Detection of nematodes in soybean crop by drone

ABSTRACT Global consumption of oilseeds has been growing progressively in the last five growing seasons, in which soybean represents 60% of this sector. Thus, in order to maintain a high production in the region of Rio Verde, State of Goiás, against the phytopathological problems, this study aimed to define the best spectral range for the detection of H. glycines and P. brachyurus by linear regressions in soybean at R3 stage, as well as the elaboration of mathematical models through multiple linear regressions. For this, soil and root were sampled in the experimental area, as well as a flight was performed with the Sentera sensor. Data were used for the elaboration of regressions and for the validation of 2 mathematical models. Significant values were observed in simple linear regression only for cysts, in the visible range, with a good R² value for the Green, Red and 568 nm bands, to nonviable cysts. When working with the stepwise statistics, better results are found for H. glycines, which now has an R²(aj) of 0.7430 and P. brachyurus is then detected. From the mathematical model obtained with the multiple linear regression for non-viable cysts with an R²(aj) of 0.7430, it is possible to detect the spatial distribution of nematodes across the soybean field, in order to perform a localized management, optimizing the applications. Good results are also possible using the mathematical model obtained by simple linear regression.


INTRODUCTION
Global consumption of oilseeds has been growing steadily since the last fi ve growing seasons, in which soybean represents 60% of this sector (UNITED STATES DEPARTMENT OF AGRICULTURE, 2019).However, just like any other crop, soybean suff ers from reduced productivity due to plant pathogens.Among them, the two main ones are, Heterodera glycines, known as soybean cyst nematode, and Pratylenchus brachyurus, the root lesion nematode.
H. glycines is a pathogen that can survive in the soil for many years even without the presence of a host.It is possible because of the dormant stage called cyst, which is a means of survival, under unsuitable conditions for juveniles (MASONBRINK et al., 2019).Symptoms are similar to nutrient and water shortage, which impairs plant development.As a consequence, soybean may exhibit stunted growth, chlorosis, reduced productivity, and even death due to the number of nematodes (BAJWA et al., 2017;NIBLACK, 2005;SONG et al., 2017;ZHANG et al., 2017).
P. brachyurus, a migratory endoparasitic nematode (HOMIAK et al., 2017), causes darkening of the main root system, reduced plant size, reduced grain number and size, and chlorosis in the canopy of the plant, which are similar to the symptoms induced by H. glycines (SANTANA-GOMES et al., 2014).In order to avoid large yield losses, farmers have used non-host crop rotation, nematode-resistant cultivars and chemical and biological nematicides (DUTTA et al., 2019).However, the application of nematicides throughout the area has a very high cost, which somehow encourages the search for alternative measures for the spot mapping of the occurrence of patches.
Precision agriculture is a suitable tool for mapping nematode patches, as it integrates tools such as: Global Navigation Satellite System (GNSS), Geographic Information Systems (GIS), Remote Sensing (RS), Wireless sensors (RSSF) and other techniques, equipment and software for obtaining useful information for agriculture (LÓPEZ et al., 2015;LÓPEZ-RIQUELME et al., 2017;ZHANG et al., 2017).It enables decision making, ongoing crop monitoring, cost savings, increased productivity and intelligent control actions.
Precise detection and mapping of trouble spots in a fi eld are products obtained using RS and GIS.The latter allow for better management practices, with localized treatment and within the timing of the crop.Several studies show the ability to detect the spatial distribution of a pathogen through thermography, spectroradiometers, multispectral and hyperspectral sensors on plant canopy spectral response (BAJWA et al., 2017;JOALLAND et al., 2017;MARTINELLI et al., 2015;MARTINS et al., 2017).
Orbital RS allows for a diagnosis of large areas in a short time and the determination of severity levels of nematode-infected plants (MARTINS et al., 2017).With the miniaturization of sensor systems, drones allow timely mapping for detection of pathogens in large areas (YANG et al., 2016), mapping of soil fertility responses (SCHUT et al., 2018) and yield estimates (JEONG et al., 2018).
For the infestation scenario present in our study areas, soybean health characterization maps obtained through unmanned aerial vehicles (UAVs), associated with directed sampling, can lead to fast and reliable methods for georeferenced detection of levels of nematode infestation.Thus, this study aimed to defi ne the best spectral range for the detection of H. glycines and P. brachyurus by simple linear regressions in soybean crop at R3 stage, as well as the elaboration of mathematical models by multiple linear regressions.

Study area characterization
The municipalities of Rio Verde and Montividiu, located in the State of Goiás and in places where their lands are valued for high fertility and high rainfall, have good topography conditions for grain management.However, the fi eld located to the south of the state is marked by the presence of H. glycines, P. brachyurus and Helicotylenchus dihystera nematodes, as evidenced by the nematological analysis carried out by the Nematology Laboratory of the Federal Institute of Goiás (Figure 1).The experimental field was characterized by a topography with a slope of less than 5% to the southeast.No-till was predominant with the use of autopilot.The main location of the patches was in an old road that crossed the field, of which the last one has an area of approximately 330 hectares.

Material
A total of 45 sites were sampled to the elaboration of regressions and 60 sites for validation of two mathematical models: i. simple linear regression; and ii.multiple linear regression.An Unmanned Aerial Vehicle (UAV), known as Inspire 2, and a modifi ed sensor with a total of 12 spectral bands were used to fl y over the experimental area.
The camera aboard Inspire 2, known as Sentera, had 8 sensors, some of which were in the visible and others outside this spectral range.The device have the wavelength band of 615, 586, 661, 825 and 775 nanometers (nm), a composition sensor R (650 nm), G (548 nm), B (446 nm), a NDVI sensor containing one Red (625 nm) and one Nir (850 nm) band and a NDRE sensor containing another Nir band (840 nm) and one known as RedEdge (720 nm).
The 45 sampling sites, mapped by means of a georeferenced orthomosaic from the 2017/18 growing season, were identifi ed in the fi eld by a white sheet so that after the fl ight they could be precisely located.To reach the sites to be sampled, we used a Garmin eTrex 20 receiver.The FieldAgent application was used to perform the fl ight over the patches, the Pix4d software to obtain the orthomosaic and the Qgis software for extracting the information from the pixels of the sampled areas and mapping.Statistical analysis and data validation were performed in Minitab statistical software.

Nematode sampling
The soybean fi eld used for soil and root sampling was planted on October 10, 2018 with the early cultivar MONSOY 7198, resistant to soybean cyst nematode races 1 and 3.The history of the area was known for the presence of H. glycines of races 1,3 and 6, P. brachyurus and H. dihystera.
A total of 105 nematode samples were collected, 45 for the regression calculation and 60 for the validation of mathematical models, with 3 subsamples for each site.The validation samples refer to an experimental area containing 60 plots.All samplings were obtained with a maximum of one day difference.The samples for the elaboration of the single and multiple linear regression models were divided into 5 patches, so that one sampling site was at the center and the other in two different directions, 10, 20, 40 and 80 m from the center of the patch (Figure 2).Samples were taken during the R3 stage of soybean at a depth of 0 to 20 cm.This stage was shrunk due to the reduction of soil interference with the value of the pixels used in the statistical analysis.For each site, besides the soil, the root of the cultivar was collected.Therefore, the samples were sent to the Phytopathology Laboratory of the Federal Institute of Goiás -Campus Rio Verde, for the extraction of juveniles, females and cysts and identifi cation of nematodes.Races of H. glycines were identifi ed in another laboratory, and they were forwarded by the farmer who allowed the research in his property.
The extraction of juvenile P. brachyurus, H. dihystera and H. glycines in soil and root were performed according to the methods of Jenkins (1964), Coolen and D'Herde (1972) and Alfenas and Mafi a (2007).For viable and non-viable cysts in soil, we used the methodology adapted by Araújo (2009) and for females in root, the method adapted by Tihohod (2000).

Flight planning and orthomosaic construction
The fl ight was performed on the same sampling days, between 10:00 and 14:00, so that all sites were overfl own.In the FieldAgent application, we set a 60% overlap between front and side of the images at a fl ight height of 120 meters, so that a battery was suffi cient to fl y over all the patches.Heterogeneous lighting conditions during photo capture were avoided as they alter the spectral response of plants and the results of statistical analysis, even with the use of the irradiance sensor.
Prior to the fl ight over the patches, some blank targets were placed at the sites to be sampled to facilitate identifi cation in the image.It is important that the image information is extracted as close as possible to the soil and root collection, such that the spectral response refers to what is present at the site of the collected plant.For the experimental area, the plots were delimited by targets at their ends.
Orthomosaic was constructed by photogrammetric processing in the Pix4d software, with the provision of the camera's interior and exterior orientation parameters and the calculated point cloud.Therefore, no calibration panel was used for the refl ectance calculation.

Image information extraction and prediction map construction
Through the orthomosaic, in each plant referring to the nematode data sampling sites, information on the amount of light refl ected by the leaf canopy was extracted so that the value could be used in single and multiple linear regressions.The extraction for the 45 sites was performed in each of the 12 Sentera bands with the feature identify tool.The latter allows, with just one click on the image, the pixel value is expressed.
For each of the 45 samples of patches, the value of the light refl ected by the canopy, used in the statistical analysis of single and multiple linear regressions, was the result of an average of 10 pixels extracted by the option identify features, without the presence of shade and soil.
As for obtaining the estimated value of non-viable cysts from the prediction map, we considered the pixel coincident with or overlapping the validation point located at the center of the plots.The prediction map of the single and multiple linear regression was generated using the raster calculator, with the mathematical equation of each of the two best regressions for non-viable cysts.

Statistical analysis and validation of two mathematical models
The correlations performed were between the nematode data and the pixel value of the images.Relationships between juveniles, cysts and females were analyzed with the amount of light refl ected by the plant canopy for each of the sites and bands.
However, H. dihystera was not included in the statistical analysis because it is an ectoparasite that does not cause significant damage to soybean crop (ANTÔNIO, 1992).For simple linear regressions, only those significant at 5% were considered.The normality and independence test was applied to the residuals as well as the homogeneity of variance.
For multiple linear regressions, we used the forward stepwise methodology, which starts with a simple linear regression, and new variables are inserted into the model according to the adopted signifi cance level (CHATTERJEE;HADI, 2015).It is only completed when the best mathematical model found is repeated again.A p-value of 0.05 was adopted in the model, uncorrelated independent variables and with the lowest possible Mallows Cp.Finally, normality tests and residual independence and homogeneity of variance were also performed.
For validation of the prediction model, from the measured and estimated data, the root-mean square error (RMSE) and the ERROR (%) were calculated.RMSE and ERROR (%) are obtained through equations 1 and 2. In the case of viable cysts, the best detection band is the same as for non-viable cysts, but with an R² of 0.3766.Green and Red also stood out, with an R² of 0.3198 and 0.3481, respectively.However, cysts that have not yet hatched should not be taken into consideration for the detection of soybean cyst nematode, as they do not cause symptoms.
Table 2 presents a mathematical model elaborated using stepwise multiple linear regression, which presents better results than simple linear regressions for non-viable The prediction variables of the mathematical model for non-viable cysts consider two bands that showed good correlation results in simple linear regression, and one that is out of the visible range, which is 825 nm.None of the independent bands was correlated with each other, with the calculated VIF equal to 1 (TAMURA et al., 2019).Mallows Cp was 6.8, lower than other multiple linear regression models for non-viable cysts, which had an R² less than 0.7430.Thus, it was prioritizing Mallows Cp as close as possible to the number of predictors (MALLOWS, 2000).
In the case of P. brachyurus, although the multicollinearity of the model is adequate, the Mallows Cp of 8.5 indicates a low model precision and a high variance of regression coeffi cients.In Figure 3, there is the true-color composition of 3 patches, where samples were taken.The only patch of all studied, which was evident in the image, was the patch -1, due to the larger amount of bare soil due to a smaller canopy size.
The subtle symptoms to human eyes present in the RGB composite image become more evident when a thematic map for predicting non-viable cysts is drawn through the 586 nm band for simple linear Detection of nematodes in soybean crop by drone  They have similar but distinct results.The model using simple linear regression presented an RMSE of 14.81, with an error of 53.97%.In turn, the mathematical model, somewhat better, presented an RMSE of 12.78 and an error of 46.59% (Figure 5).Such values are satisfactory, considering that the nematode population can vary greatly from site to site within a few tenths of a centimeter.They all had a p-value less than 0.05.
Although prediction maps have a high error, they can be used for preliminary detection of sites with higher and lower occurrence of H. glycines, so that soil sampling can be directed and quantification improved.This is valid as the areas to be monitored are extensive and the soil sampling is expensive and time consuming (MARTINS et al., 2017).In summary, with some soil samples and the georeferenced thematic map, it is possible to reduce the costs of nematicide application (ALJAAFRI, 2017), which makes control sustainable and viable, with intelligent application.

CONCLUSIONS
1.For detection of H. glycines, the best spectral range was 586 nm, which presented an R² of 0.649 for non-viable cysts in soil.In simple linear regression, nothing was significant for the detection of P. brachyurus; 2. However, when the Green and Nir-NDVI bands are combined in a mathematical model, P. brachyurus in soil becomes significant and detected.In the case of the mathematical model for detection of H. glycines, specifically the non-viable cysts, a significant improvement in R² is achieved, combining the Red, Green and 825 nm bands; 3. By validating the two mathematical models, both for detection of non-viable cysts in soil, it was possible to prove the efficiency in the detection of sites with larger and smaller amounts of nematodes.This is important so that the number of soil samples can be reduced and targeted so as to produce a good application map at a variable rate.The best prediction result was the one that used multiple linear regression.

Figure 1 -
Figure 1 -Location of the experimental area and the patches

Figure 2 -
Figure 2 -Sites outside the plots of the experimental area comprise the samples used for the elaboration of linear regressions and sites inside the plots for the validation of regressions and in situ measured value of non-viable cysts in 100 cm³ soil, respectively; and n is the number of samples.B. H. T. Arantes et al.

Figure 3 -
Figure 3 -RGB composition of some sampled areas

Figure 5
Figure 5 -A-RMSE and ERROR of simple linear regression with 586 nm band for detection of non-viable cysts; B -RMSE and ERROR of the mathematical model for detection of non-viable cysts

Figure 5
Figure5presents the performance of the prediction model expressed by the root mean square error (RMSE).They have similar but distinct results.The model using simple linear regression presented an RMSE of 14.81, with an error of 53.97%.In turn, the mathematical model, somewhat better, presented an RMSE of 12.78 and an error of 46.59% (Figure5).Such values are satisfactory, considering that the nematode population can vary greatly from site to site within a few tenths of a centimeter.They all had a p-value less than 0.05.