2.1 Methodology

As mentioned above, this study begins with the identification of the spectral reflectance signature of brown rust (P. melanocephala) and orange rust (P. kuehnii), which contaminate sugarcane crops (Saccharum spp.). Our methodology focuses on the identifying of spectral bands capable of differentiating between healthy and contaminated leaves. These bands were finally used to obtain multispectral images, which allows the calculation of a vegetation index capable of distinguishing the difference between the presence and absence of rust. To do this, we follow the steps below:

The steps of this methodology are summarized in Figure 1 and described in the following subsections.

Figure 1 Experimental methodology 

2.2 Preparation of culture samples

The samples of brown rust were collected in an experiment considered as proof of resistance to this disease of cultivars under study. This experiment was planted in September 2016 in the experimental block of the Estación Territorial de Investigaciones de la Caña de Azúcar Centro Villa Clara (ETICA Centro Villa Clara), province of Villa Clara, Cuba. This station belongs to the Instituto de Investigaciones de la Caña de Azúcar (INICA). The collection of the leaves was done close to 11:00 am in the month of March 2017, having an age of six months. The samples of orange rust were collected in an experiment for the evaluation of resistance of commercial cultivars to orange rust and monitoring of the evolution of this disease. Samples of this experiment were planted and harvested equally in ETICA Centro Villa Clara, in the same dates of the brown rust. For the study, samples of leaves infected with the brown rust of cultivar B4362 (susceptible to P. melanocephala) and of leaves infected with the orange rust of cultivar C01-227 (susceptible to P. kuehnii) were collected.

For the selection of samples of infected leaves, in each case a leaf diagnosis was first adopted, leaves +3 (third leaf with visible dewlap) were collected, using the scale of five degrees of attack severity (estimating the percentage of area foliar affected by pustules), according to [³⁴], where: grade 1 is applied when the affected leaf area (ALA) reaches up to 5% (ALA ≤ 5%), grade 2 (5% < ALA ≤ 15%), grade 3 ( 15% < ALA ≤ 25%), grade 4 (25% < ALA ≤ 50%) and grade 5 (50% < ALA). Subsequently, the presence of the causative agent was confirmed through an optical microscope, taking into account the characteristics and morphological differences of the uredospores of both rusts described by [³⁵]. Cultivars with grade 1 are considered resistant to brown rust and orange rust, and grade 5 are considered highly susceptible [³⁶]. Five leaves of cultivar B4362 were collected, one healthy (not infected by P. melanocephala) and four infected by brown rust with ALA grade between 1 and 4. In the same way, five leaves of the cultivar C01-227 (susceptible to P. kuehnii) were collected: one healthy (not infected by P. kuehnii) and four infected by orange rust with ALA grade between 1 and 4.

2.3 Use of the spectrometer to obtain spectral signatures

To obtain the spectral reflectance signature, the Corona Plus Remote spectrometer [Figure 2], marketed by the German firm Carl Zeiss, was used. This is a laboratory equipment that has a measuring head that is connected to the central unit using fiber optics and this unit in turn connects to a PC for the analysis of the data and visualization of the results via the Ethernet interface. Its reflectance measurement range is from 398 nm to 1,702 nm in 422 bands, with a spectral resolution lower than 10 nm, sensing a circular area of sample with approximately 2.5 mm radius, equivalent to 20 mm². The curves were obtained reading in points along each leave with separations between 20 and 30 mm approximately, without taking into account that the reading is made on a contaminated area or not.

The data obtained by the spectrometer were processed by the Aspect Plus software, version 1.76 (C) (https://www.zeiss.com/spectroscopy/products/software/aspect-plus.html). This is a modular and flexible spectral analysis software package that works under MS Windows. Through Aspect Plus, spectral measurements and processing can be performed for the Zeiss MCS 5xx and Corona spectrometers.

Figure 2 Spectrometer Corona Plus Remote (https://www.zeiss.com/spectroscopy/products/spectrometer-systems/corona-plus-remote.html) 

2.4 Multispectral images production

To validate the practical application of the analysis of the obtained spectral reflectance signatures, images obtained with a multispectral camera were used. The multispectral images were obtained using the Parrot Sequoia camera (https://www.parrot.com/business-solutions-us/parrot-professional/parrot-sequoia). This camera contains five independent sensors that operate simultaneously to obtain their respective images in the different spectral bands, a sunshine intensity sensor to adjust the cameras to the existing lighting conditions, a GPS module, an inertial measurement unit that registers the orientation of the cameras, as well as a magnetometer that acts as a compass and indicates the direction in the flight. Designed for agriculture with excellent precision, it can be adapted to all types of drones. The data provided by all these components are integrated as metadata in captured images.

The five cameras obtain respectively a panchromatic image (RGB) of 16 Mpx (4,608x3,456), 24 bits per pixel (8 bpp for each channel of color), with a pixel of 1.34 (m, focal length of 4.88 mm, and are saved in JPG format, as well as four monochromatic images in: 1) range of green (G) at 550 nm with 40 nm bandwidth (BW), 2) range of red (R) at 660 nm with 40 nm BW, 3) range of red edge (RE) at 735 nm with 10 nm BW, and 4) range of near infrared (NIR) at 790 nm with 40 nm BW. The monochromatic images are 1.2 Mpx (1,288x960), 16 bits per pixel (65,536 gray levels), with 3.75 (m pixel, 3.98 mm focal length, and are saved in TIF format.

Using the Parrot Sequoia multispectral camera with sun lighting in the morning of a clear day, images of the sugarcane leaf samples were taken. The distance between the camera and the leaves was approximately 1.5 m, which gives a resolution at ground level, or ground sample distance (GSD), of 1.4 mm per pixel. The GSD determines the spatial resolution of the images. This is defined as those fine details that can be captured by the images or what portion of the ground can cover a pixel of the image. The smaller the GSD in a digital image, the smaller the details that can be captured and the greater the spatial resolution. According to [³⁷], this distance can be calculated using Equation (1),

where H is the height at which the image was taken, P _s is the distance between pixels in the camera sensor (pitch), and f the focal length of the camera lens. From the GSD used, it is achieved that the spots of contamination with rust in the leaves can be covered by several pixels in the image, making clearly visible the pixels of the image that correspond to the affected regions when calculating the indexes. This analysis must be taken into account for images taken at distances of several tens of meters, in which the distance covered by a pixel of the image is greater than the size of the spot on the contaminated region of the leaf. Figure 3 shows an example of some of the leaves used contaminated with orange rust and brown rust.

Figure 3 Examples of leaves used contaminated with brown rust and orange rust 

A second validation of the analysis of the obtained spectral reflectance signatures was performed at field level. For this purpose, multispectral images were taken with the Parrot Sequoia camera, inserted in an unmanned aerial vehicle that flew at a height of 80 m over the parcels from which the sugar cane leaves were extracted with each type of rust.

2.5 Calculating the vegetation index

The normalized difference vegetation index (NDVI) is used to extract vegetation abundance from remote sensing data by isolating the dramatic increase in reflectance over the visible wavelengths from red ρ(R) to the reflectance from near infrared ρ(NIR), and normalizes it by dividing by the total reflectance of each pixel at those wavelengths using Equation (2).

In this experiment, the NDVI is obtained by taking multispectral images in the red (R) and near infrared (NIR) bands. Thus, each pixel of the image calculated using Equation (2) indicates the value of the NDVI at its location on the leaf as shown in sections 3.2 and 3.3.

In theory, the value of this index ranges between -1 and 1. However, in practice the measurements generally range between -0.1 and 0.8 [³⁸, ³⁹]. Clouds, water, snow and ice give negative values of NDVI. Naked soils and other background materials produce NDVI values between -0.1 and 0.1. The highest NDVI values occur when the amount of green vegetation in the observed area increases, ranging from 0.2 to 0.8. Healthy plants have a high NDVI value due to their high infrared reflectance and relatively low red light reflectance. Phenology and vigor are the main factors that affect the NDVI, so it is possible to estimate the state of the health of the plant in general, the photosynthetic activity, and the possible deficit of nutrients, as well as to study the spatial and temporal tendencies in the dynamics of vegetation, productivity and distribution. In [⁴⁰] the NDVI is used to identify areas of sugarcane and evaluate their health condition from images of the IRS LISS-II satellite, which makes it possible to calculate the area of the land surface covered by the crop and which portion is diseased. The authors of [⁴¹] used the NDVI, among other indices, to demonstrate the influence of water deficit in the low fruit production in apple trees. The work in [⁴²] measures sugarcane yield potential across a large number of genotypes using canopy reflectance measurements, such as NDVI. Recently, Johansen et al. [⁴³] investigated the use of high-spatial resolution satellite image data and geographic object-based image analysis, based on NDVI, to map putative sugarcane grub damage. In [⁴⁴] also focused on the task of classifying sugarcane, taking as input NDVI time series extracted from remotely sensed images. In addition, the authors of [⁴⁵] evaluated the potential of the NDVI to monitoring changes in sugarcane yield imposed by different straw removal rates. Finally, [⁴⁶] explored the potential of the Sentinel-2 satellite to discriminate between the severities of yellow rust infection in winter wheat using some vegetation indices, such as NDVI. However, research relating NDVI to brown rust and orange rust in sugarcane has been poorly evaluated in the literature.

3.1 Analysis of spectral signatures

Using the Corona Plus Remote spectrometer and the samples of sugarcane leaves contaminated with brown rust and orange rust, we obtained the spectral reflectance curves that are shown in Figure 4. For healthy leaves, the low spectral reflectance in the portion of the blue (450 nm) and red (670 nm) bands is observed in Figure 4, due to the strong absorption by the leaf pigments (mainly chlorophyll). On the contrary, a peak of reflectance in the green portion (550 nm) is manifested. In the NIR region (700-1,300 nm), the high reflectance of the leaves can be seen, a characteristic in which cellular structures influence. This step type pattern between the visible regions of the electromagnetic spectrum and NIR is characteristic of green vegetation. Beyond the NIR, the short wave infrared region (1300 nm and above), an absorption peak at 1,450 nm can be perceived as a result of the presence of water. On the contrary, in the contaminated leaves there is a noticeable dispersion in the spectral curves, which is due to the fact that the contamination covers at most 50% of the leaf, and the readings of each curve are made randomly along the leaf between contaminated areas and not contaminated. The main difference between healthy and contaminated leaves is seen, for orange rust, in a decrease of ((() for the region between 750 and 900 nm (RE~NIR), and for brown rust in an increase of ((() in the region between 600 and 700 nm (R), a decrease of ((() in the region between 750 and 900 nm (RE~NIR), and an increase of ((() in the region between 1,400 and 1,600 nm. Furthermore, the difference between leaves contaminated with orange and brown rust is mainly seen in the R, RE~NIR and between 1,400 and 1,600 nm regions.

Quantifying the differences in the spectral bands of the Parrot Sequoia multispectral camera, it can be observed that in the leaves contaminated with orange rust (top row of Figure 4], the average reflectance decreases in the G band from 22.6%, in the healthy leaf, to 20.01% in the contaminated leaf. In addition, the R band remains almost unchanged, and both the RE and NIR bands decrease from 57.67% and 65.85%, to 55.35% and 63.42% respectively. On the other hand, the difference in signatures of spectral reflectance between healthy leaf and contaminated with brown rust is much greater. In the G band it decreases from 20.56% reflectance in the healthy leaf to 16.3% in the contaminated leaf, while in the R band an increase is produced from 10.73% up to 16.78%. However, the greatest differences are seen in the RE and NIR bands, where the reflectance values decrease from 56.21% and 68.2% in the healthy leaf to 42.63% and 48.85% in the contaminated leaf. Additionally, the spectral signatures of the leaves contaminated with both types of rust also exhibit differences. It is important to emphasize that those leaves that are infected with orange rust will present a higher reflectance than the leaves infected with brown rust in the G, RE and NIR bands.

Figure 4 Spectral reflectance signatures of healthy leaves (0% of ALA), and contaminated leaves (50% of ALA) with orange rust (upper row) and brown rust (lower row) 

These differences can be better appreciated in Figure 5 for the case of brown rust. In this figure, the mean values and standard deviations of the reflectance values for healthy leaf (0% ALA) are plotted, as well as contaminated in grade 1, 2, 3 and 4 (up to 50% of ALA) with this type of rust. In addition, the mean values of the spectral signatures for the five grade are included in the lower right corner, indicating the wavelength values corresponding to the G, R, RE and NIR sensors of the Parrot Sequoia multi-spectral camera. These mean values and standard deviations were obtained by taking between 17 and 24 readings on each leave. As can be seen, not only is there evidence of the notable difference that spectral signatures have between healthy and contaminated leaves, mainly from 15% of contamination (grade 2 of ALA), but also that in the four monochromatic bands that can sense the Parrot Sequoia multispectral camera most of these differences occur. This same analysis carried out for the leaves contaminated with orange rust shows less significant differences between the different grade of contamination and between these and the healthy leaves.

These results were obtained with higher spectral resolution than that previously published by [³³], which used the Hyperion EO-1 satellite which has only 220 spectral bands in the range of 400 nm to 2,500 nm and a spatial resolution (GSD) of 30 m.

Figure 5 Mean values and standard deviation of spectral signatures of healthy leaf and contaminated with brown rust in ALA grades of 1 to 4. In the lower right part of figure, the mean values are compared, and the G, R, RE and NIR bands of the Parrot Sequoia camera are indicated 

3.2 Multispectral images analysis

The images obtained did not undergo any adjustment process. As an example, the images in the four multispectral bands obtained from the leaves contaminated with brown rust can be seen in Figure 6 without radiometric correction, as well as the reflectance levels that emit a healthy region of interest and another contaminated region. The radiometric correction can equalize the sunlight used to obtain these multispectral images with the light of the halogen lamp used to obtain the spectral reflectance curves.

In spite of the numerical differences due to the uneven illumination, and the non-radiometric correction, the proportionality that exists between them and the reflectance characteristic of the spectral reflectance signatures can be verified. For healthy regions: low reflectance in the R band, with high reflectance in the RE and NIR bands. For the regions contaminated with both types of rust: decrease of the spectral reflectance in the G, RE and NIR bands, with an increase of reflectance in the R band. This increase in the R band is more appreciable for the leaves contaminated with brown rust. Figure 7 shows these values comparatively in the four bands of the Parrot Sequoia camera for the cases of healthy and contaminated leaves in grade 4 of ALA with orange and brown rusts. This figure confirms the result obtained previously by spectrometry, in which differences in reflectance between healthy and contaminated leaves are significant for brown rust, but not for orange rust.

Figure 6 Images obtained by the four bands of the multispectral camera in leaves contaminated with brown rust. Reflectance values showing a point of a healthy region and a contaminated region 

Figure 7 Comparative values of reflectance between healthy and contaminated leaves at degree 4 of ALA for the two types of rust in the Parrot Sequoia camera bands 

3.3 Vegetation index

The NDVI calculated from the reflectance values measured by spectrometry shows clear differentiation between healthy and contaminated zones, mainly for brown rust, been greater the index value in healthy areas than in contaminated areas. From Figure 7, the minimum difference between NDVI values in healthy and contaminated areas for the leaves with orange rust is 0.1327, while for the leaves with brown rust is 0.2398.

Using the reflectance values obtained by the multispectral images [Figure 6], the NDVI was calculated to obtain the images shown in Figure 8 for both types of rust. In this figure, the numerical values of the NDVI are shown in a pixel of a healthy area and in another of a contaminated area, the numerical difference between these zones being corroborated for both types of rust. In this case, an average difference between points of the uncontaminated leaf and contaminated points of 0.2081 for orange rust and 0.2379 for brown rust is obtained. Taking into account that the NDVI varies between -1 and 1, this represents a ratio of relative decrease of the index between contaminated areas with respect to non-contaminated areas in approximately the 11.9% for brown rust and 9.9% for orange rust.

Unlike what was published in [³³], where the R and NIR bands are combined with the short wave infrared (SWIR) at 1,600 nm from the Hyperion EO-1 satellite to achieve a high differentiation between sugar cane plants contaminated with orange rust, this work demonstrates the possibility of not using the SWIR band. This makes it possible to use small multispectral cameras, which have only the spectral bands R and NIR, placed on board vehicles flying at low altitude, to monitor with high spatial resolution parcels susceptible to brown rust and orange rust.

Figure 8 NDVI calculated from the multispectral images for each type of rust indicating the numerical value of the index in one pixel of a healthy region and in another of a contaminated region for each type of rust 

The second validation, carried out at field level, flying over sugar cane parcels susceptible to each type of rust, also confirms the validity of the NDVI obtained by multispectral imaging to differentiate areas contaminated with rust from healthy areas. Figure 9 shows an example of an image with the NDVI obtained. The validation of the aerial images was carried out by an expert from the ETICA Centro Villa Clara, where the experiments were carried out.

Figure 9 Example of NDVI calculated on parcels of sugarcane susceptible to rust