Aerial multispectral imagery for plant disease detection: radiometric calibration necessity assessment

This paper focused on the necessity of radiometric calibration to distinguish diseased trees in orchards based on aerial multi-spectral images. For this purpose, two study sites were selected where multispectral images were collected using a multirotor UAV. The impact of radiometric correction on plant disease detection was assessed in two ways: 1) comparison of separability between the healthy and diseased classes using T-test and entropy distances; 2) radiometric calibration effect on the accuracy of classification. The experimental results showed the insignificant effect of radiometric calibration on separability criteria. In the second strategy, the experimental results showed that radiometric calibration had a negligible effect on the accuracy of classification. As a result, the overall accuracy and kappa values for un-calibrated and calibrated orthomosaic classifications of the citrus orchard were 96.49%, 0.941, 96.57% and 0.942, respectively, using five spectral bands as well as DVI, NDRE, NDVI and GNDVI vegetation indices using a random forest classifier. The experimental results were also similar at the other study site. Therefore, the overall accuracy and kappa values for the un-calibrated and calibrated orthomosaic classifications were 95.58%, 0.913, 95.56% and 0.913, respectively, using five spectral bands as well as NDRE, BNDVI, GNDVI, DVI, and NDVI vegetation indices. ARTICLE HISTORY Received 17 February 2019 Revised 13 June 2019 Accepted 8 July 2019


Introduction
In recent years, the prediction of effective challenges on the crop yield such as climate change effects, diseases, and pests has become a critical issue in food managing strategies (Onisimo, Timothy, & Omer, 2017). Furthermore, because of the population increase and instability of agricultural production due to climate variability, some methods and techniques have been developed for monitoring the condition of plants as well as the quality and quantity of agricultural products.
Apparently, the inspection procedure for the early detection of the diseases is extremely time-consuming. Therefore, it is necessary to find a detection process which will be faster than the field inspection by experts (Psirofonia, Samaritakis, Eliopoulos, & Potamitis, 2017). The limitations of direct field inspection methods have led researchers to investigate remote sensing data to quickly and inexpensively obtain plant health information. Non-destructive remote sensing data permits measurement of biophysical and biochemical parameters of plants for nondestructive monitoring of plant growth and health (Lu et al., 2015;Sankaran, Ehsani, Inch, & Ploetz, 2012;Sankaran, Mishra, Ehsani, & Davis, 2010). These methods are based on the optical properties of plants and can be used for identifying nutrient deficiencies, plants diseases, water stress and excess water, insect damage, weeds, etc. (Baret, 2016;Haboudane, Miller, Pattey, Zarco-Tejada, & Strachan, 2004;Hede, Kashiwaya, Koike, & Sakurai, 2015;Sankaran et al., 2010). Combination of plant knowledge and remote sensing data can provide information for plant management and even initial advices about plant stress to prevent the spread of disease or pest infestation by adopting appropriate reactions in the early stages of stress.
One of the main steps of image post-processing in remote sensing applications is the extraction of absolute reflectance measurements from the data in the radiometric calibration process (Kelcey & Lucieer, 2012). It should be noted that the necessity of camera calibration is an open question. Therefore, this study investigated the potential of UAV-based multispectral imagery in plant disease detection and especially concentrated on the necessity of radiometric calibration.
The first study area was a citrus orchard with the Greening disease. The citrus Greening disease has become one of the greatest challenges for citrus growers across the world (Cevallos-Cevallos, Futch, Shilts, Folimonova, & Reyes-De-Corcuera, 2012). Citrus Greening, transmitted by grafting from citrus to citrus, can severely weaken citrus trees, causing reductions in fruit yield and quality. A sample of Greening symptoms on citrus tree leaves is presented in Figure 1(a).
The other study area is a peach orchard with the peach leaf curl (PLC) disease. The PLC fungal disease is one of the most common diseases of peach trees in the north of Iran that mostly affects peaches leaves and nectarines. PLC is caused by the ascomycetous fungus Taphrina deformans and causes parts of individual peach leaves to become curled and pale green or reddish in color (Figure 1(b)) (Moscatello et al., 2017). Taphrina deformans primarily affects a foliage, but may also affect fruits and young twigs (Pecknold, 2015).
To investigate the potential of the nutrient deficiency using UAV-based multispectral images, Agüera et al. compared the correlation between the calculated NDVI from UAV-based multispectral images and the ground-based multispectral radiometer with the nitrogen status in a sunflower field (Agüera et al., 2011). Results proved a higher correlation coefficient (R = 0.80) using the UAV platform than a ground-based measurement (R = 0.71). In another study, Garcia-Ruiz et al. investigated the capability of UAV-based multispectral images using for citrus greening detection in comparison with similar images captured based on piloted aircraft (Garcia-Ruiz et al., 2013). They achieved a classification accuracy of 67-85% and 61-74% based on UAV-based and aircraft-based datasets, respectively, using six spectral bands as well as NDVI, GNDVI, SAVI, NIR-R, R/NIR, green(G)/R and NIR/R vegetation indices. To detect the citrus Greening disease, Sarkar et al. also presented an UAS sensor set including an onboard active spectrally selective light source and an RGB-D sensor that measures distance and reflectance of rotated polarized light (Sarkar et al., 2016). They used a liner SVM classifier and reported 93% classification accuracy.
De Castro et al., evaluated the spatial and spectral of multispectral requirements of images for quick and accurate detection of laurel wilt (LW) disease in an avocado field, considering the change in the flight altitudes and spectral sensors. Results demonstrated that most effective spectral wavelengths were 580-10 nm, 650-10 nm, 740-10 nm, 750-10 nm, 760-10 nm and 850-40 nm with 15.3 cm spatial resolution and the optimum vegetation indices were TCARI760-650, GNDVI, NIR/G, Redge/G and VIGreen using Redge or NIR bands (De Castro et al., 2015).
Albetis et al. evaluated the capability of spectral features computed from the UAV multispectral imagery including spectral bands, vegetation indices and biophysical parameters for discriminating the Flavescence doréesymptoms disease in red and white cultivars from healthy vine vegetation. Results presented best with red cultivars, but were not acceptable for white cultivars (Albetis et al., 2017).
The radiometric calibration of the UAV-based multispectral images is one of the main steps of image processing in remote sensing applications (Kelcey & Lucieer, 2012). For the purpose, especially in UAV-based remote sensing applications, usually an image-based empirical approach, called empirical lines based methods, is used (Berni, Zarco-Tejada, Suárez, & Fereres, 2009). This technique is based on the assumption of an empirical relationship between the image DNs and the at-surface reflectance which is made as a field measure of the appropriate targets (Smith & Milton, 1999).
Kelcey and Lucieer for radiometric calibration of UAV-based multispectral images of mini-MCA camera, used three radiometric targets made of a plywood and backing with a white paperish fabric (Tyvek), grey fabric and black fabric covers. The targets' at-surface reflectance were measured by an ASD HandHeld 2 Portable Spectrometer (Kelcey & Lucieer, 2012). In the study, two black and white calibration targets with 3% and 82% reflectance, respectively, were used during flight for radiometric correction in empirical line calibration method which was carried out in ENVI software (De Castro et al., 2015).
On the other hand, along with a UAV-based multispectral cameras, a white calibration panel is usually provided for radiometric calibration. Miura and Huete evaluated three white calibration panel-based radiometric calibration methods in the hypersonic airborne data, called "reflectance mode", "linear interpolation" and "continuous panel" (Miura & Huete, 2009). These methods used white calibration panel to convert DN to at-surface reflectance. Clemens modified the "reflectance mode" method from Miura and Huete (2009) to reduce the reflectance value conversion bias (Clemens, 2012). In this study, an average of the pre-and post-flight panel readings in the reflectance conversion is used which assumed a linearity due to the short time of flight. Garcia-Ruiz et al. used whiteboard for radiometric calibration of the UAV-based images to ensure that 100% reflectance corresponding with a digital number of 255 (Garcia-Ruiz et al., 2013). Albetis et al. also used the calibrated ground panel, before and after the flight, to convert the DN to at-surface reflectance in each band (Albetis et al., 2017).
As shown in the literature, radiometric calibration is considered as the main step in UAV-based plant disease detection, and measurements of reference targets are implemented in empirical line-based methods by default, while the efficacy and necessity of this step is unclear. Therefore, the efficacy and necessity of the radiometric calibration step in plant disease detection using UAV-based multispectral images were investigated in this study.

Study area and imaging system
Two different agricultural fields located in the north and south of Iran were selected as the study area. The first one, a citrus orchard outbreak with Greening disease, is located in Fasa County, Fars Province, Iran (Figure 2(a)). The second field is located in Neka City, Mazandaran Province, Iran, which is a peach orchard affected by the PLC disease ( Figure  2(b)).
A small multirotor platform that is equipped with Micasense RedEdge multispectral camera, which is especially designed for agriculture applications, is selected for imaging of both test fields. A summary on the specifications of both platform and imaging system is presented in Table 1.
To achieve reliable and efficient multispectral images, accurate flight planning is needed in order to achieve optimum image resolution and accuracy at the best flight time and cost. Some details on the flight planning parameters in the both study areas are presented in Table 2.
Disease and healthy trees as ground truth were selected by the experts' visual inspection and labeled using special targets ( Figure 5) before image acquisition to train and test the classifier. Therefore, 15 citrus Greening infected and 10 healthy trees at the Fasa study area as well as 20 PLC infected and 12 healthy trees at the other study area were selected as ground truth data. Figure 3 summarizes the workflow of proposed methodology for plants disease detection based on multispectral imagery and assessment of radiometric calibration evaluation necessity.

Processing methodology
Acquired multilens images were registered bandto-band in the first step and then aligned in the bundle adjustment process. Regarding the image acquisition under different illumination conditions, the normalization of image blocks is necessary in order to produce homogeneous block which is done based on radiometric block adjustment. Finally, Band registered aligned Images stitched together to produce orthomosaic. In (Albetis et al., 2017;De Castro et al., 2015;Garcia-Ruiz et al., 2013;Honkavaara et al., 2013;Kelcey & Lucieer, 2012;Wijesingha, 2016), commonly radiometric calibration is done as the main step that converts DN to reflectance. In order to evaluate the radiometric calibration necessity of multispectral images and to inspect its effects on the quality of final results, classification process is performed both on DN (Strategy II) and reflections based data (Strategy I) and results were evaluated and discussed. The classification process consisted of three main steps: feature space generation; train classifier and performs classification to generate disease map.

Band-to-band registration
Recent developed UAV-based multispectral cameras are usually composed of multilens structure ( Figure 4). Due to the low altitude of flight in small UAV, Band missregistration of multispectral images acquired by multilenses cameras is a great challenge (Jhan, Rau, & Huang, 2016). In this study, SIFT 1 algorithm is applied to extract corresponding features from images. It is well suited for the feature detection and description because it is able to handle the special characteristics of the input data such as low image contrast as well as rotation, translation and scale differences (Lindeberg, 2012).   After the extraction of feature points using SIFT algorithm, the matching process can be done by comparing the minimum Euclidean distance between the extracted feature vectors. With the matched features, the mathematical relationship between the two images will be computable. Finally, the images were registered and resampled based on the projection transformation (Shi, Wang, & Xu, 2011).

Image alignment
Image alignment provides interior and exterior orientation parameters in the bundle adjustment process that use tie points in overlapping images (Nissen, Johnson, Fitzgerald, Morgan, & White, 2014). Tie points could be extracted by the SIFT algorithm. The SIFT algorithm detects features and matches corresponding features in overlapping images; moreover, outliers and improper matches are eliminated simultaneously (Nesbit, 2014). Thus, a sparse point cloud will be an output of the bundle adjustment (Wijesingha, 2016). This process is carried out effectively due to consideration of sufficient overlap and side laps of images during a flight mission.

Radiometric normalization
Regarding the image acquisition under different illumination conditions, the radiometric normalization of image blocks is necessary in order to produce homogeneous orthomosaic. According to Gehrke  and Beshah (2016), the radiometric normalization parameters for all images in the block are computed in the least-squares adjustment using radiometric tie points. The main observation equation for each radiometric tie point (extracted geometric tie points in image alignment step) between two images can be expressed as Equation (1): (1) where c xy and b xy are the normalization parameters.
For the radiometric control points on the reference image, Equation (1) is rewritten as Equation (2): Radiometric control points were applied to define the "radiometric datum" (Gehrke & Beshah, 2016).

Orthomosaic generation
Because of the small footprint of each image, the acquired images were geometrically corrected and merged to produce a seamless orthomosaic of the whole area.
In UAV imagery-based applications, dense point cloud extraction based on the aligned images is a main step commonly (Candiago et al., 2015;Näsi et al., 2015;Nevalainen et al., 2017;Wijesingha, 2016). Finally, orthomosaic was generated based on aligned images and dense point cloud; Thus, the topographic relief of trees was considered there. However, in this study, an orthomosaic without trees topographic relief consideration satisfied our goals. Therefore, orthomosaic is generated based on aligned images and one coarse DSM producing from the sparse point cloud.

Radiometric calibration
One of the main steps in image processing for remote sensing applications is the radiometric calibration process where the absolute reflectance measurements from the DN data is extracted (Kelcey & Lucieer, 2012). Acquired images could be calibrated using radiometric reference images that captured from a calibrated reflectance panel before and after each flight (Clemens, 2012).
In this study, "reflectance mode" method is applied for radiometric calibration based on Clemens (2012). In the Clemens method, an average of the pre-and post-flight panel measurements is used in the reflectance conversion. Equation (3) presents the basis of the reflectance factor calculation in the reflectance mode method.
where DN T is digital numbers of the target; DN R is the average of the pre-and post-flight in reference in images; and R R is the reflectance factor of the white panel which will determine R T , the reflectance factor of the unknown surface.

Feature space generation
One of the simplest techniques to map the vegetation is to use vegetation indices that are easily calculable and understandable (P. C. Joshi, 2011). Numerous vegetation indices are calculated as the difference or ratio of two or several bands in the wide range of the spectrum (e.g. NDVI, RVI, etc.). They constitute the simple and straight method to extract highly correlated information to biophysical parameters of plants from remotely sensed data. After extraction of ortho images, spectral features including five spectral bands and 12 vegetation indices were extracted to be applied in the classification (Table 3). In Table 3, B, G, R, RE, and NIR are the blue, green, red, red edge, and near infrared bands of the multispectral images, respectively. After calculating the spectral features, the most suitable features that discriminated between the healthy and unhealthy trees were selected in this study. Decreasing the number of features, increases the generalization capability and reduces the computational complexity of the classification algorithm (Guyon & Elisseeff, 2003). The T-test and entropy distances measure the separability between two classes and can be used as criteria to evaluate the efficacy of each feature for discriminating between healthy and unhealthy trees classes.

Classification
After feature space generation, the final step towards health map generation is the classification of images based on selected training data. For this purpose, a random forest classifier, developed by Breiman (2001), is used due to its simplicity and high performance (Feng, Liu, & Gong, 2015;A. Joshi, Monnier, Betke, & Sclaroff, 2015). Random forests non-parametric classifier is an ensemble-based machine-learning algorithm that engages the multiple decision tree classifiers in a voting strategy to provide the final prediction. It consists of several decision tree classifiers that build using a bootstrap sample of the data with a random variable set at each node to split on. Implementation of random forest classifiers requires a few parameters, including the number of trees and the number of randomly selected predictor variables (Rodriguez-Galiano, Ghimire, Rogan, Chica-Olmo, & Rigol-Sanchez, 2012). So random forest offer less complex computations and running time, as well as high performances in the computer vision domain especially (REZAEIAN, AMIRFATTAHI, & SADRI, 2014). In this study, we used a random forest classifier consisting of 100 decision trees.
The final accuracy in detection of the diseased and healthy trees is tested by calculating error matrices of the test data (references). The error matrix presents the error of omission, error of commission, overall accuracy and kappa (κ) values (references).

Experiments and results
In this study, we evaluated the impact of radiometric calibration on UAV-based multispectral plant disease detection using the two datasets with different diseases and environmental conditions. For this, 557 images from the Fasa test area and 683 images from the Neka test area were captured based on the flight plan.

Band-to-band registration
The first processing step is spectral bands registration of images. Due to the multilenses structure of a multispectral camera, a considerable geometric displacement between images captured by lenses appeared. To remove the displacements, the R band was selected as the master and the rest were considered as slaves which should be registered to the master. The band registration of all the images were conducted based on the proposed strategy and the RMSE value of 0.4 pixels was available based on the target measurement. Figure 5 shows two false color images of the same scene: original image and band registered image as a sample.

Image alignment and seamless orthomosaic generation
After band registration step, a seamless orthomosaic of the whole area covering in the imagery step should be generated with five spectral bands. For this purpose, multispectral images must be aligned. In this step, geometric tie points were extracted, and image interior and exterior parameters were estimated. Table 4 presents the summary of the block adjustment parameters.
Based on the aligned band registered images, a dense point cloud was produced and used in orthomosaics generation. The specification of the dense point cloud and orthomosaics are presented in Table  5. The generated orthomosaics are depicted in Figure 6. As is depicted in Figure 6, a seamless orthomosaics was generated after radiometric normalization which was prepared for the feature extraction and classification processing step.

Radiometric calibration
For the radiometric calibration process, reference images of a white panel were captured immediately before and after each flight. A sample image acquired from the calibrated reflectance panel is presented in Figure 7.
Maximum DNs in the images captured from the white panel, before and after the flight (BF and AF, respectively), are listed in Table 6 for the both study data sets. Note that the sensor outputs 12-bit data are stored as 16-bit TIFFs.
DNs values are related with at-sensor radiance directly; Thus, Table 6 shows the decrease in at-   sensor radiance during the both flights. The rate of this reduction was much higher at the Neka site, which indicated a significant change in illumination conditions during that flight. The mean DN values of BF and AF were used in Equation (3) to calculate the reflectance-based values and new orthomosaics were generated based on these new values for each study data.
In order to investigate the necessity of sensor calibration for vegetation disease detection, separability between healthy and un-healthy tree classes, related features were investigated using T-test and entropy criteria using ground truth data collected from the calibrated and un-calibrated data for the both study areas. The experimental results are presented in Figure 8.
Considering the achieved values presented in Figure 8, it can be concluded that the camera calibration did not have any effects on the result based on Ttest and entropy distances. However, comparing the extracted feature values in ground truth data   measured in calibrated and un-calibrated orthomosaics states partly different values in the both study areas (Figure 9).

Classification
Three main classes were determined as un-healthy trees, healthy trees and non-trees at the study sites.
Using the red and near infra-red spectral bands, the non-tree class was easily separated exactly from the trees classes. However, the detection of healthy and un-healthy trees classes remained as the main challenge. Five spectral bands, as well as some VI that have high discrimination between healthy and un-healthy classes based on T-test and entropy criteria ( Figure 8) and low correlation together (Table 7) were used in a random forest classifier consisting of 100 decision trees.
From Table 7, it is obvious that there were higher correlations between RVI, NDVI, IPVI, OSAVI, and CI, between BNDVI and SIPI, and between GNDVI and GI; similar results were obtained for the other study site.
At the Fasa study site, five spectral bands as well as DVI, NDRE, NDVI, and GNDVI vegetation indices were used in the classification process of un-calibrated and calibrated orthomosaics. Figure 10 shows the generated map of the healthy and Greening infected trees and non-trees. There was an insignificant difference in the classification maps ( Figure 10). Small differences in results were observable, although were not significant.
The overall accuracy and kappa values for uncalibrated and calibrated based orthomosaic classifications were 96.49%, 0.941, 96.57% and 0.942, respectively. The confusion matrix and estimated user's and producer's accuracy measures are also presented in Table 8 (un-calibrated orthomosaic) and Table 9 (calibrated orthomosaic).
At the Neka study site, classification of the uncalibrated and calibrated orthomosaics is performed using five spectral bands as well as NDRE, Figure 9. The feature values in ground truth data collected from calibrated and un-calibrated based orthomosaics. BNDVI, GNDVI, DVI, and NDVI vegetation indices. Figure 11 the classification result of healthy, Leaf Peach Curve infected and non-trees at the Neka study site. As in the Fasa data set, no significant difference in the classification map is observable ( Figure 11). Achieved overall accuracy and kappa values for the un-calibrated and calibrated orthomosaic classifications were 95.58%, 0.913, 95.56% and 0.913, respectively. The confusion matrix and estimated user's and producer's accuracy measures are also computed and presented in Tables 10 and 11, respectively. Inspecting results, it can be concluded that radiometric calibration of images has a minor effect on the classification accuracy. In the both statuses, almost the same precision was achieved and the confusion matrix components were not significantly different.

Discussion
Radiometric calibration is mentioned as one of the main steps in UAV-based plant disease detection, and measurements of reference targets are implemented in empirical line-based methods by default. To study the necessity of this process, in this paper, two study data with different environmental conditions and affected by different diseases were classified to generate a tree health state map. Therefore, different processing steps were analyzed and the focus was on the necessity of radiometric calibration.
The multi-lenses structure of the multispectral camera in low flight altitude missions created a considerable geometric displacement between the spectral bands ( Figure 5) which is up to 25 pixels for our data. Here, the maximum height difference is about 10 m. Ignoring this displacement will result in different spectral features for the same objects and will lead to unsuitable training data. Therefore, the band-toband registration step is important in the application; here, even leaf size is important and will reduce the registration error up to an average RMSE of 0.4 Pixel. It has to be mentioned that in images with extremely small relief variances with respect to flight height or with not much small focusing elements, the registration process can be neglected. In this situation, the ortho generation process can reduce the displacement to an acceptable level.
Orthomosaic is generated based on aligned images with the accuracy of 0.85 pixel and generated a dense point cloud. However, different illumination conditions in each capturing station heterogeneous final   orthomosaics (Figure 6-Left) and corrupts training data like the previous step. The next step to prepare the training data is radiometric calibration which is performed based on empirical lines and changes pixel values linearly in each band. Therefore, outputs of this step can change the spectral features values. As shown in Figure 9, variation in most of the VIs is small due to their relative nature.
The detection of the healthy and unhealthy tree classes was performed using a random forest classifier. For this purpose, feature space was produced, including five spectral bands and 12 vegetation indices. The best features were selected based on the separation criteria of the tree classes and the correlation between the features. The correlation ranges from 94.5% to 100% between RVI, NDVI, IPVI, OSAVI, and CI, is 94.7% between BNDVI and SIPI, and is 97.6% between GNDVI and GI (Table 7); thus, six features can be eliminated. Moreover, at the Fasa study site, the spectral bands DVI and NDRE made the highest separation whereas the spectral band BNDVI made the lowest separation between the healthy and unhealthy tree classes. On the other hand, at the Neka study site, NDRE and BNDVI made the highest separation while GRVI made the lowest separation between classes. Classification was carried out using the selected features, and results are presented in Table 8 to Table 11. At the Fasa study site, citrus orchard with the greening disease, the overall accuracy and kappa values for the un-calibrated and calibrated orthomosaic classifications were 96.49%, 0.941, 96.57% and 0.942, respectively. Similarly, the experimental results at the Neka study site with PLC yielded the same conclusions. The overall accuracy and kappa values for the un-calibrated and calibrated orthomosaic classifications were 95.58%, 0.913, 95.56% and 0.913, respectively.   Figure 11. The classification map of healthy trees (green), leaf peach curve infected trees (yellow) and non-trees (red) or the uncalibrated (left) and calibrated (right) orthomosaics at the Neka study site.
However, because of the linear structure of radiometric calibration, separation of the healthy and unhealthy tree classes based on t-test and entropy criteria did not change using the most spectral features ( Figure 8); consequently, there were insignificant differences in the classification results after radiometric calibration.

Conclusions
This study investigated the necessity of radiometric calibration in UAV-based multispectral imagery for plant tree disease detection and classification applications. For this purpose, two study sites with different conditions of climate, date, tree species and disease types were selected and multispectral images were captured using a RedEdge camera mounted on a multirotor UAV. Due to the multi-lens structure of the camera and the low flying altitude, the band-to-band miss-registration error was not negligible; therefore, in the first step, they were registered. Then, to generate seamless orthomosaics, the registered images were aligned and radiometric normalization was implemented.
In order to produce a homogeneous orthomosaic, radiometric normalization as a part of the post-processing step is essential for generating uniformity in the block of images that acquired under a wide variety of environmental conditions. Therefore, in the radiometric calibration step, only the conversion of digital numbers to reflectance values is applied.
Radiometric calibration effects on plant classification and disease detection were evaluated in two scenarios. First, discrimination of the healthy and un-healthy classes was compared using T-test and entropy distances considering ground truth data collected from calibrated and un-calibrated orthomosaics. The experimental results showed that the effects of radiometric calibration on separability criteria using most spectral features were insignificant . In the other scenario, the radiometric calibration effects on the accuracy of classification were investigated. For this purpose, the best features were selected based on the separation criteria of the tree classes and the correlation between the features. The random forest classifier was applied to detect diseased trees. Experimental results showed insignificant differences in the classification results after radiometric calibration. At the Fasa study site, the overall accuracy and kappa values for un-calibrated and calibrated orthomosaic classifications were 96.49%, 0.941, 96.57% and 0.942, respectively. On the other study site, the overall accuracy and kappa values for un-calibrated and calibrated orthomosaic classifications were 95.58%, 0.913, 95.56% and 0.913.
As the results showed, radiometric calibration had a minor effect on the classification results. This can be attributed to the fact that images used for the classification were acquired at a single date with consistent atmospheric and sun illumination condition. Exclusion of radiometric calibration would substantially simplify UAV-based small-scale agricultural monitoring and disease detection processes. However, in case images from different dates or large areas are used for classification, radiometric calibration or image normalization might be required to improve classification accuracy.

Disclosure statement
No potential conflict of interest was reported by the authors.