channels) and their widths [17]. While multispectral imagery includes between 5 and 12 bands, with each band captured using a remote sensing radiometer and represented in pixels, hyperspectral imagery features a much larger number of bands—hundreds or even thousands—organized within narrower bandwidths of about 5 to 20 nm each. As a result, such imagery is widely applied to precision agriculture [18,19].

Several commercial solutions are currently available to estimate the application efficiency of copper-based fungicides; however, these systems generally rely on indirect measurements and lack the capability for real-time monitoring of product deposition, similar to the limitations observed in Water Sensitive Papers (WSP). Conversely, spectral analysis techniques have proven to be effective tools for the non-invasive assessment of substances accumulated in grapevine foliage, particularly in the evaluation of nutrient content [20].

As the images released in this dataset are on standard and open formats, they could be used by other researchers in order to perform further analysis and for the training of different Machine Learning models for different purposes, including pesticide applications, among others. In particular, these data offer significant potential for advancing the detection and characterization of fungicide compounds using hyperspectral imaging techniques. Possible applications include pesticide deposition detection, quantitative assessment of fungicide coverage on leaf surfaces, and the identification of relevant spectral features or wavelength selection strategies. Additionally, the dataset may support comparative studies, methodological benchmarking, and the development of new algorithms for precision agriculture and plant protection monitoring.

Despite being conducted under controlled laboratory conditions to ensure high internal consistency, this study has certain limitations. In particular, measurements were restricted to a single phenological stage, and therefore, the reported spectral patterns may not capture variability across the full vegetative cycle. Future research, including multiple developmental stages, would strengthen the broader applicability of these spectra.

2. Data Description

Leaves corresponding to the images in this database were treated with: (i) five microliters of the 18 preparations and 2 blanks of distilled water, making a total of 20 randomized drops per batch and 6 repetitions of batches per leaf; (ii) with a pressurized mist nozzle spraying both sides of the leaves, and (iii) with a conventional spray, trying to imitate real conditions of pesticide application. Images were taken with the wet treatment and after 25 min of unforced natural drying. The taken images are organized in 4 folders: ‘TEMPLATE’, ‘NOZZLEUP SIDE’, ‘NOZZLE LOWER SIDE’ and ‘BIG DROPS’. All in all, the dataset is composed of 372 hyperspectral images.

The dataset comprises hyperspectral imagery of grapevine leaves (Vitis vinifera L., cv. Tempranillo) collected from a vineyard situated in the D.O. Cigales region (Valladolid, Spain), where the plants were treated with varying doses of copper- and sulfur-based pesticides.

The folder named “Hyperspectral images of vine leaves treated with pesticides” contains a total of 372 hyperspectral images corresponding to hypercubes in .bil (Band Interleaved by Line) image format. Also, a .bil.hdr text file corresponds to each image, including the whole metadata, and a .tiff archive shows a preview image of each leaf.

The metadata included in the bil.hdr file provides valuable information for understanding the hyperspectral image format and calibration specifications, such as the number of scanned lines and the number of bands (300 spectral bands), resulting from a spectral binning of 2, with their exact units in nm. The reflectance factor scale is presented with values from 0 to 10,000. The data is stored with data type 12, corresponding to 16-bit unsigned integer values, and with byte order 0 (little-endian), without a header offset. During acquisition, a shutter time of 15.314 ms and a gain of 12.0 were used, allowing an acquisition rate of approximately 64.96 frames per second. The file includes image rotation information for correct spatial orientation, as well as the pixel size in m. Finally, the system is identified by the serial number of the sensor, 100121-178.

The dataset is organized according to the structure shown in

Table 1. Dataset structure and sample images.

FOLDER: TEMPLATE	TOTAL LEAFS: 24		TOTAL IMAGES:48			TOTAL FILES: 142
Wet treatment (EMw100723)		Dry treatment (EMd100723)
Deposition Template
FOLDER: NOZZLE UP SIDE	TOTAL LEAFS: 48		TOTAL IMAGES:144		TOTAL FILES: 570
No treatment (EM7210723)	Wet treatment (EM7w210723)		Dry treatment (EM7d210723)
FOLDER: NOZZLE LOWER SIDE	TOTAL LEAFS: 48		TOTAL IMAGES:144		TOTAL FILES: 320
SUBFOLDER: NOZZLE LOWER SIDE
No treatment (EG5220823)	Wet treatment (EG5folpetw220823)		Dry treatment (EG5folpetd220823)
FOLDER: BIG DROPS	TOTAL LEAFS: 12		TOTAL IMAGES: 36	TOTAL FILES: 72
No treatment (EM1300823)	Wet treatment (EM1cuprantolw300823)		Dry treatment (EM1cuprantold300823)
Sample images channels set: 639.1 (Red), 801.4 (Green), 459.2 (Blue)

.

2.1. Structure of the Dataset

Leaves corresponding to the images in this database were treated with: (i) five microliters of the 18 preparations and 2 blanks of distilled water, making a total of 20 randomized drops per batch and 6 repetitions of batches per leaf. Images were taken with the wet treatment and after 25 min of unforced natural drying, (ii) with a pressurized mist nozzle spraying both sides of the leaves, and (iii) with a conventional spray, trying to imitate real conditions of pesticide application. The taken images are organized in 4 folders: ‘TEMPLATE’ (containing 48 images corresponding to the drop treatment done at 3 positions in the shoots, canopy orientation, replications, dry and wet of 24 leaves), ‘NOZZLEUP SIDE’ (containing 144 images corresponding to 48 leaves without treatment, with nozzle application of the products, and when this application is dried), ‘NOZZLE LOWER SIDE’ (containing 144 images corresponding to 48 leaves without treatment, with nozzle application of the products, and when this application is dried), and ‘BIG DROPS’ (containing 36 images of 12 leaves without treatment, with nozzle application of the products, and when this application is dried). All in all, the dataset is composed of 372 hyperspectral images.

As the data are released on standard and open formats, they could be used by other researchers in order to perform further analysis and for the training of different Machine Learning models to select spectral wavelengths related to pesticide applications.

Hyperspectral technology, using information provided by 300 spectral channels, allows for the study of the surface of the Tempranillo vine leaves that appear in the dataset in three different formats: untreated, with wet treatment, and with dry treatment. The systemic Cu-containing antifungal products used for the study were ZZ Cuprocol (70% w/v Copper oxychloride) and Cuprantol Duo (Copper oxychloride 14% w/w, copper hydroxide 14% w/w), both from Syngenta (Syngenta, Switzerland). A contact antifungal compound, such as Folpet (2-[(Tricloromethyl)sulfanyl]-1H-isoindole-1,3(2H)-dione), widely used in intensive and traditional viticulture and containing sulfur and chloride as heteroatoms, has also been tested in the commercial form of Vitipec Azul (Cimoxanil 6% w/w, Folpet 37.5% w/w, Ascenza, Portugal). The concentration of the different products was varied from 0.20, 0.33, 0.40, 0.66, 0.80 and 1.33 g L⁻¹ prepared by diluting the product in distilled water. These concentrations correspond to applications of 100, 200 and 400 g ha⁻¹ of active product at 2 volumes of application: 300 L and 500 L.

Figure 1 shows how the spectra of the same whole leaf when the product is not applied and when it is in wet status are very similar in appearance, because the untreated part of the leaf has a determining incidence in proportion to the area covered by the product droplets. However, there are slight differences that will become more evident in the individual treatment of the individual spectra resultant of the product deposition points shown in the technical validation section. In general, the main values are lower in the wet status treatment graphic the variation is bigger.

When we look at the dry product graphic, we can see that the deviation variation is smaller and there is a clear difference in the values up to the 850 nm channel.

Low values between 400 nm and 500 nm are mainly related to carotenoid and chlorophyll (a + b) contents, and the characteristic large peak around 550 nm is attributed to the anthocyanin content [8]. The spectral region between 620 nm and 680 nm is related to the chlorophyll content of the leaves.

The red edge between 680 and 750 nm is also typical for vegetation, separating the visible spectral region related to pigments and the plateau between 750 and 1000 nm related to the leaf structure. The region 700–740 nm is of special interest, as it is found to be the most sensitive to plant stress [8,9].

2.2. Average Spectra per Product and Concentration

An average spectrum per product and concentration has been calculated for all of the deposition points (Figure 2 and Figure 3). All of the spectra corresponding to each product and concentration plot have the same shape but differ in the intensity of the reflectance. Higher concentrations have higher reflectance in the spectral region between 400 and 700 nm and lower reflectance in the next range from 700 to 900 nm. In this case, the channels are cropped from 450 nm to 900 nm as we observed spectral noise appearing on the limit channels according to similar investigations [10].

Figure 2 shows the spectra for the dry treatments for each product and concentration. Product 1, Cuprantol duo (Syngenta), shows spectral variations in the 550 nm region and in the infrared zone. In the first region, the spectra corresponding to the highest concentrations show higher readings. However, in the infrared region, the spectral readings for the highest concentrations are lower.

For the spectra of the dry treatment of the second product, Cuprocol (Syngenta), we find a similar behaviour, but less pronounced, in both regions. There is therefore a distinctive spectral difference for both products and also a difference between concentrations.

The spectra belonging to product 3 have significantly lower values at the beginning of the graph, and the trend of values approaching the IR is clearly different from that of copper-based products.

Figure 3 shows the spectra of the wet treatments. They behave in a similar way to the dry treatments, with higher readings in the 550 nm zone and lower readings in the infrared region as the concentration of the product droplet increases. Copper and non-copper-based products differ in the behaviour of the graph, as the sulphur-based product presents lower readings, in line with the wet status data.

If we compare the spectra of the copper-based products with the spectra of the sulphur-based product, although the shape of the spectral graph is similar, it presents clear differences in the 450–550 nm region, in the region around the peak of 550 nm and, especially, in the infrared region. In both product spectra, the highest concentrations show lower spectral values in the infrared zone and higher values up to 550 nm.

We have differentiated spectra for the treatments, depending on the product, the concentration and its state of application, wet or dry. These results have great potential for researchers seeking to better understand the application of fungicide products. This can help determine accurately and in real time the percentage of coverage of the products applied to the vine leaf, helping to improve the effectiveness of applications.

3. Methods

3.1. Study Area

Leaf material was collected from a vineyard located within the Cigales Denomination of Origin in north-central Spain (41°49′17″ N, 4°35′49″ W), at an elevation of 770 m above sea level. The area presents a Mediterranean climate characterized by warm summers (Csb) according to the Köppen classification.

The vineyard consists of Vitis vinifera L. cv. Tempranillo vines grafted onto 110-Richter rootstock, planted at a spacing of 3.0 × 1.5 m (equivalent to 2222 vines per hectare). The training system employed is a vertically positioned double Royat cordon. The vines have an average age of 25 years and are established on calcareous, sandy-textured soils classified as Calcaric Cambisol (CMca) following FAO guidelines. The plot is oriented northeast–southwest, with southern exposure and an average slope of 7%. Leaves of three size categories (small, medium, and large) were sampled from both the eastern and western sides of the canopy during the 2023 growing season, specifically in July and August.

The samples were transported to the laboratory under refrigerated conditions at approximately 10 °C. Immediately after collection, the leaves were treated with the three pesticides to prevent dehydration.

3.2. Experimental Design

The systemic copper-based fungicides evaluated in this study were ZZ Cuprocol, containing 70% w/v copper oxychloride, and Cuprantol Duo, formulated with 14% w/w copper oxychloride and 14% w/w copper hydroxide; both products are manufactured by Syngenta (Basel, Switzerland). In addition, the contact fungicide Folpet (2-[(trichloromethyl)sulfanyl]-1H-isoindole-1,3(2H)-dione), which is extensively applied in both conventional and traditional viticulture and incorporates sulfur and chlorine as heteroatoms, was also included in the experiments. Folpet was assessed through its commercial formulation Vitipec Azul, composed of Cimoxanil (6% w/w) and Folpet (37.5% w/w), produced by Ascenza (Portugal).

The application of the products has been carried out on the beam and the underside of sampled leaves, with: (i) a misting nozzle, (ii) with localized depositions of 5 µL droplets using a precision micropipette with the help of a template for the location of droplets, and (iii) using a manual sprayer with bigger drops, with a similar effect to real field conditions.

The concentration of the different products was varied from 0.20, 0.33, 0.40, 0.66, 0.80 and 1.33 g L⁻¹ prepared by diluting the product in distilled water. These concentrations correspond to applications of 100, 200 and 400 g ha⁻¹ of active product at 2 volumes of application: 300 L and 500 L.

3.2.1. Misting Nozzle

To carry out this experiment, a misting nozzle was attached to a compressor machine. The vine leaves were displayed in an extended manner and labelled as shown in Figure 4. The application of products was carried out at a uniform pressure of 4 Bars.

3.2.2. Deposition Template

As shown in Figure 5, a circular PVC template was designed and manufactured for this assay, for the deposition of 5 µL drops of product. The template consists of 120 deposition points separated by 1 cm. The template has deposition sites for 6 repetitions of the 18 treatments (3 products and 6 concentrations, as explained in the previous section), and 2 more spaces for applying the distilled water blanks.

3.3. Data Acquisition

3.3.1. Experimental Setting

As illustrated in Figure 6, the configuration used to acquire the hyperspectral images included in the dataset is based on a motorized platform synchronized with the optics of the hyperspectral camera during image capture. The camera, together with the halogen spotlights used to illuminate the samples, is securely mounted on a fixed support structure. Adjacent to the acquisition frame, a computer system is installed, equipped with dedicated control software for both the camera and the moving platform, as well as an external hard drive for data storage.

3.3.2. Sample Movement Platform Hardware Structure and Control Device

The system incorporates a motor-driven platform that translates at low speed along two parallel rails, with its motion precisely synchronized to the hyperspectral camera through the use of a stepper motor. Communication and control are handled by Arduino-based software connected to the platform via a USB cable. This interface enables automated coordination of the acquisition process, allowing the user to set the translation velocity and to define the initial and final positions for image capture.

3.3.3. Illumination Setting

The illumination setup is composed of four halogen light sources designed to enhance performance in the visible spectral region. These lamps are installed on two lighting rails fixed to the camera support frame and are oriented with a 25° beam angle, as shown in Figure 6.

3.4. Hyperspectral Image Acquisition

Hyperspectral images were acquired using a Resonon Pika L hyperspectral camera (Bozeman, MT, USA) (CMOS push broom sensor; spectral range 388–1024 nm; 300 bands; spectral binning = 2) under controlled laboratory conditions with constant artificial illumination, fixed sensor-to-sample distance, and stable geometry throughout the 2023 campaign (July–August). The camera was thermally stabilized prior to acquisition.

All images were captured using fixed acquisition parameters (shutter time: 15.314 ms; gain: 12.0; framerate: 64.96 fps), which remained unchanged across sampling dates.

Reflectance conversion was performed using the manufacturer’s proprietary software. A calibrated white reference panel provided by the manufacturer was acquired once per day under the same acquisition settings as the samples and measured twice to ensure stability. Dark signal was recorded at the beginning of the experiment by covering the camera lens. Daily recalibration with the white reference ensured inter-date consistency.

No spectral smoothing was applied, and noisy spectral extremes were removed during data processing, while original hyperspectral images were preserved. Reflectance values (scale factor = 10,000) correspond to the average of replicate measurements. Complete acquisition metadata is provided in the ENVI headers to ensure reproducibility.

The system features a spectral resolution of 2.7 nm and a spatial resolution of 900 pixels. During acquisition, the distance between the camera lens and the sample tray was set at 62 cm.

4. User Notes

The hyperspectral dataset is publicly available for reuse and supports research on fungicide detection and characterization using hyperspectral imaging. It enables machine and deep learning model development, quantitative assessment of leaf coverage, spectral feature selection, and methodological benchmarking for precision agriculture and plant protection applications.