1 Introduction

The coffee plant (Coffea arabica L.) is of great economic, social and cultural relevance in Brazil and around the world. Effective management of coffee crops can ensure good productivity, generating greater economic return, and reducing detrimental impacts to the environment.

The disease known colloquially as coffee leaf rust poses a significant threat to Brazilian coffee producers, due to favourable environmental conditions in most coffee regions of the country, and where there is still widespread use of susceptible varieties [39]. The pathogenic fungus can cause losses of up to 50% if no control measures are employed [34, 49].

Control measures for coffee rust are administered across entire plantations, carried out with the use of protective and systemic fungicides that are applied to the plant’s leaves, or by drenching the soil, respectively. Agrochemical sprayings are calendar-based starting in November/December and continuing through April [19]. The spatial and temporal distributions of coffee rust incidence are not homogeneous across a crop, thus it can be assumed that any total area leaf rust control strategy could be replaced by focused management in specific areas of disease incidence [2].

Precision agriculture techniques can result in potential benefits in profitability, yield, sustainability, crop quality and traceability, environmental protection, food security and rural economic development [35]. The efficacy of disease management strategies in the field can be accessed using geographic information system (GIS), global navigation satellite system (GNSS), geocomputation, remote sensing, and variable rate application technology integrated with advanced information processing in georeferenced maps during crop growth, development and production [3, 35].

A commonly-used remote sensing method for vegetation studies is the use of vegetation indices. Vegetation indices are mathematical operations involving two or more spectral bands designed to enhance vegetation properties, allowing reliable spatial and temporal inter-comparisons of photosynthetic activity and canopy structure [28]. The method is based on the fact that energy reflected in red and near-infrared (NIR) regions is directly related to photosynthetic activity of the vegetation canopy, as well as the assumption that using two or more spectral bands can substantially minimize the main sources of noise that affect detected variation in vegetation responses [21].

Changes observed in spectral properties within the NIR region may provide information on senescence and vegetation stress issues [31]. The Multi-Spectral Instrument (MSI) sensor embedded in the Sentinel-2 satellite obtains terrestrial surface spectral reflectance information by recording it in 13 different bands of the electromagnetic spectrum. Five of these spectral bands are positioned in distinct bands within the NIR region, and three of these five are located in the rapidly growing reflectance zone known as Red-Edge [20]. Based on reflectance measurements applied to simulate the Sentinel-2 satellite sensor bands, Chemura et al. [13] examined three levels of coffee rust severity in coffee leaves under greenhouse conditions. According to the authors, bands located at the Red-Edge spectral position can be useful for disease detection and coffee crop status evaluation.

In coffee plantations in Brazil, Landsat 7 and 8 images associated with in situ coffee rust data have been used to analyse spectral radiometry behaviour of plants under different irrigation systems [40]. The authors observed that in areas with higher coffee rust incidence under a central pivot and rainfed system, the average reflectance of NIR and green spectral regions was reduced, while reflectance increased in short wave infrared (SWIR) and red wavelengths, when compared to periods with lower coffee rust incidence. Moreover, NDVI generated by Landsat 7 ETM + sensor images were used to evaluate coffee leaf fall caused by coffee rust in Guatemalan coffee plantations [32]. It was observed that the NDVI standard deviation value in damaged farms increased as the average NDVI value decreased. By analysing annual NDVI changes, it was also reported that disease progression is not homogeneous in crops [32].

Based on the hypothesis that it is possible to detect changes in the spectral behaviour of plants due to the incidence of leaf rust through orbital remote sensing, this study aimed to evaluate the effectiveness of different vegetation indices for remote detection of coffee rust incidence, and to characterize the spectral signature of crops with different levels of disease incidence.

2 Materials and methods

2.1 Area description

The study area is located in Minas Gerais Agricultural Research Company (EPAMIG) experimental field in Três Pontas city, Minas Gerais state, Brazil. The cultivar present in the area is Catuaí Amarelo IAC-17, a variety susceptible to coffee rust. Plant spacing is 3.5 m × 2.0 m with 2 plants per pit, and a crop age of 42 years. The approximate central geographical point of the study area has a latitude of 21° 20′ 38.36″ S and a longitude of 45° 28′ 48.51″ W, and the average altitude is around 935 m. The region is classified as Cwa, with dry winters and rainy summers [1].

2.2 Georeferencing of sample points

Images obtained by adding the Sentinel-2 MSI sensor to the Spectra Precision SP60 L1/L2 RTK GNSS receiver software provided latitude and longitude data, corresponding to the center of fifty pixels in different spectral bands (Table 1). Subsequently, the GNSS RTK receiver base was placed in the field and the fifty sample points were manually located by walking with the mobile RTK receiver in the sampling area. In each of the fifty identified subareas (measuring 10 m × 10 m), a central coffee plant was georeferenced, which was then used to select a further two plants within a radius of approximately 2 m within the planting line. In total one hundred and fifty coffee plants were evaluated, three plants per sampling point.

Table 1 Spatial and spectral resolution of Sentinel-2 satellite MSI sensor bands [20] used to calculate vegetation indices
Full size table

2.3 Crop treatments

The study area was split into two sections (Fig. 1). The northern section measured approximately 0.57 ha and contained 32 sampling points, while the southern section measured around 0.43 ha and contained 18 sampling points. No rust control was performed in the northern section, however the southern section received a systemic fungicide containing triazole and strobilurin on 11 December 2018 and 20 February 2019, as well as a leaf application of copper on 8 April 2019. The application of fertiliser was carried out following Guimarães et al. [26], based on soil fertility management and leaf nutrition in coffee culture recommended for Minas Gerais state, Brazil.

Fig. 1
figure 1

Study area indicating the location of Minas Gerais state within Brazil, and the south/southwest region of the federative unit (left), the area of Três Pontas city (bottom right), and the plantation area in the EPAMIG experimental fieldsite, showing the sampling pixels (subareas) (top right)

Full size image

2.4 Field evaluation: coffee rust, defoliation and yield

Coffee rust was sampled monthly between December 2018 and April 2019 (totalling five separate samples), using a non-destructive method that takes into account plant exposure to solar radiation. For each coffee plant, eight leaves were evaluated; four leaves from different randomly selected plagiotropic branches on the north-facing side of the plant (i.e. higher insolation levels), and four leaves from different plagiotropic branches on the south-facing side of the plant (lower insolation levels) [15]. Therefore, a total of 24 leaves were evaluated per subarea. Leaves of the 3rd and 4th pair of plagiotropic branches in the middle third of the plants were assessed according to the methodology proposed by Lima [36]. The percentage of coffee rust incidence for each subarea was obtained by dividing the total number of leaves with recorded coffee rust signals by the total number of leaves evaluated in the subarea (24 leaves), and multiplied by one hundred.

Defoliation was determined using strands positioned at two points on one plagiotropic branch of each of the one hundred and fifty coffee plants. Leaf fall was recorded in each monthly assessment following the initial number of leaves counted in December 2018, totalling five samples. These leaf drop data were then transformed into a percentage of defoliation.

Fruit harvest was carried out on 2 July 2019 with a portable Nakashi DL 262-MT2L® gasoline machine. A cloth was placed on the floor around the plants to facilitate the collection of the fallen fruits. The volume of fruits from the three plants in each subarea was measured using a graduated bucket, meaning that the volume assigned to each of the fifty subareas corresponded to liters per three plants. Fruits were collected from all one hundred and fifty plants that were previously evaluated for coffee rust incidence levels.

2.5 Multispectral data

The orbital data captured by the Sentinel-2 satellite-embedded MSI sensor was obtained by a free download available from the United States Geological Survey [44] collection. Sentinel-2 scenes were selected using the criteria of least atmospheric interference and closest proximity to date of crop rust sampling (Table 2).

Table 2 Dates that in situ data were obtained, respective orbital data used for correlation analysis and cloud coverage of Sentinel-2 images [44]
Full size table

2.6 Sentinel-2 digital image processing

Digital images were processed using the Dark Object Subtraction (DOS) method [11, 12]. Correction by the DOS method is carried out by estimating the atmospheric interference in each spectral band, performing calculations to transform the digital number into radiance values, and then to surface reflectance values. Minimum and maximum radiance, top-level irradiance, and sensor gain were necessary variables for calculations, and varied for different bands. Additionally, other factors that depend on the image date were included, such as Earth-Sun distance and solar elevation angle [27].

2.7 Vegetation indices calculations and Pearson’s correlations

Vegetation indices (Table 3) using different relationships between Sentinel-2 spectral bands were calculated. The vegetation indices selected provide data on plant canopy such as leaf area, plant leaf moisture and chlorophyll content.

Table 3 Vegetation indices calculated to correlate with leaf rust incidence, defoliation and yield data obtained from field samples, spectral bands necessary for index calculations and authors who developed/adapted them
Full size table

Vegetation index calculations (Table 3) were performed for 7 selected Sentinel-2 images for correlation with coffee rust incidence, defoliation and yield data. Another Sentinel-2 image (November 2018) was selected for index calculations for use in comparison with monthly rainfall intensity data. To perform the correlations, values were extracted from each index corresponding to the 50 georeferenced subareas in the coffee crop area selected for rust sampling (i.e. the 50 pixels of highest spatial resolution bands of Sentinel-2 MSI images). Pearson’s correlations (Fig. 2) were performed between in situ data and vegetation indices values extracted from the 18 sampling points in the southern section (with coffee rust chemical control). The same procedure was performed for the 32 points in the northern section (without coffee rust chemical control).

Fig. 2
figure 2

Correlations between field data sampled in EPAMIG experimental fieldsite in Três Pontas city and vegetation indices calculated with Sentinel-2 images

Full size image

In order to evaluate spectral behaviour in the south and north sections, reflectance values of each 50 sample points were extracted from all 13 spectral bands in the Sentinel-2 images. For each of the 13 bands, monthly mean reflectance values of the 32 northern sample points were calculated. The same procedure was performed for the 18 southern sample points. Student’s t-tests were then performed to compare monthly averages of areas with and without leaf rust chemical control.

2.8 Meteorological in situ data

To better understand the coffee plant’s spectral response to coffee rust, monthly rainfall data were used to compare with the monthly vegetation indices data. Rainfall intensity (mm/month) was measured with a rain gauge from three conventional weather stations located in the cities of Machado, Lavras and São Lourenço, that surround the EPAMIG experimental fieldsite in Minas Gerais state, Brazil [30]. With these data, rainfall intensity interpolation was performed by using an inverse distance weighting method [45].

3 Results

3.1 Spatial–temporal coffee rust distribution

An increase in coffee rust incidence was recorded over the months of assessment, most notably in the area where the disease was not controlled (Fig. 3). Coffee rust incidence classes were defined in order to facilitate the visualization of rust spatiotemporal progress (Fig. 4).

Fig. 3
figure 3

Coffee rust incidence percentage between December 2018 and April 2019 in areas with and without chemical control in the EPAMIG experimental fieldsite, Três Pontas city, Minas Gerais state, Brazil

Full size image
Fig. 4
figure 4

Coffee rust incidence percentage in crop subareas over time in control and uncontrolled sections (continuous and dashed lines, respectively) in the EPAMIG experimental fieldsite, Três Pontas city, Minas Gerais state, Brazil

Full size image

3.2 Correlation of data derived from Sentinel 2 with field data

Pearson’s correlation indicated an interaction between levels of coffee rust incidence and the vegetation indices calculated for each respective month of evaluation. It was noted that in February 2019 negative correlations were reported between various vegetation indices and coffee rust incidence in the northern section of the study site (Table 4).

Table 4 Pearson’s correlations (r) between different vegetation indices, yield and defoliation coffee rust incidence between December 2018 and April 2019, in sampling areas with chemical control measures (c), and without chemical control measures (nc)
Full size table

The vegetation indices calculated with the February 2019 images using spectral bands in NIR (B08) and Red-Edge 1 (B05) presented the highest correlation values with leaf rust data sampled in the respective month [NDVI(RE1) r = − 0.562; CI(RE1) r = − 0.573; MSR(RE1) r = − 0.569]. For NDMI indices calculated using NIR band (B08) and one of the bands in SWIR position (B11 or B12), we obtained, respectively, moderate and weak correlations for February 2019 (r = − 0.518 and r = − 0.481), and moderate for April 2019 (r = − 0.502 and r = − 0.519).

The yield harvested in July 2019 had a moderate positive correlation (r = 0.555) with coffee rust incidence reported in April 2019 in the sampling section without any chemical control. Furthermore, the IRECI vegetation index generated from images recorded in August and September 2018 also correlated positively with leaf rust assessment data obtained in February and March 2019 (Tables 5 and 6).

Table 5 Pearson’s correlations (r) between vegetation indices calculated with August images with coffee rust incidence data and yield data from sample points in the area without chemical control
Full size table
Table 6 Pearson’s correlations (r) between vegetation indices calculated with September images with coffee rust incidence data and yield data from sample points in the area without chemical control
Full size table

3.3 Reflectance and vegetation index averages in areas with and without coffee rust control

By analysing average monthly reflectances for the different bands of the electromagnetic spectrum in both the northern and southern sections, it was observed that there were higher average reflectances in the Red-Edge/NIR region (Fig. 5). It was also noted that for the spectral region in Red-Edge/NIR, there were reflectance peaks for the month of February 2019 in both sections. However, average reflectances at RE2 (740 nm), RE3 (783 nm), NIR-8 (842 nm) and NIR-8a (865 nm) were higher in the northern area with no rust control. By analysing isolated bands, two SWIR bands were found to have the highest reflectance in December 2018, indicating a lower moisture content in the vegetation canopy.

Fig. 5
figure 5

Average monthly reflectance of Sentinel-2 spectral bands in the two areas (with and without coffee rust control measures) in the EPAMIG experimental fieldsite at Três Pontas city, Minas Gerais state, Brazil

Full size image

Student’s t test showed that the average reflectance of RE2 (B06), RE3 (B07), NIR-8 and NIR-8a bands in the northern and southern sampling sections did not differ from each other in December 2018. However, reflectance averages became statistically different when comparing the north and south sample areas due to increased occurrence of coffee rust in January and February 2019, corresponding to higher reflectance in the northern section where the incidence of leaf rust was higher (Table 7).

Table 7 Spectral bands reflectance averages of areas with (c) and without (nc) coffee rust control and Student’s t-test results (‘ns’ equal to not significant) for independent samples (comparison of means)
Full size table

The tested vegetation indices responded differently to rainfall intensity (Fig. 6). The curve of indices NDVI, NDMI(B11), NDMI(B12), RNDVI, NDVI(RE1), CI(RE1), MCARI/OSAVI and MSR(RE1) are similar and appeared to correspond with the rainfall intensity curve with a delay of around 3 months, however, with different intensities. For indices calculated with SWIR bands, the pattern was observed more sharply. The vegetation index TCARI/OSAVI presented this same pattern in inverted form. For the IRECI case, the response to rainfall seemed to be faster, following the rainfall intensity curve with an interval of less than 1 month.

Fig. 6
figure 6

Rainfall intensity and vegetation index curves of monthly mean values in areas with and without chemical control of coffee leaf rust at Três Pontas city, Minas Gerais state, Brazil

Full size image

4 Discussion

This study aimed to evaluate the effectiveness of different vegetation indices calculated from Sentinel-2 MSI images for remote detection of coffee rust incidence symptoms, and to characterize the spectral signature of crops with different coffee rust incidence levels. Field data were obtained for comparative analysis with orbital data on several dates. Rainfall intensity data were estimated for the crop area to help understand the spectral behaviour of coffee plants with different disease intensity.

4.1 Spatial–temporal leaf rust distribution

Outbreaks of coffee rust were detected at the start of field evaluations, and from these affected spots the area that had not received chemical control was further infected by the disease. Hemileia vastatrix uredospores are dispersed inside the canopy by raindrops, through the spread of water flow in the infected leaves during rainfall [34]. In short, the sudden increase in the magnitude of disease incidence is the result of its effective dispersal through water [6, 38]. However, wind gusts also promote uredospore dispersal across crops in full sun. Such dispersion is favoured during dry days with low relative humidity [7]. As a result, through the kriging map analysis it was observed that distribution and incidence intensity of coffee leaf rust were not homogeneous throughout the coffee field [2].

4.2 Correlation of data derived from Sentinel 2 with field data

Vegetation indices for remote disease monitoring may aid crop management. Through correlations between vegetation indices and field data for the same month, it was found that February 2019 is the best month to extract information on conditions of H. vastatrix infestation in coffee crops. In addition, it seemed appropriate to use vegetation indices that make use of bands in the NIR/Red-Edge region. Spectral bands positioned in the Red-Edge region appear to have greater sensitivity for identifying changes in the plant canopy when dealing with leaf diseases [9, 13, 18, 50].

The IRECI vegetation index calculated using MSI Sentinel-2 images from August and September 2018 was positively correlated with the February 2019 rust incidence data (IRECIaug r = 0.566 and IRECIsep r = 0.493). This observation may help in the localized control of coffee leaf rust, since in the subareas with higher IRECI values, higher levels of leaf rust incidence were observed. Furthermore, coffee leaf rust incidence data from April 2019 correlated positively with yield data collected in July 2019. The positive relationship between plant yield and coffee rust incidence has been widely discussed in the specialized literature [4, 5, 8, 10, 16, 37, 43, 48].

4.3 Reflectance and vegetation index averages in areas with and without leaf rust control

This study reported higher average reflectances in the area with chemical control measures in February 2019 for bands RE2, RE3 and the two NIR plateau bands, relative to the area without chemical control. According to Knipling [33], reflectance in the NIR region is a result of the scattering of electromagnetic radiation that occurs due to existing interface areas in the transition between cell walls. Furthermore, the mesophyll deterioration leads to the reduction of existing air cavities, with a consequent reduction of reflectance in the NIR region. However, the author also notes that at the beginning of the leaf senescence process, there is an increase in these interface areas due to the separation of adjacent cells, as well as the distancing of living cell content from cell walls. Therefore, February 2019 may have marked the beginning of the leaf senescence process, as evidenced by increased reflectance in the Red-Edge/NIR region, most notably in the northern fieldsite (without chemical control) which had higher coffee leaf rust incidence.

High reflectance values were observed for bands in the SWIR region for both areas in December 2018. The wavelengths in the SWIR region are absorbed by water, meaning that when vegetation suffers from a water deficit, there is increased reflectance in the spectral bands positioned in this region of the electromagnetic spectrum [29, 31, 46]. Therefore, despite the rainfall regime beginning in October 2018, the leaf structure had not yet shown to have water at its maximum potential.

Vegetation indices apparently responded to rainfall after approximately three months, with the exception of IRECI, which appears to have responded in less than a month. IRECI provides information on the chlorophyll content of leaves at a ratio close to 1:1 g/m2 and still works well as a leaf area index [22]. Decreasing leaf water potential leads to decreased chlorophyll content, as this pigment is sensitive to increased environmental stress, especially salinity and drought [25]. Therefore, there may be a relationship with the occurrence of this observed behaviour and more detailed studies are needed to obtain better understanding.

5 Conclusions

This study demonstrated the varying effectiveness of vegetation indices for remote detection of coffee rust signs and symptoms through correlation analyses of different indices. Vegetation indices obtained through relationships between bands in the Red-Edge and Near Infra-Red regions were more sensitive to changes caused by coffee leaf rust. Sentinel-2 orbital data transformed by calculating the IRECI vegetation index for August and September may be useful for indicating areas for coffee rust field sampling, and could be used to determine appropriate disease control according to variation in incidence rates. Future studies should examine the applicability of this vegetation index as a useful tool to aid in coffee plantation management.

Through this work it was also possible to characterize areas with different levels of disease incidence. The relationships between the monthly average spectral signature and levels of leaf rust incidence vary depending on the different regions of the electromagnetic spectrum. This may be due to the fact that the onset of the leaf senescence process reflectance peaks in most bands in the Red-Edge/NIR region were higher in areas with higher incidence of coffee leaf rust.