A study on the use of UAV images to improve the separation accuracy of 1 agricultural land areas 2

10 Classifying satellite images with medium spatial resolution such as Landsat, it is usually difficult to 11 distinguish between plant species, and it is impossible to determine the area covered with weeds. In this 12 study, a Landsat 8 image along with UAV images was used to separate pistachio cultivars and separate 13 weed from trees. In order to use the high spatial resolution of UAV images, image fusion was carried out 14 through high-pass filter, wavelet, principal component transformation, BROVEY, IHS and Gram Schmidt 15 methods, and ERGAS, RMSE and correlation criteria were applied to assess their accuracy. The results 16 represented that the wavelet method with R2, RMSE and ERGAS 0.91, 12.22 cm and 2.05 respectively had 17 the highest accuracy in combining these images. Then, images obtained by this method were chosen with 18 a spatial resolution of 20 cm for classification. Different classification methods including unsupervised 19 method, maximum likelihood, minimum distance, fuzzy artmap, perceptron and tree methods were 20 evaluated. Moreover, six soil classes, Ahmad Aghaei, Akbari, Kalleh Ghoochi, Fandoghi and a mixing 21 class of Kalleh Ghoochi and Fandoghi were applied and also three classes of soil, pistachio tree and weeds 22 were extracted from the trees. The results demonstrated that the fuzzy artmap method had the highest 23 accuracy in separating weeds from trees, differentiating various pistachio cultivars with Landsat image and 24 also classification with combined image and had 0.87, 0.79 and 0.87 kappa coefficients respectively. The 25 comparison between pistachio cultivars through Landsat image and combined image showed that the 26 validation accuracy obtained from harvest has raised by 17% because of combination of images. The results 27 of this study indicated that the combination of UAV and Landsat 8 images affects well to separate pistachio 28 cultivars and determine the area covered with weeds.


Introduction 32
Accurate data and statistics could be really important to manage agricultural land areas well (Wardlow et 33 al., 2007) Also, the accurate classified information on a variety of agricultural crops plays a significant role 34 in managing agricultural land areas and it can help evaluate net national product. Precision agriculture (PA) 35 can also help experts maximize production efficiency by providing instant information on cultivated land 36 (Hamidy et al., 2016). 37 The traditional methods applied only through observation of the land to estimate the cultivation area and 38 classification of tree cultivars were very high-priced, time consuming, and not widely applicable. Experts 39 used remote sensing data to discover the type and level of cultivation of each crop, which could give proper 40 information to decision makers (Tatsumi et al., 2015). Satellite data decreases not only human error, but 41 also it can affect in various agricultural programs and lower costs and time. 42 Since there is a balance in the design of satellites between spatial, temporal, and spectral separation power 43 (Emelyanova et al., 2015), because of technical limitations, most satellites cannot simultaneously bring 44 images together with high spatial, temporal, and spectral resolution, and this is a major limitation in using 45 satellite images. 46 Nowadays, as science advances, there has been access to aerial images taken by UAVs 1 (Chianucci et  An OLI 8 image of Landsat 8 satellite and a UAV image were used in order to classify different cultivars of 91 the pistachio tree and also to separate the weeds around the trees. OLI sensors of Landsat 8 gather data for 92 spatial resolution of 30 meters and 8 bands in the visible spectrum, near-infrared, infrared short wavelength 93 and a panchromatic band with a spatial resolution of 15 meters. The UAV image used is an RGB color 94 image, the general specifications of which are given in Table 1. Images of the UAV were taken using Canon 95 8 Operational Land Imager EOS M3 18-55 Digital Camera, the general specifications of which are given in Table 2. The date of 96 imaging was chosen in summer and at the peak of vegetation period. Pistachio phenology includes steps; 97 bloom, leaf out, shell expansion, shell hardening, nut filling, shell splitting, null split, harvest and 98 postharvest. In order to classification, the images related to the nut filling stage have been used, which 99 according to the studies done by Goldhammer, (2005) the peak of vegetation period is related to this stage. 100 Figure 2 shows the diagram of the present study steps. 101 Image fusion is a useful way to provide a more accurate classification which could be an efficient tool to 107 raise the spatial resolution of multispectral images through two images with different spatial, spectral, and 108 temporal resolution. The history of image fusion goes back to the 1950s and 1960s, and it was started to 109 identify the natural and artificial topography, and also the image fusion of different sensors (Wald, 1999). 110 111 2.2.1. Gram-Schmidt method 112 are as follows: 1) simulating a panchromatic image of a spectral band with low spatial resolution 2) 115 Applying GS 9 to a simulated panchromatic image and spectral band using simulation panchromatic band 116 as the first band 3) replacing the high-resolution panchromatic band with the first GS band 4) Using reverse 117 GS to create a panchromatic spectral band (Maurer, 2013 In this method, a high-pass filter is used to get the details of the spatial information of the image with high 127 spatial resolution and to apply those details to the multispectral image (Pohl & van Genderen, 2014). The 128 image created this way is the same as the original multidimensional image, to which the details of the spatial 129 information of the panchromatic image have been added. This method includes the following steps: 1) 130 Applying the high-pass filter on the panchromatic image with high spatial resolution 2) Adding the filtered 131 image to all multispectral images by applying a weighted coefficient on the standard deviation of 132 multispectral bands 3) Adapting the histogram of the combined image with original multispectral image. 133 The HPF method is based on increasing the spatial resolution of a multispectral image using a high-pass 134 filter that extracts high-frequency information and then applies a multispectral image to each band. Where σMSi is the standard deviation of multispectral image bands and σPAN_HPF is the standard deviation of 147 the panchromatic image by applying a high-pass filter. In order to implement the HPF method successfully, 148 the size of the main core filter must be specified, which depends on the R factor. Where PRMS is a multispectral image and PRPAN is a panchromatic image and the optimal size of the core 152 is R2 (Aiazzi et al., 2007). 153

IHS Method 155
The IHS fusion method is one of the most common methods for combining remote sensing images, and this 156 algorithm has been used widely due to the high spatial resolution of the output image and the high efficiency 157 of this algorithm in satellite images (Carper et al., 1990). In fact, IHS is a spectral replacement method that 158 extracts spatial (I) and spectral information (H, S) from a standard RGB image. This method converts the 159 multispectral image color space from RGB space to IHS space, replaces its spatial component with

BROVEY method 169
Brovey is a numerical method in which images are combined by normalizing the pixel values in 170 multispectral image bands and then multiplied by the value of the corresponding pixels in the panchromatic image. In numerical methods, addition and multiplication and the ratio between different bands of 172 multispectral image and panchromatic image are used (Aiazzi et al., 2007). The general equation of this 173 method is as follows: 174 In which the BTii of the band i from combined image, the MSi of the band i from the multispectral image, 176 and PAN is a panchromatic image with high spatial resolution. 177 178

Wavelet method 179
In this method, the spatial information in the panchromatic and multispectral image is extracted by direct 180 wavelet conversion, and the spatial information in the panchromatic image is replaced with or added to the 181 spatial information in the multispectral image. Then, reverse wavelet conversion is done on the conversion 182 coefficients of the converted wavelet of multispectral image (Park & Kang, 2004). The basis of this method 183 is resembles the IHS method and includes the following 6 steps: 184 1) Converting pixel dimensions of multispectral image to panchromatic image 2) Applying IHS conversion 185 to multispectral image and using I, H and S parameters 3) Creating new "P" panchromatic image according 186 to figure I 4) "P" decomposition through wavelet decomposition, also two components of the wavelet image 187 y1 (p) and y2 (p) , and an approximate image of P2 are estimated. Moreover, it is repeated for I. Two components 188 of wavelet image y1 (1) and y2 (1) , and an approximate image of I2 are estimated. 5) Calculation of the 189 difference: Multispectral data can be visualized in a multidimensional space. The dimensions of this space will be the 196 same as the number of image bands, in which each pixel is considered as a vector. The main goal in principal 197 component transformation is to get new components in which the data variance is higher and the 198 dependence between the components is less than the initial state of the images. The fusion of data at the 199 pixel level, which is also called image fusion, has a great variety of algorithms. For this reason, in various 200 applications, researchers have tried to study and analyze the methods used to combine images, and consider 201 classifying the methods, their advantages and disadvantages (Rockinger, 1996). In this research, PCS 202 method is used as one of the main methods of principal component transformation.
  F and R are basic and combined images, μ (R) and μ (F) are the mean of the two images The closer this value is to 1, the greater the degree of correlation between the two images. In order for the data to be more homogenized with the mean, this index provides a better estimation to compare the combination result (Choi et al., 2013).
The closer this value is to zero, the better combination and the less error is (De Carvalho & Meneses, 2000). It is sensitive to the displacement mean and change of dynamic rate. If the value is less than 3, it means that the result of the combination is satisfactory and the combined image is of good quality. Because this index is independent of the unit, it somehow involves the spatial resolution of the source images (Alparone et al., 2004). 216 217

Classification methods 218
In this section, in order to summarize the article, classification methods in this study will be explained 219 briefly. Readers are kindly asked to follow available references in each section. 220 The teaching and learning process requires a set of educational models with optimal inputs and outputs 240  The ultimate goal of image fusion is to get an image that has a higher spatial resolution. The main necessity 269 of all proposed methods in the process of image fusion is to maintain or make the least change in the spectral 270 information of the input images. The purpose of quality assessment is to obtain quantitative and qualitative 271 estimation of the image and also to compare the relative efficiency of different image fusion algorithms. 272  The results of the evaluation of three criteria showed that the wavelet method increases the spatial resolution 302 accuracy by maintaining the spectral information of the image.
Since spatial resolution is one of the factors which determines the accuracy of classification, the combined 304 image was used by the wavelet method to classify and separate weed cover and pistachio tree cultivars. In 305 fact, it will help to classify different vegetation. 306 Table 4. Values of evaluation indices between the corresponding bands in the combined images and 307 Landsat image bands 308  Ghoochi and Fandoghi were extracted to classify pistachio cultivars. Also, three soil classes, pistachio tree 325 and weed were chosen to separate weeds from trees. In order to evaluate the accuracy of the classified maps 326 by different methods, classified maps were compared with the map obtained from the field studies. Then, 327 confuse matrix was formed, and the overall accuracy and kappa coefficient were calculated (Tables 5 and  328 6). It is impossible to identify the weed-covered area and separate it from pistachio trees through Landsat 329 images, and this classification was done only with the combined images and UAV images. The results of 330 the accuracy assessment indicated that the kappa coefficient, overall accuracy and validation using 331 A section from the area under study with six classification methods is shown in Figure 6. In all classification 348 methods, soil contains the highest area, then pistachio and weed are in the following. In the fuzzy artmap 349 method, which is known as the optimal method to separate weeds from pistachio trees, 6% of the area is 350 covered with weeds, 22% contains pistachio trees and 70% is soil. Figure 7 shows a map of separation of 351 pistachio trees from weeds by fuzzy artmap method. 352   )A) )C) )D) 383 of image fusion is to study agricultural uses, natural resources, and to separate plant species, in addition to 392 increasing the spatial resolution of the image, spectral characteristics must also be kept. Therefore, in order 393 to combine images, a method must be used that has acceptable accuracy and can, in addition to improving 394 the location, maintain the spectral content of multispectral images well. Using the appropriate method 395 through image quality evaluation indices depends on the researcher's goal of combining images. Since the 396 accuracy of classification depends on the spatial information in the image, by comparing the results of 397 combining images, it can be observed that by keeping the spectral information of the image, the spatial 398 accuracy is increased to 20 cm. 399 The results of comparison between different classification methods to determine different pistachio 400 cultivars and separate weed from trees indicated that the fuzzy artmap method has the highest accuracy 401 following the maximum likelihood method. This study demonstrated that the product resulted by combining 402 UAV and Landsat images gives the chance to separate weeds that cannot be identified with Landsat images, 403 and also increases the accuracy of classification of pistachio tree cultivars. Moreover, it has a high accuracy 404 of land area and cultivation pattern. The present investigation corrected the map of different forest types, 405 which has prevented the achievement of more desirable results because of the openness of the canopy, 406 mixing of soil, and vegetation reflections. In addition, it showed that by combining Landsat and UAV 407 images and increasing spatial resolution, it would be possible to stop the mixing of soil reflection and 408 vegetation. The study of the area under cultivation of different cultivars through satellite data and preparing 409 land maps and determining the area covered by weeds can be effective in optimal management of these 410 land areas and it is a great way to increase efficiency in the area as well. 411 412