Analysis, Modeling and Multi-Spectral Sensing for the Predictive Management of Verticillium Wilt in Olive Groves Blekos Kostas Xouris Christos Katakis Sofoklis Tsakas Anastasios Gaia Robotics S.A. Makedonas Andreas Evdokidis Ioannis 25002 Patras, Greece Theoharatos Christos Alexandropoulos Dimitris IRIDA Labs S.A. Alexakos Christos 26504 Patras, Greece Lalos Aris {blekos,lalos}@isi.gr Industrial Systems Institute, Athena Research Center 26504 Patras, Greece ABSTRACT be, considering territorial characteristics such as low fertility and The intensification and expansion in cultivation of olives have con- sloping, difficult or impossible to exploit from other crops. tributed to the significant spread of verticillium wilt, which is the Verticillium wilt is the biggest fungal problem of olive cultivation. most important fungal problem affecting those trees. Recent studies It contributes to serious reduction in olive productivity, plant capital confirm that practices such as the use of innovative natural miner- destruction and soil degradation. Verticillium wilt causes a gradual als (Zeoshell ZF1) and the application of beneficial microorganisms malfunction and eventually a complete blockage of the vessels of (Micosat F BS WP), restore health in infected trees However, for the tree, in part or in whole, interrupting the movement of water their efficient implementation the above methodologies require the from the roots to the leaves, resulting in interruption of the water marking of trees in the early stages of infestation; a task that is supply in the affected part of the tree. This reduction in water supply impractical with traditional means (manual labor) but also very leads to nutritional deficiencies and even starvation of the leaves. difficult as early stages are difficult to perceive with the naked eye. Before the complete blockage and total necrosis of the affected In this paper we present the results of the MyOliveGroveCoach tissue associated with the part of the root that has been infected, project which uses multispectral imaging from unmanned aerial there precedes a stage of temporary water stress, a reversible stress, vehicles to develop an olive grove monitoring system that is able which is due to the closure of the mouths of the affected plant to a) collect large amount of data that is particularly important in tissue [4]. relation to the evolution of tree infestation b) quickly detect the In this stage of temporary stress, a deregulation is caused in problem, using innovative signal processing methods, multispectral the process of photosynthesis which results in a slight light-green imaging and computer vision, in combination with machine learn- discoloration of the leaves; a discoloration that is very subtle and ing techniques, providing accurate spatial identification of affected very difficult to detect with the naked eye. trees c) guide the farmer / agronomist when required, with a com- Thermal and multispectral surveying has shown high correla- munication and decision-making support system, with appropriate tions of leaf’s spectral characteristics to the degree of infestation, as interventions and providing maps of quantitative and qualitative measured in the 11 point scale of Table 1 [39]. On this basis, using characteristics of the grove. aerial imaging by unmanned aerial vehicles, we create the platform “My Olive Grove Coach” (MyOGC) KEYWORDS The main goal of MyOGC is the development of an intelligent system that will monitor olive groves and support farmers in the Precision Agriculture, Intelligent Management of Agriculture Pro- detection and treatment of Verticillium, using multispectral sen- duction, multi-spectral sensing, co-registration and fusioning of sors and spectrophotometers. With MyOGC it will be possible to a) multispectral and spectroscopy data in agriculture collect important data on the progress of tree infestation b) quickly detect the problem, using innovative signal processing methods, 1 INTRODUCTION multispectral imaging and computer vision, in combination with machine learning techniques, providing accurate spatial identifica- Olive cultivation in Greece is widespread. Olive groves occupy tion of affected trees c) guide the farmer / agronomist when required, an area of 7.16 million acres, numbering about 130 million olive with a communication and decision-making support system, with trees. This represents a large percentage of the total agricultural appropriate interventions and providing maps of quantitative and land and a very large percentage of agricultural land that would qualitative characteristics of the grove. AIT-CPS2020, September 02–04, 2020, Athens, Greece Copyright ©2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 0 Healthy tree main sources, non provide a clear advantage, rather they comple- 1 Tree looks healthy (slight crown discolloration) ment each other on the fronts of cost, resolving power, easiness 2 Chlorotic hair (yellow-bronze color) of use and other relevant metrics. MyOGC uses Unmanned Aerial slight twisting or curving of the extreme leaves Vehicles (UAVs), recognizing their low cost, ability for regular up- 3 Dry inflorescence - inflorescence - dehydration of twigs dates and resolving power as key advantages towards the goal of 4 Drying of a twig or a branch early infestation detection. 5 Dry arm - section or half of the tree 6 A main branch or arm of a tree retains vegetation 7 75 % of the tree has died 3 CONCEPTUAL ARCHITECTURE 8 A branch of the tree retains vegetation MyOGC integrated system provides an overall automation solution 9 A small section or a branch of the tree retains vegetation for detecting the Verticillium wilt from aerial multi-spectral images. 10 Drying of the whole tree The basic user requirements for the MyOGC platform is to support Table 1: Verticillium infection scale [39] different ways of data insertions, manually from the user or direct from the camera located to the drone. Thus, it combines Cloud and Edge Computing technologies, ensuring a highly efficient and scalable high-demanding data processing system and the execution 2 RELATED WORK of adapted AI prediction models in an embedded platform in user’s Remote sensing of agricultural crops that facilitate the timely pre- edge devices. The functional module of MyOGC platform is depicted diction of plant infestation by diseases has developed rapidly. Both in Figure 4. in Greece and abroad, companies have been set up to provide ser- MyOGC system consists of four main sub systems: a) The Core vices that monitor and support farmers and their fields. Figure 1 has coordination role, it provides the interfaces to the users and edge presents platforms available to farmers and producers that offer devices and it accepts and schedules data processes requests for remote monitoring of their fields and collection of agricultural data execution in the other subsystems; b) The Data Storage combines a by remote sensing. Figure 2 presents a comparison of some systems classical RDBMS and File System to store metadata, multi-spectral and data capturing sensors that are on the market, available as images and calculated results; c) Containers Execution Engine initi- commercial solutions for monitoring vegetation and crops. ates containers which execute a specific data processing task during Apart from commercial applications targeting farmers and other a data processing pipeline; and d) the Drone which hosts the Edge specialists of the field, there is significant research interest in using device, a Coral Edge TPU device from Google, deployed for execut- remote sensing data [9, 17, 19, 22, 35] in relation to automating and ing region of interest detection and classification tasks. facilitating all aspects of crops management, like disease monitor- In the Core subsystem, the Process Orchestrator is the module ing, predicting and preventing [4, 16], to crops yield monitoring that receives input data and requests for processing. Such requests and optimization [18, 20, 37] can be either the process of multi-spectral images of an olive field Specific applications include computer vision algorithms target- and the prediction of the spread of the disease on it, or use the ing productivity monitoring through tree counting / tree crown stored data in order to train the AI prediction models (both cloud delineation [9, 17, 22, 36] and health assessment through calcula- and embedded). According to the request, it selects the appropriate tions of vegetation indices [15, 24]. analysis workflow, it calculates the required resources and proceeds In the detection and delineation of individual tree crowns, deep to create the execution plan. The plan contains the data processes learning and machine learning approaches [6, 22, 36] also exhibit microservices that must be used and a workflow that defines the commendable results. A recent semi-supervised approach [36], em- execution order of the analysis tasks. The Process Orchestrator, ploying a convolutional neural network (CNN), combines LIDAR coordinates and monitors the analysis workflow initiating each and RGB data, yielding similar outcomes with classical unsuper- step and passing the intermediate results between the tasks. vised algorithms. CNNs were also used with multi-spectral imaging The two interfaces of the Core subsystem are a Graphical User data [6, 33]. In [6], a deep network was employed to differentiate Interface and a HTTP-based Application Programming Interface trees, bare soil and weeds. Li et al. [21] developed a CNN framework (API). The GUI is the point of interaction of the users with the to detect oil palm trees. Even though they provide accurate results, system. It is implemented using Python Django Framework and they need a large amount of training data. Angular Js library for the frontend. The user can define fields, upload There is also significant research going on the use of visible new multispectral images of a field, and ask for their processing and infrared spectroscopy for disease detection in plants in a fast, while the results are depicted in a GIS-based interactive map. The non-destructive, and cost-effective manner. The visible and infrared HTTP API is mainly used for the interoperability between the portions of the electromagnetic spectrum provide the maximum cloud platform and the edge device, the embedded installed in the information on the physiological stress levels in the plants, even drone. The HTTP API uses the GET, POST methods for allowing before the symptoms can be perceived from the human eye. Dif- the invocation of methods that support various tasks such as image ferent studies have been conducted for disease detection in plants uploading, new trained IA models downloading, image process using this technology [5, 7, 8]. execution, prediction upload, etc. Lastly, on sourcing data for remote sensing, there exist a variety The Data Storage, as mentioned before, is the centralized sub- of active, passive and mixed sources like GEOSATs, LIDARs and, system, responsible to securely store all the data of the MyOGC more recently, UAVs [11, 14, 15, 35, 37] (Figure 3). Of those three integrated system. A RDBMS is used, the proposed approach utilises Figure 1: Platforms available to farmers and producers for remote monitoring of fields. PostgresSQL, for storing users and fields metadata, pre-processed data, devices connection info, prediction results, etc. On the other hand, the filesystem is used to save binary files such as the input and processed images and the AI trained prediction models. The Containers Execution Environment takes the advantage of the virtual containers’ technology providing on demand data process functionalities in a cloud infrastructure. Each container is independent of computational resources and provides a specific data analysis task in the notion of the microservices architectural model [12]. There are four microservices in the MyOGC archi- tecture: a) the Tree Crown Detection, b) the Vegetation Indices Calculation, c) the AI Prediction for the Verticillium wilt disease presence and spread and d) the AI prediction model training. All these microservices are running independently and they are exe- cuting specific tasks which are invoked as services by the Process Figure 2: Comparison of indicative remote sensing systems Orchestrator. The Container Orchestrator’s main role is the in- available on the market. stantiation of the appropriate containers to be available for the execution of an analysis task. It executes a Credit-based algorithm [27] for scheduling the instantiation of the containers according to the amount of user’s requests and the available computational resources of the cloud infrastructure. This approach ensures both the scalability and the reuse of the cloud resources for serving the on-demand user’s requests in the most efficient manner. Finally, the Drone sub-system, aims to bring the intelligence provided by the AI prediction models near to the user’s main de- vice. In MyOGC, the drone with a multi-spectral camera, is used to capture the aerial image datasets. These datasets contain over- lapping images that can be merged to create a reflectance map, which is a mosaic of the area of interest where each of pixels in the Figure 3: Research by remote sensing source image represents the actual reflectance of the imaged object used for plant health analysis and the detection of the Verticillium wilt in olive trees. The classic procedure is to upload the images in the MyOGC platform for further processing and algorithmic analysis. The MyOGC system provides and additional feature. An embedded board with GPU capabilities is installed with the camera in the drone. A compact version of the AI prediction models is installed in the embedded, which is able to perform the data process anal- ysis on the spot. The results are sent to the MyOGC platform for presentation to the user. 4 MULTIMODAL PROCESSING APPROACHES Plant leaves contain information which is highly associated to their health. Optical leaf properties such as reflectance and transmittance are useful in remote sensing techniques for disease detection. They allow early detection, well before they can be perceived by the human eye, in a non-invasive manner. In assessing a plant’s health, the most basic and common metric used is the reflection of vegetation, i.e the ratio of the reflected Figure 4: MyOGC Overall Architecture. radiation to the incident radiation. An assumption is made that the reflection of vegetation at a certain electromagnetic wavelength, or spectral reflectivity, depends on the properties of the vegetation due to factors such as the type of each plant, its water content, its chlorophyll content and its morphology [10]. However, there may be a need to compare measurements that are more related to biophysical variables than to the spectral reflectivity itself. For these reasons, Vegetation Indices are often calculated. These indi- Based Image Analysis - GEOBIA) to achieve fast and accurate crown cators are obtained when two or more wavelength bands are used delineation (Figure 9). in an equation to calculate the corresponding vegetation index. In A second method is used for the same purpose but targeting a addition, vegetation indicators can help minimize problems related different platform, an embeddable device tuned for running ML to reflectivity data, such as changes in viewing angles, atmospheric applications. This device can be mounted on the UAV and connects distortions and shadows, especially as most vegetation indicators to the multispectral sensors, allowing real time processing of the are calculated as ratios of two or more wavelength bands [10, 31]. captured multispectral images. To make possible the on-the-fly Different vegetation markers use different wavelength zones and processing of incomplete and noisy data, we use a Convolutional provide information on different biophysical variables [38]. For ex- Neural Network (CNN), a class of NN that is ideal for tasks involving ample, one of the most commonly used indicators is the Normalized image segmentation and classification, trained on ground truth data Difference Vegetation Index (NDVI). NDVI uses the wavelength that where automatically generated from classically processed mul- corresponding to the red color band and is absorbed to a very large tispectral images. The main trade-offs between the simple computer extent by the chlorophyll in the foliage of the plants, and the wave- vision and the CNN methods are on implementation complexity, length band corresponding to the near-infrared (NIR) in which the accuracy and efficiency. On one hand, the CV approach is much chlorophyll shows the most intense reflection simpler to implement, shows high and consistent accuracy but is not efficient enough and therefore not a good choice for embed- NIR − RED NDVI = ded devices. The CNN approach, on the other hand, is significantly NIR + RED more complex and requires much more work to get it to satisfactory NDVI values range from -1 to 1 with the values closest to 1 results; furthermore, the accuracy of segmentation is not as con- corresponding to healthier and denser vegetation. NDVI can be sistent as in the CV case and the CNN may need some fine-tuning calculated using reflexivity or non-physical measurements for the and readjustment between runs or between fields. The deciding wave bands. advantage, though, of the CNN method is that it gives very good Another example where a vegetation index can provide biophys- results when deploying data from fewer or even one band, eliminat- ical information is the Green Exceedance Index (GEI), which is ing the preprocessing overhead, and making the method suitable calculated using the red, blue and green wavelength bands. Re- for low power and low memory embedded platforms, especially on search for GEI showed that the measured gross primary product in ML-tuned devices that further enhance the efficiency benefits of a deciduous forest was significantly correlated with GEI. Thus, a the method. specialized vegetation index can be used as a substitute for measur- able biophysical variables that are important when evaluating the phenology of a particular site or plant. Calculation of vegetation indices is usually done on a pixel-by- pixel basis and is, therefore, very sensitive to even slight image distortions. In order to calculate the real changes in biochemical and 5 MYOLIVGROVECOACH PLATFORMS physiological parameters of vegetation, collected multispectral data MyOGC system consists of two basic platforms: a) the cloud plat- have to be geometrically and radiometrically aligned, calibrated and form that contains the most of the MyOGC subsystems and b) the corrected, so as to ensure that the pixels in two images represent edge platform which is an embedded board (Coral’s Dev Board) ca- the same soil characteristics and the same soil point. Thus, a crucial pable to execute complex AI and image processing techniques. The part of MyOGC is the correct design and implementation of appro- role and interconnection between them is depicted in the Section 2 priate geometric transformations and spatial-temporal image filters of the current article. which include, characteristically, algorithms for image registra- Cloud platforms GUI is the main access point for the users to tion and alignment, image stitching, creation of orthomosaic with the MyOGC system. It provides the basic authorisation and au- photogrammetry techniques, spectral and luminosity corrections thentication mechanism and the forms for managing the fields and noise filtering. Classical computer vision techniques are, in related meta-data, such as location, photography sessions, owner, most cases, adequate for the implementation of the aforementioned prediction results. processes. Regarding the last, in order to demonstrate the condition of the Another class of processing algorithms relates to the removal fields to their respective farmers, the platform generated multi- of image noise due to data acquisition and enhancing the distinc- ple colored layers which are presented as overlays on the original tion between the objects under detection (i.e. tree crowns) and the map of the field. When the end-user decides to spectate a field, the background (i.e. shaded area). platform redirects to a specific interactive map screen where the pre- The next stage in processing of the multispectral data concerns processed orthomosaic with three basic colors (red, yellow, green) the extraction of useful macroscopic characteristics of the grove, in is presented. Green represents healthy trees without phytopatho- an individual tree basis. A key part of this process is the detection logical stress signs, yellow represents stress which is quantified of individual olive trees and the delineation of their crown. This is by reduced photosynthetic activity of the affected plant’s canopy achieved by using state of the art classical computer vision tech- and therefore possible onset of disease symptoms and finally red niques [9, 15]. More specifically, MyOGC employs a combination indicates sick trees and/or ground. The end-user can zoom-in and of pixel-based methods like Local Maximum Filtering and Water- out the map, in order to preview every single tree on the map, with shed Segmentation and object-based methods (Geographic Object high detail. 60 50 40 Reflectance (%) 30 20 10 0 400 500 600 700 800 900 1000 1100 Wavelength (nm) Figure 6: Typical Reflectance 6 DATA, TRIALS AND EVALUATION MyOGC uses two main sources of data: a) data from direct re- flectance measurements on leaves, collected from fields and used as samples for training the assessment- and prediction- algorithms, and b) data from aerial surveying with multispectral cameras. Olive leaves’ reflectance measurements are performed in cer- tain bands of the electromagnetic spectrum, mainly in the visible and near-infrared wavelengths. A typical reflectance spectrum of a healthy plant is shown in Figure 6. The reflectance of healthy leaves is usually low in the visible spectrum (400–700 nm) due to the significant absorbance from chlorophyll. Healthy plants have Figure 5: MyOGC GUI interactive map where (a) the ortho- high chlorophyll concentration since this substance is crucial for mosaic is depicted as overlay in original satellite field image photosynthesis, allowing plants to absorb light energy. Chlorophyll and (b) its zoom to the level where the trees are clearly de- reflects the green portion of the spectrum, producing the charac- picted . teristic green color of the leaves. Healthy leaves reflect strongly in the near-infrared spectrum, as absorbance of infrared light would cause overheat and consequently damage of the plant tissue. However, when a plant dies, the process of photosynthesis slows down, chlorophyll content is reduced, allowing other pigments For the map representation, Google’s Leaflet library was utilized to appear. These pigments reflect light on wavelengths which are with Google Map’s satellite image tiles. The overlay is a prepro- perceived as yellow or orange by the human eye. A diseased plant’s cessed orthomosaic that was constructed with open source pho- leaves absorb infrared light while they reflect the visible portion of togrammetry software (OpenSFM and GDAL libraries), ensuring the spectrum; the plant gradually dries up and eventually dies. It the maintenance of the spectral reflectance accuracy (reflectance has been observed that the effect of a disease on a plant changes map) and the exact geographical coordinates of the original multi- its leaf reflectance in a specific manner. Consequently, reflectance spectral images. Consequently, the image is rendered with a level of change of plant leaves is correlated to certain diseases. Remote sens- transparency, and the map is initialized based on the orthomosaic’s ing techniques combined with vis/near-infrared spectroscopy are coordinates. In this manner, only the farmers’ fields which can be capable of diagnosing diseases at an early stage without observable stretched with map zooms, are visualized (Figure 5). indications, by simply measuring the reflectance of a plant leaf. The edge platform used in MyOGC is the “Dev Board” by Coral. When light incidents on a plant leaf, two types of reflectance are It is a development board for prototyping on-device ML products. observed, specular and diffuse. Specular reflectance takes place in The device’s Edge-TPU is ideal for running embedded ML appli- the plant epidermis—air interface. Specular reflectance does not con- cations. In this project a dev-board is employed on the drone in tain useful information for the health of a plant as the reflected light order to assist and assess the data collection procedure in real time, does not penetrate the interior tissue of the leaf and therefore has bypassing the need for the cpu-intensive and time consuming step not interacted with biochemical constituents (such as chlorophyll, (uploading images to the server and processing), at least for pre- carotenoids etc.).In contrast, light collected by diffuse reflectance liminary data analysis. More specifically, algorithms are run on the has interacted with the mesophyll, the inner part of the leaf, where dev-board that delineate the olive trees and provide preliminary multiple processes of scattering and absorption of light by its bio- info for their health status. chemical constituents takes place. Therefore, the light from diffuse reflectance contains information about the biochemistry of the leaf: diffuse reflectance plays an important role in determining the health status of a plant, while specular reflectance acts as noise. The diffuse reflectance component of a leaf is usually measured using a spectrophotometer and an integrating sphere. The diffused reflectance component is diffused inside the integrating sphere, while the specular reflectance component exits to the outside of the integrating sphere. In the scope of MyOGC, leaf reflectance measurements are per- formed using Lambda 35 UV/Vis Spectrophotometer along with an integrating sphere and Spectralon as the reflectance standard. Leaf samples are collected from olive trees infected with verti- cillium wilt at different stages over a long period of time starting from March up to June at 15-day intervals. Five leaf samples are usually collected from randomly selected branches of each tree. Each olive leaf is mounted in a special sample holder provided by Figure 7: Typical flight path of a UAV while collecting data the spectrophotometer’s manufacturer. The sample holder is placed from a field. at the exit port of the integrating sphere. A light source of wave- length range of 190 nm to 1100 nm is at the entrance port of the integrating sphere. We collect leaf reflectance spectra from about 400 nm to 1100 nm, calculate the mean reflectance for each tree,and perform a first- and a second-order derivative analysis. Due to the high sensitivity of derivative analysis to noise, a Savitzky – Golay filter is applied for smoothing the data with a polynomial order of 4 and a frame length of 17. The first and second-order derivative analysis provides informa- tion for the reflectance slope in the red-edge position. The slope in the red-edge is highly associated with the chlorophyll content of (a) (b) the leaf. If the slope of a leaf reflectance spectrum is low, then the leaf has a low chlorophyll content. This means that leaf is infected with a disease or it slowly dies. Peaks in the second-order derivative are correlated to certain issues such as nitrogen deficiency. The aerial multispectral images are collected using a Pix4d Parrot Sequoia camera mounted on a C0 class drone. The Parrot camera is a multispectral camera capturing images on the four characteristic bands: green (550nm), red (660nm), red edge (735nm) and near- infrared (790nm). Figure 7 visualizes a typical drone flight pattern, (c) (d) at a height of 70m. A sample of the collected images is presented in Figure 8. An early processing stage takes place on the dev board mounted Figure 8: Sample of raw input images. One image per spec- on the drone, providing some real-time preliminary analysis of the tral band, taken at the same time using a multispectral cam- olive grove. Notably, this first analysis includes visualization of era. (a) GRE - Green 550nm (b) RED - Red 660nm (c) REG - olive trees crowns (Figure 9) and vegetation indices. Red Edge 735nm (d) NIR - Near Infrared 790nm intention to use them on real-world data, or even train generative 6.1 Synthetic Data Generation models that refine synthetic data for making them more suitable The effectiveness of deep learning algorithms significantly relies for training. In addition, synthetic data can be used to increase on the proper acquisition and manual annotation of a large amount real-world datasets, or even be generated from existing data using of good quality data. In many cases, limitations occur that have to generative models, in order to produce a hybrid dataset able to effec- do with the lack of expert knowledge for data labeling, difficulties tively cover the data distribution that is not adequately represented in capturing large quantities of data with sufficient variety, or even in the real dataset and, therefore, alleviate dataset bias. the ability to capture good quality data volumes might be extremely In this line of research and due to lack of high volumes of proper expensive and under privacy restrictions. In such cases, the lack of olive tree data in different environmental conditions, generation real-world data can be tackled by generating synthetic data that of synthetic data is investigated here with the use of Blender tool. share the same basic characteristics with the real ones. Blender is an open source software for creating 3D environments, The use of synthetic data can be twofold. For example, synthetic able to run on any operating system and having the ability to data can be initially used to train a deep learning model with the write scripts and addons in Python programming language. In our Figure 9: Automatic delineation of olive trees (overlay). Figure 11: Olive branch creation procedure, (a) leaves frontal and back view, (b) leaf texture extraction, (c) leaf 3d model, (d) branch image, (e) branch texture, (f) final branches. Figure 10: Olive trees synthetic data creation chain. case, scripting was used in the Blender environment for generating multiple olive trees with great variability. From a single leaf and the use of specific textures of the tree branches, trunks and the soil, a close-to-real synthetic tree as well as a number of synthetic trees were created, using the sequential approach shown in the block diagram of Figure 10. Initially, the appropriate textures needed for the olive tree cre- ation (healthy/ill leaves, branches, trunk), as well as the position of the soil of the trees were gathered (Figure 11 (a,b)). The 3D model of the leaf was then produced (Figure 11 (c)), followed by the creation of the branch by replicating the created leaf model, or combining multiple leaf models (Figure 11 (d-f)). Using the created branches and combining them with the olive tree trunk texture, an olive tree can be created. By replicating the Figure 12: Creation of multiple trees: (a) olive tree branches, same methodology a random number of trees can be positioned combined with the trunk texture, produces the tree (d) onto the given soil, as shown in Figure 12. placed onto the soil having the texture inherited by (b), with (b) being the final olive trees’ creation. 7 DISCUSSION AND CONCLUSIONS Monitoring vegetation using drones can provide important data for the assessment of the condition of crops. However, it is vital that data collection with today’s media be done as carefully as possible, sensors, flight planning, weather and sun conditions, as well as as it will be the basis for future studies of Precision Agriculture aspects of geographic positioning. and ecological monitoring. Despite the plug-and-play nature of the By using multispectral imaging from UAVs and employing in- latest generation of multispectral sensors, such as Parrot Sequoia novative signal processing methods in combination with machine and MicaSense RedEdge, a number of factors require careful consid- learning techniques, MyOGC offers an olive grove monitoring sys- eration if the goal is to collect high quality data that are comparable tem that is useful in the early detection and prediction of verticil- between sensors, geographically and over time. lium wilt spread, and provides a platform that helps the farmer MyOliveGroveCoach has developed and is implementing a stan- asses the condition of their fields through maps of important char- dard workflow for processing agricultural multispectral data, taking acteristics of the grove and guides the agronomist through a com- into account the technical aspects and challenges of multispectral munication and decision-making support system. ACKNOWLEDGMENTS [19] Aftab Khan, Umair Khan, Muhammad Waleed, Ashfaq Khan, Tariq Kamal, Safdar Nawaz Khan Marwat, Muazzam Maqsood, and Farhan Aadil. 2018. Remote MyOliveGroveCoach (MIS 5040498) is implemented under the Ac- Sensing: An Automated Methodology for Olive Tree Detection and Counting tion for the Strategic Development on the Research and Technolog- in Satellite Images. IEEE Access 6 (2018), 77816–77828. https://doi.org/10.1109/ ACCESS.2018.2884199 ical Sector, co-financed by national funds through the Operational [20] Ilya Levner and Vadim Bulitko. [n.d.]. Machine Learning for Adaptive Image Programme of Western Greece 2014-2020 and European Union Interpretation. ([n. d.]), 7. funds (European Regional Development Fund). [21] Weijia Li, Haohuan Fu, Le Yu, and Arthur Cracknell. 2017. Deep learning based oil palm tree detection and counting for high-resolution remote sensing images. Remote Sensing 9, 1 (2017), 22. [22] Weijia Li, Conghui He, Haohuan Fu, Juepeng Zheng, Runmin Dong, Maocai Xia, REFERENCES Le Yu, and Wayne Luk. 2019. A Real-Time Tree Crown Detection Approach for [1] Thomas W. Brakke. 1994. Specular and diffuse components of radiation scattered Large-Scale Remote Sensing Images on FPGAs. Remote Sensing 11, 9 (April 2019), by leaves. Agricultural and Forest Meteorology 71, 3 (Nov. 1994), 283–295. https: 1025. https://doi.org/10.3390/rs11091025 //doi.org/10.1016/0168-1923(94)90016-7 [23] Yingying Li, Yaoliang Chen, and Jingfeng Huang. 2019. An Approach to Improve [2] Cédric Bravo, Dimitrios Moshou, Jonathan West, Alastair McCartney, and Leaf Pigment Content Retrieval by Removing Specular Reflectance Through Herman Ramon. 2003. Early Disease Detection in Wheat Fields using Spec- Polarization Measurements. IEEE Transactions on Geoscience and Remote Sens- tral Reflectance. Biosystems Engineering 84, 2 (Feb. 2003), 137–145. https: ing 57, 4 (April 2019), 2173–2186. https://doi.org/10.1109/TGRS.2018.2871830 //doi.org/10.1016/S1537-5110(02)00269-6 Conference Name: IEEE Transactions on Geoscience and Remote Sensing. [3] Craig R. Brodersen and Thomas C. Vogelmann. 2007. Do epidermal lens [24] Shan Lu, Fan Lu, Wenqiang You, Zheyi Wang, Yu Liu, and Kenji Omasa. 2018. A cells facilitate the absorptance of diffuse light? American Journal of Botany robust vegetation index for remotely assessing chlorophyll content of dorsiventral 94, 7 (2007), 1061–1066. https://doi.org/10.3732/ajb.94.7.1061 _eprint: leaves across several species in different seasons. Plant Methods 14, 1 (Dec. 2018), https://bsapubs.onlinelibrary.wiley.com/doi/pdf/10.3732/ajb.94.7.1061. 15. https://doi.org/10.1186/s13007-018-0281-z [4] R Calderón, Juan Antonio Navas-Cortés, C Lucena, and Pablo J Zarco-Tejada. 2013. [25] Petr Lukeš, Pauline Stenberg, Miina Rautiainen, Matti Mõttus, and Kalle M. High-resolution airborne hyperspectral and thermal imagery for early detection Vanhatalo. 2013. Optical properties of leaves and needles for boreal tree of Verticillium wilt of olive using fluorescence, temperature and narrow-band species in Europe. Remote Sensing Letters 4, 7 (July 2013), 667–676. https: spectral indices. Remote Sensing of Environment 139 (2013), 231–245. //doi.org/10.1080/2150704X.2013.782112 Publisher: Taylor & Francis _eprint: [5] Don H. Card, David L. Peterson, Pamela A. Matson, and John D. Aber. 1988. https://doi.org/10.1080/2150704X.2013.782112. Prediction of leaf chemistry by the use of visible and near infrared reflectance [26] Matti Mõttus, Madis Sulev, and Lea Hallik. 2014. Seasonal Course of the Spectral spectroscopy. Remote Sensing of Environment 26, 2 (Nov. 1988), 123–147. https: Properties of Alder and Birch Leaves. IEEE Journal of Selected Topics in Applied //doi.org/10.1016/0034-4257(88)90092-2 Earth Observations and Remote Sensing 7, 6 (June 2014), 2496–2505. https://doi. [6] Ovidiu Csillik, John Cherbini, Robert Johnson, Andy Lyons, and Maggi Kelly. org/10.1109/JSTARS.2013.2294242 Conference Name: IEEE Journal of Selected 2018. Identification of citrus trees from unmanned aerial vehicle imagery using Topics in Applied Earth Observations and Remote Sensing. convolutional neural networks. Drones 2, 4 (2018), 39. [27] Andreas Pegkas, Christos Alexakos, and Spiros Likothanassis. 2018. Credit-based [7] Paul J Curran, Jennifer L Dungan, Bruce A Macler, Stephen E Plummer, and algorithm for Virtual Machines Scheduling. In 2018 Innovations in Intelligent David L Peterson. 1992. Reflectance spectroscopy of fresh whole leaves for the Systems and Applications (INISTA). IEEE, 1–6. estimation of chemical concentration. Remote Sensing of Environment 39, 2 (Feb. [28] Josep Peñuelas and Iolanda Filella. 1998. Visible and near-infrared reflectance 1992), 153–166. https://doi.org/10.1016/0034-4257(92)90133-5 techniques for diagnosing plant physiological status. Trends in Plant Science 3, 4 [8] D. Wu, L. Feng, C. Zhang, and Y. He. 2008. Early Detection of Botrytis cinerea on (April 1998), 151–156. https://doi.org/10.1016/S1360-1385(98)01213-8 Eggplant Leaves Based on Visible and Near-Infrared Spectroscopy. Transactions [29] V. P. Polischuk, T. M. Shadchina, T. I. Kompanetz, I. G. Budzanivskaya, A. L. of the ASABE 51, 3 (2008), 1133–1139. https://doi.org/10.13031/2013.24504 Boyko, and A. A. Sozinov. 1997. Changes in reflectance spectrum char- [9] Michele Dalponte, Lorenzo Frizzera, and Damiano Gianelle. 2019. Individual tree acteristic of nicotiana debneyi plant under the influence of viral infection. crown delineation and tree species classification with hyperspectral and LiDAR Archives of Phytopathology and Plant Protection 31, 1 (July 1997), 115–119. data. PeerJ 6 (Jan. 2019), e6227. https://doi.org/10.7717/peerj.6227 https://doi.org/10.1080/03235409709383221 Publisher: Taylor & Francis _eprint: [10] Donald Wayne Deering. 1978. Rangeland reflectance characteristics measured by https://doi.org/10.1080/03235409709383221. aircraft and spacecraft sensors. Ph. D Dissertation. Texas A&M Universtiy (1978). [30] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. 2019. Fast-SCNN: [11] Anderson Aparecido dos Santos, José Marcato Junior, Márcio Santos Araújo, Fast Semantic Segmentation Network. arXiv:1902.04502 [cs] (Feb. 2019). http: David Robledo Di Martini, Everton Castelão Tetila, Henrique Lopes Siqueira, //arxiv.org/abs/1902.04502 arXiv: 1902.04502. Camila Aoki, Anette Eltner, Edson Takashi Matsubara, Hemerson Pistori, [31] William E Reifsnyder. 1981. Analytical Methods in Biophysical Ecology. Raul Queiroz Feitosa, Veraldo Liesenberg, and Wesley Nunes Gonçalves. 2019. [32] Sindhuja Sankaran, Ashish Mishra, Reza Ehsani, and Cristina Davis. 2010. A Assessment of CNN-Based Methods for Individual Tree Detection on Images review of advanced techniques for detecting plant diseases. Computers and Captured by RGB Cameras Attached to UAVs. Sensors (Basel, Switzerland) 19, 16 Electronics in Agriculture 72, 1 (June 2010), 1–13. https://doi.org/10.1016/j.compag. (Aug. 2019). https://doi.org/10.3390/s19163595 2010.02.007 [12] Lang Feng, Prabhakar Kudva, Dilma Da Silva, and Jiang Hu. 2018. Exploring [33] Anderson Aparecido dos Santos, José Marcato Junior, Márcio Santos Araújo, serverless computing for neural network training. In 2018 IEEE 11th International David Robledo Di Martini, Everton Castelão Tetila, Henrique Lopes Siqueira, Conference on Cloud Computing (CLOUD). IEEE, 334–341. Camila Aoki, Anette Eltner, Edson Takashi Matsubara, Hemerson Pistori, et al. [13] David M. Gates, Harry J. Keegan, John C. Schleter, and Victor R. Weidner. 1965. 2019. Assessment of CNN-Based Methods for Individual Tree Detection on Spectral Properties of Plants. Applied Optics 4, 1 (Jan. 1965), 11–20. https: Images Captured by RGB Cameras Attached to UAVs. Sensors 19, 16 (2019), 3595. //doi.org/10.1364/AO.4.000011 Publisher: Optical Society of America. [34] Serge Savary, Andrea Ficke, Jean-Noël Aubertot, and Clayton Hollier. 2012. Crop [14] Emilio Guirado, Siham Tabik, Domingo Alcaraz-Segura, Javier Cabello, and Fran- losses due to diseases and their implications for global food production losses cisco Herrera. 2017. Deep-learning Versus OBIA for Scattered Shrub Detection and food security. Food Security 4, 4 (Dec. 2012), 519–537. https://doi.org/10. with Google Earth Imagery: Ziziphus lotus as Case Study. Remote Sensing 9, 12 1007/s12571-012-0200-5 (Nov. 2017), 1220. https://doi.org/10.3390/rs9121220 [35] Jorge Torres-Sánchez, Francisca López-Granados, Irene Borra-Serrano, and [15] David Gómez-Candón, Sylvain Labbé, Nicolas Virlet, and Audrey Jolivot. [n.d.]. José Manuel Peña. 2018. Assessing UAV-collected image overlap influence on HIGH RESOLUTION THERMAL AND MULTISPECTRAL UAV IMAGERY FOR computation time and digital surface model accuracy in olive orchards. Precision PRECISION ASSESSMENT OF APPLE TREE RESPONSE TO WATER STRESS. Agriculture 19, 1 (Feb. 2018), 115–133. https://doi.org/10.1007/s11119-017-9502-0 ([n. d.]), 10. [36] Ben. G. Weinstein, Sergio Marconi, Stephanie Bohlman, Alina Zare, and Ethan [16] George Iatrou, Spiros Mourelatos, Zois Zartaloudis, Miltiadis Iatrou, Sandra White. 2019. Individual tree-crown detection in RGB imagery using semi- Gewehr, and Stella Kalaitzopoulou. [n.d.]. REMOTE SENSING FOR THE MAN- supervised deep learning neural networks. bioRxiv (May 2019). https://doi. AGEMENT OF VERTICILLIUM WILT OF OLIVE. Fresenius Environmental Bulletin org/10.1101/532952 25, 4 ([n. d.]), 7. [37] Xiangqian Wu, Xin Shen, Lin Cao, Guibin Wang, and Fuliang Cao. 2019. Assess- [17] K G Karantzalos and D P Argialas. [n.d.]. TOWARDS AUTOMATIC OLIVE TREE ment of Individual Tree Detection and Canopy Cover Estimation using Unmanned EXTRACTION FROM SATELLITE IMAGERY. ([n. d.]), 5. Aerial Vehicle based Light Detection and Ranging (UAV-LiDAR) Data in Planted [18] Christos Karydas, Sandra Gewehr, Miltiadis Iatrou, George Iatrou, and Spiros Forests. Remote Sensing 11, 8 (April 2019), 908. https://doi.org/10.3390/rs11080908 Mourelatos. 2017. Olive Plantation Mapping on a Sub-Tree Scale with Object- [38] Jinru Xue and Baofeng Su. 2017. Significant remote sensing vegetation indices: Based Image Analysis of Multispectral UAV Data; Operational Potential in Tree A review of developments and applications. Journal of Sensors 2017 (2017). Stress Monitoring. Journal of Imaging 3, 4 (Dec. 2017), 57. https://doi.org/10. [39] Z Zartaloudis, M Iatrou, G Savvidis, K Savvidis, D Glavenas, S Theodoridou, 3390/jimaging3040057 K. Kalogeropoulos, and S Kyparissi. 2014. A new 11-point calibration scale of Verticillium wilt of olive verified by thermal remote sensing and plant analysis. hyperspectral measurements. Computers and Electronics in Agriculture 85 (July In 17th PanHellenic Plant Pathology Conference. Volos. 2012), 13–23. https://doi.org/10.1016/j.compag.2012.03.006 [40] Jing-Cheng Zhang, Rui-liang Pu, Ji-hua Wang, Wen-jiang Huang, Lin Yuan, and Ju-hua Luo. 2012. Detecting powdery mildew of winter wheat using leaf level