Remote Sensing to Detect Nests of the Leaf-Cutting Ant Atta sexdens (Hymenoptera: Formicidae) in Teak Plantations

Leaf-cutting ants of the genus Atta are an important insect pest in forest plantations in many countries of South America. The objective of this work was to evaluate the potential for using Landsat-8 images, with medium spatial resolution and distributed free of charge, to detect leaf-cutting ant nests in Tectona grandis plantations in Brazil, using partial least squares discriminant analysis (PLS-DA). The regression model adjusted by PLS-DA selected three principal components with a cross-validation error of 0.275 to map and predict the presence of leaf-cutting ant nests in these plantations. The most important bands and vegetation indices were selected using the main variables in the projection (VIP) and predicted pixels with the presence or absence of leaf-cutting ant nests with an accuracy of 72.3% on an independent validation data set. The study indicates that Landsat-8 OLI images have the potential to detect and map leaf-cutting ant nests in T. grandis plantations.


Introduction
Tectona grandis L.f. (Lamiaceae), known as teak, is a deciduous tropical tree species, native to southern and southwestern Asia [1]. The wood of this plant has high commercial value for sawmill purposes [2,3]. The area planted with T. grandis in Mato Grosso state, Brazil, is over 74 thousand hectares [4]. Insect pests such as leaf-cutting ants limit the productivity of T. grandis plantations in monocultures [5,6]. Ants of the genus Atta are generalist herbivores with widespread distribution in the Americas, from the southern United States to northern Argentina [7]. These ants are considered pests throughout the Neotropical region because they cut and transport plant fragments to cultivate their symbiotic fungus Leucoagaricus gongylophorus, from which their individuals feed [8,9]. A single nest of Atta laevigata (Smith, 1858) (Hymenoptera: Formicidae) can cut about 5 kg of plant material per day [10], and 89 nests of the genus Atta can collectively forage 825 kg of plant biomass per hectare [11].

Study Area
The study areas were located in the municipalities of Porto Esperidião (15 • 51 10" S, 58 • 27 37" W and 160 m mean altitude, approximately) (area 1) and Cáceres (16 • 04 14" S, 57 • 40 44" W and 118 m mean altitude, approximately) (area 2), with four and three T. grandis stands of seminal origin, respectively, both 4.5 years old, planted in 4 × 4 m spacing, totaling 104.73 and 152.05 hectares, respectively, in the western region of Mato Grosso state, Brazil ( Figure 1). These areas were part of the Amazonia and Pantanal biomes transition, with a tropical climate with a dry winter, type Aw according to the Köppen classification [29]. The mean annual temperature and precipitation of the region are 24.9-25.3 • C and 1360-1565 mm, respectively [29].
All leaf-cutting ant nests in the study area were located (census), photographed, and measured with a tape measure (at the largest width and length of the area containing loose soil) and the center of each one georeferenced with GPS navigation using the Garmim Etrex 20 (accuracy of 3.6 meters). All leaf-cutting ant nests in the study area were of the genus Atta and there was no other defoliator species present which could mask the results.
Ten individuals of the soldier caste were collected per leaf-cutting ant nest and placed in 70% alcohol containers, photographed, identified at the species level [30], and kept in the reference collection of the Laboratório de Fitossanidade (FitLab) of the Instituto Federal de Educação, Ciência e Tecnologia de Mato Grosso (IFMT), Campus Cáceres, Mato Grosso state, Brazil.
A 90 meter buffer was generated from the geographic coordinates around each leaf-cutting ant nest in T. grandis stands using the QGIS software [31]. This distance guaranteed the generation of points or pixels without interference from other leaf-cutting ant nests mapped in the T. grandis stands [32]. Random points, in locations not covered by the buffers generated, were allocated and considered to be sites without the presence of leaf-cutting ant nests in order to allow balanced statistical analyses of the data. Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 12 All leaf-cutting ant nests in the study area were located (census), photographed, and measured with a tape measure (at the largest width and length of the area containing loose soil) and the center of each one georeferenced with GPS navigation using the Garmim Etrex 20 (accuracy of 3.6 meters). All leaf-cutting ant nests in the study area were of the genus Atta and there was no other defoliator species present which could mask the results.
Ten individuals of the soldier caste were collected per leaf-cutting ant nest and placed in 70% alcohol containers, photographed, identified at the species level [30], and kept in the reference collection of the Laboratório de Fitossanidade (FitLab) of the Instituto Federal de Educação, Ciência e Tecnologia de Mato Grosso (IFMT), Campus Cáceres, Mato Grosso state, Brazil.
A 90 meter buffer was generated from the geographic coordinates around each leaf-cutting ant nest in T. grandis stands using the QGIS software [31]. This distance guaranteed the generation of points or pixels without interference from other leaf-cutting ant nests mapped in the T. grandis stands [32]. Random points, in locations not covered by the buffers generated, were allocated and considered to be sites without the presence of leaf-cutting ant nests in order to allow balanced statistical analyses of the data.

Acquisition and Pre-processing of Satellite Images
The multispectral images of the Landsat-8 Operational Land Imager (OLI) satellite (30 meter resolution), one week after or before the leaf-cutting ants nests census, free from clouds and shadows, were obtained free of charge from the United States Geological Survey (USGS) ( Table 1) [33].

Acquisition and Pre-processing of Satellite Images
The multispectral images of the Landsat-8 Operational Land Imager (OLI) satellite (30 meter resolution), one week after or before the leaf-cutting ants nests census, free from clouds and shadows, were obtained free of charge from the United States Geological Survey (USGS) ( Table 1) [33]. The digital numbers of the images were initially converted to radiance in order to radiometrically correct the images. The radiance was converted to reflectance with the following equation: where λ is the solar radiance at the top of the atmosphere; M p the multiplicative scale factor of reflectance for the band; Q cal the digital pixel value; A p the reflectance additive scaling factor for the band; and θ the solar elevation angle [34]. The radiometric correction of the topographic effects induced in the Landsat-8 OLI images was performed using the Cosine correction method, using a 30 m digital elevation model of the terrain acquired from the National Institute for Space Research (INPE) of the Brazilian geomorphometric database project, Topodata. The radiometric correction of the induced topographic effects was calculated using the following equation: where H is the horizontal surface reflectance; T the sloping surface reflectance; θz the solar zenith angle; and IL the sun angle on a sloped surface [34]. Eight vegetation indexes (VI) were calculated from the reflectance data of the Landsat-8 OLI satellite bands: atmospherically resistant vegetation index (ARVI), land surface water index (LSWI), modified normalized difference water index (MNDWI), normalized burn ratio (NBR), normalized difference vegetation index (NDVI), photochemical reflectance index (PRI), red green ratio index (RGI), and simple ratio (SR) ( Table 2). These indexes were used to assess the phytosanitary status of the forest areas [35,36]. Table 2. Vegetation indexes (VI) calculated from the reflectance data of the Landsat-8 Operational Land Imager (OLI) satellite bands for the detection and monitoring of leaf-cutting ant nests in Tectona grandis plantations.

Vegetation Indexes (VI) Reflectance
Red green ratio index (RGI) Red/Green Simple ratio (SR) Near infrared/Red The forest stands in geometric location format were clipped at the end of the preprocessing step, resulting in reflectance images of the teak tree canopy target per spectral band. Corrections and processing were performed with the R program [37] and raster [38], rgdal [39], splancs [40], landsat [41], landsat8 [42], and shapefile [43] packages.

Spectral Characterization of Leaf-cutting Ant Nests
The canopy reflectance values of areas with or without damage by leaf-cutting ant nests, for each of the six spectral bands, and the values of the different vegetation indexes were subjected to the permutational multivariate analysis of variance (PERMANOVA) [44] with Mahalanobis distance [45] to reduce the bias of the analysis. The p-value was calculated by: where pseudoF is obtained from the original data and Fp is a permutation pseudoF [44], with a significance level of p < 0.05 and 9999 replicates. This allowed the evaluation, with Landsat-8 OLI satellite bands, of the differences between the pixels with a presence or absence of leaf-cutting ant nests. The analyses were performed with the R program [37] and vegan [46] package.

Prediction and Mapping of the Leaf-cutting Ant Nests
The presence of leaf-cutting ant nests was predicted by partial least squares discriminant analysis (PLS-DA), used to predict categorical data [47]. The six multispectral bands (2, 3, 4, 5, 6, and 7) and the eight vegetation indexes (ARVI, LSWI, MNDWI, NBR, NDVI, PRI, RGI, and SR) were introduced in a PLS-DA regression algorithm to predict the presence or absence of leaf-cutting ants nest in the study areas. This algorithm reduced the collinearity effect between the input variables and maximized the correlation between the predictors of the independent variables and the dependent one [48]. Partial least squares regression was performed by the following equations: where X is the matrix of the predictors; Y is the matrix response; T = X-scores; U = Y-scores; P = X-loadings; Q = Y-loadings; E = X-residues; and F = Y-residues [49]. The number of principal components was defined according to the classification error rate calculated by the leave-one-out cross-validation approach for an accurate estimation of the PLS-DA regression model [50]. The predictive variables, chosen and included in the PLS-DA regression model, were evaluated by calculating the important one in the projection (VIP) [50], which determined the contribution of each one to the data set and identified the most important independent variables to predict the presence of leaf-cutting ant nests. The adjustment and calibration of the PLS-DA regression were conducted with a subset of 75% of the database and the prediction with the remaining 25% of the unused pixels in the last step and used as cross-validation data, and the procedure was replicated 999 times.
The leaf-cutting ant nests were mapped by adjusting the PLS-DA regression model to the 25% subset of the dataset in the seven T. grandis stands, for pixel-to-pixel prediction. The general accuracy (%) of the maps was obtained by the simple proportion of the number of correctly classified pixels divided by the total ones of the stands. The analyses were performed with the R program [37] and mixOmics [51] and raster [38] packages.

Spectral Caracterization of the Leaf-cutting Ant Nests
A total of 299 nests of the leaf-cutting ant Atta sexdens (Linnaeus, 1758) (Hymenoptera: Formicidae) was found in the seven T. grandis planted stands and, therefore, 299 geographic coordinates without them were generated, allowing a balanced analysis of the data. The average area of loose soil from A. sexdens nests was 79.8 ± 4.73 m 2 . The permutational multivariate variance analysis (PERMANOVA) of the six Landsat-8 OLI sensor spectral bands and the eight vegetation indexes used in this study differed by the Mahalanobis distance for pixels with presence or absence of leaf-cutting ant nests (pseudo-F = 16.701, p = 0.001, DF = 1).

Predicting and Mapping Leaf-cutting Ant Nests
The PLS-DA model allowed for the separation of the spectra according to the presence or absence   The highest VIP values were 1.3088, 1.2961, 1.2363, and 1.2239 for bands 6 and 2, with the vegetation indexes MNDWI and PRI, respectively (Figure 2b).
The overall accuracy of the cross-validation model for predicting the presence of leaf-cutting ant nests in T. grandis stands was 72.3% with three principal components, using the validation data set (Figure 3).
The estimated average cost of conventional monitoring of leaf-cutting ants (plots or transects) is US$1.23/ha, while that of estimated remote monitoring was US$0.08/ha-about 15 times lower. The overall accuracy of the cross-validation model for predicting the presence of leaf-cutting ant nests in T. grandis stands was 72.3% with three principal components, using the validation data set (Figure 3).
The estimated average cost of conventional monitoring of leaf-cutting ants (plots or transects) is US$1.23/ha, while that of estimated remote monitoring was US$0.08/ha-about 15 times lower.

Discussion
The differences, identified by PERMANOVA, between the bands and vegetation indexes of the Landsat-8 OLI satellite for areas with or without the presence of the nests of the leaf-cutting A. sexdens indicate that there are specific spectral signatures to explain the phenomena studied. Similar results were obtained for the monitoring and detection of areas attacked by T. peregrinus in eucalyptus plantations in Brazil with the use of Landsat-8 OLI satellite bands [19], and in eucalyptus, pinus, and acacia plantations in South Africa using bands and vegetation indexes of the WorldView-2 satellite [35].
The pixel spectral characterization, using bands and vegetation indexes of the Landsat-8 OLI satellite and PLS-DA model to detect the presence or absence of leaf-cutting ant nests in T. grandis plantations, was efficient for initially evaluating their existence in the area. Similar results were obtained with the PLS-DA model for discriminating samples of Swietenia macrophylla King., Carapa guianensis Aubl., Cedrela odorata L., and Micropholis melinoniana Pierre, with predictive errors of 0.14, 0.09, 0.12, and 0.6, respectively [52], and to detect areas attacked by T. peregrinus, using the spectral bands and vegetation indexes of WorldView-2 satellite, with RMSE of 4.27% [35]. The choice of the best PLS-DA model with three components and a classification error of 0.275 showed that the increase in the number of components in the model does not increase precision [53]. Similar results were obtained in the mapping of eucalyptus plantations damaged by the bronze bug, T. peregrinus, with the best PLS-DA model having three components and a prediction error of 0.248 [19]. The error variation of the PLS-DA regression model for predicting the presence of Uromycladium acacia (Cooke) P. Syd. & Syd. (rust) in Acacia mearnsii trees was lower, with six components, with a prediction error of 4.09% using Landsat-8 OLI bands and vegetation indexes [54]. This reinforces the robustness of the PLS-DA regression model in discriminating categorical data in forest plantations.
The highest VIP values for predicting leaf-cutting ant nests in T. grandis plantations were from band 6 (shortwave infrared1), band 2 (blue), and the MNDWI and PRI vegetation indexes, corroborating other studies [19,35,[54][55][56]. Band 6 was the best for explaining the reflectance of the T. grandis canopy cover [55], and the reflectance of this band was higher in eucalyptus areas attacked by T. peregrinus than in the healthy ones, evidencing a change in the physiological behavior of the plant due to the presence of the insect [19]. Band 2 (blue) was the most efficient for explaining soil reflectance [56] and, therefore, it is linked to the loose soil surface composition of A. sexdens nests. The MNDWI index was developed for the identification of water bodies [36], and it was used in the present work because the vegetation reflects a little more of shortwave infrared1 (band 6) than of green (band 3) [57]. This would allow the detection of defoliation, even light defoliation, by leaf-cutting ants. The PRI index is highly correlated with the canopy and the plant physiological activity [58], and can also detect changes in the structure of the younger leaves [59], which are usually preferred by leaf-cutting ants.
The accuracy of detection of A. sexdens nests in T. grandis plantations corroborates values of 76.7% in the mapping of areas attacked by T. peregrinus in eucalypt plantations in Brazil using Landsat-8 OLI [19]. However, this accuracy was lower than the 96.2% for the mapping of areas attacked by T. peregrinus in South Africa using WorldView-2 [35], a satellite with a spatial resolution of 50 cm, 60 times greater than that of Landsat-8 OLI, and without free images. Another freely available option compared to Landsat-8 OLI is the Sentinel 2 MSI satellite for monitoring insects' pests, with a 2-10 days revisiting period, similar spectral bands, and better spatial resolution (10~20 meters) [60], but calibration and validation studies are yet to be tested. The detection of leaf-cutting ant nests by conventional monitoring, with field workers visually locating and quantifying the nests in Pinus sp. plantations, with sampling intensity of 10.5% and plot of 700 m 2 , showed an expected error of 24% [61]. The estimation error of the number of leaf-cutting ant nests by the Prodan sampling method with a sampling intensity of 10.47% was 13.05% [62]. This reinforces the potential for the use of freely distributed spatial resolution images such as those of the Landsat-8 OLI satellite, with a 16 day revisiting period, to detect and map Atta sexdens leaf-cutting ant nests in forest plantations, as a solution for monitoring these plantations [54].

Conclusions
Medium spatial resolution images of the Landsat-8 OLI satellite using six bands and eight vegetation indices allow for the detection and mapping of the presence of leaf-cutting ant nests in Tectona grandis plantations, with 72.3% accuracy and a low cost (US$0.08/ha), making possible its operational use for monitoring this insect pest in teak forest plantations.