This is the accepted version of the article: González, Alejandro; Sevillano, Xavier; Betegón Putze, Isabel; [et al.]. MyROOT 2.0 : An automatic tool for high throughput and accurate primary root length measurement. DOI 10.1016/j.compag.2019.105125 This version is avaible at https://ddd.uab.cat/record/218482 under the terms of the license MyROOT 2.0: An automatic tool for high throughput and accurate primary root length measurement Alejandro Gonzáleza,∗, Xavier Sevillanoa , Isabel Betegón-Putzeb , David Blasco-Escámezb , Marc Ferrera , Ana I. Caño-Delgadob a GTM - Grup de recerca en Tecnologies Mdia. La Salle, Universitat Ramon Llull. Quatre Camins, 30. 08022. Barcelona, Spain b Centre for Research in Agricultural Genomics (CRAG). Edifici CRAG - Campus UAB. 08193 Cerdanyola, Barcelona, Spain Abstract The automatic measurement of external physical traits (i.e. phenotyping) of plant or- gans, such as root length –which is highly correlated with plant viability– is one of the current bottlenecks in academic and agricultural research. Although many root length measurement software tools are available to the community, plant scientists often find their usability is limited, the measurements they provide are not accurate enough or they are too limited to specific image characteristics. In response to that, this work describes MyROOT 2.0, an automatic software tool jointly developed by plant scien- tists and computer vision engineers to create a high throughput root length measure- ment tool that reduces user intervention to a minimum. Using Arabidopsis thaliana seedlings grown on agar plates as a case study, MyROOT 2.0 is capable of detecting the root regions of interest in a fully automatic manner with an accuracy of 98%. Fur- thermore, this work also presents previously unreported experiments to evaluate several constituting modules of MyROOT 2.0, such as the ability to determine image scale au- tomatically with subpixel accuracy, or the influence of training the hypocotyl detector using wildtype or mutant samples. Finally, when compared to state-of-the-art root length measurement software tools, MyROOT 2.0 achieves the highest root detection rate, obtaining measurements which are four times more accurate than its competitors. This makes MyROOT 2.0 an attractive tool for high throughput root phenotyping. ∗ Correspondingauthor Email address: a.gonzalez@salle.url.edu (Alejandro González) Preprint submitted to Journal of LATEX Templates November 14, 2019 Keywords: Plant phenotyping; Root length measurement; Imaging tools; Plant science; High throughput Root phenotyping. 1 1. Introduction 2 In the quest to create genetically improved crop varieties which can better adapt to 3 future climate conditions and agricultural management techniques, plant scientists have 4 leveraged increasingly cheaper high throughput sequencing technologies to make huge 5 amounts of genomic data of different plant species available to the scientific community 6 [1, 2]. 7 However, to evaluate genetic improvements it is necessary to analyze how the in- 8 teraction between the genotype and the environment modifies the observable traits of 9 the plant, a task that goes by the name of phenotyping. 10 When compared to the boost in genomic sequencing, the relatively slower develop- 11 ment of phenotyping tools may hinder the advances in plant breeding and fundamental 12 plant science [2]. In this sense, plant scientists need high throughput tools that allow the 13 automatic analysis of large volumes of phenotypic traits of plants grown in laboratory 14 conditions, in order to later transfer their conclusions to agriculture. 15 Among all plant organs, the root is crucial for plant functioning, as some of its ob- 16 servable features reveal critical aspects related to plant viability [2]. As pointed out by 17 Wasson et al., the analysis of root system architecture (which encompasses morpho- 18 logical traits such as root length, root density, root branching or total root surface) is 19 key to obtain higher crop yields [3]. This makes the root one of the best options when 20 it comes to plant phenotyping. 21 For this reason, the development of automatic tools for high throughput root phe- 22 notyping constitutes one of the most significant areas of research and innovation in this 23 field. However, many recent works in this area prove that the accurate measurement 24 of significant root traits (like root length) is still carried out manually to a large extent 25 (e.g. [4, 5, 6]), despite the fact that several root length measurement software tools are 26 available to the plant science community. 27 This can be due to three main hurdles that plant scientists face when using available 2 28 root length measurement software tools: i) the ease of use (often related to the usability 29 of the tools and the degree of manual intervention required), ii) the constraints imposed 30 on the image acquisition process, and iii) the accuracy of the root length measurements. 31 Considering those, we recently introduced MyROOT [7], developed as a joint effort 32 between plant scientists and computer vision engineers to obtain an accurate and usable 33 software tool to measure primary root length. 34 In that work, we tested MyROOT by measuring the primary root length of Ara- 35 bidopsis thaliana (Arabidopsis) seedlings grown in agar plates (see Figure 1), given its 36 use as a model organism in plant science with significant implications in agricultural 37 research. 38 We highlighted how its main features and usability could benefit plant scientists in 39 need of fast and accurate primary root length measurements, providing experimental 40 evidences of its accuracy by comparison against manual measurements made by ex- 41 perts. Moreover, we also analyzed how the use of MyROOT reduces labour time by 42 half when compared to manual measurements using ImageJ. 43 In our quest to make our tool usable and appealing to the widest possible audi- 44 ence among plant scientists, we identified two features that could limit their interest in 45 MyROOT. 46 First, the need to place a piece of measuring tape on the plate lid is not a widely 47 used technique among plant scientists. Indeed, many image capture systems (e.g. fixed 48 camera setups, or flatbed scanners) allow plant scientists to know the exact equivalence 49 between pixels and centimeters. For this reason, we facilitated the introduction of this 50 data manually through MyROOT 2.0 graphical user interface. 51 And secondly, many plant scientists are interested in conducting measurements of 52 many roots to validate their scientific hypotheses, which requires processing batches 53 of images. These measurements should be processed automatically, and as quickly 54 and efficiently as possible, saving time and effort. The previous version of MyROOT 55 allowed batch image processing, but still required manual intervention from the user 56 to define the region of the image where roots are located. For this reason, we have 57 developed an automatic and efficient algorithm to detect root regions in the image, 58 reducing manual intervention and making high throughput root length measurement 3 (a) (b) Figure 1: Two images of Arabidopsis seedlings grown on an agar plate. Note that seeds were planted in two rows. Notice the differences in illumination conditions and scale. 59 feasible with MyROOT 2.0. 60 It is important to highlight that these improvements have not made the use of the 61 tool more complex. Moreover, the accuracy of the root length measurements remains 62 unchanged from MyROOT to MyROOT 2label Root:fig:placa .0, as the label added Root:fig:placa features do not affect t he 63 measurement procedures. 64 To provide the reader with a comprehensive view of MyROOT 2.0 capabilities, 65 we present a series of previously unreported experiments dedicated to the evaluation of 66 each of the image processing algorithms that constitute the main modules of the system. 67 Thanks to these experiments, we are able to provide a global view of the performance 68 of the tool. 69 In particular, we evaluate i) the precision of automatic image scale estimation, ii) 70 the accuracy of root ROI detection, iii) the influence of wild-type and mutant models 71 on the hypocotyl detection algorithm accuracy, and iv) the accuracy of root length 72 measurements, comparing them against other state-of-the-art root length measurement 73 software tools on images provided by the authors of those. 74 The obtained results reveal that MyROOT 2.0 determines image scale with sub- 75 pixel accuracy, and that it is capable of detecting nearly 98% of the regions of interest 76 of the image where roots are located. Moreover, MyROOT 2.0 yields average absolute 77 differences with respect to manual measurements four times smaller than its competi- 4 78 tors, which makes it a very appealing alternative for plant scientists requiring accurate 79 automatic root length measurements. 80 The rest of the paper is organized as follows: Section 2 reviews related work in 81 the literature. Then, the architecture and constituting modules of MyROOT 2.0 are 82 thoroughly described in Section 3. Section 4 presents the experimental evaluation, and 83 finally Section 5 discusses the conclusions and proposes lines for further work. 84 2. Related work 85 The literature presents several approaches to the automatic analysis of root system 86 architecture, mostly based on acquiring, processing and obtaining quantitative data 87 from root images. These approaches differ on several aspects, such as i) the medium 88 in which the plant is grown (e.g. soil or artificial growth media like agar), ii) the use of 89 2D or 3D imaging, iii) the imaging modality (e.g. optical, X-ray, magnetic resonance, 90 ground penetrating radar) and iv) the degree of manual intervention required from the 91 user. 92 In this section, we provide an overview of some of the most recent contributions in 93 the area of (semi-)automated optical imaging tools developed for 2D root phenotyping 94 in artificial media, which is the specific area of our work. The reader interested in a 95 broader view of this topic is referred to the comprehensive survey on root phenotyping 96 platforms by Kuijken et al. [2]. 97 The WinRHizo system is regarded as one of the pioneering efforts in the field of 98 root phenotyping [8]. Images captured with a scanner (which simplifies subsequent 99 image processing due to the constant illumination conditions) were processed to obtain 100 global root measurements (such as average root diameter, total root length or area), and 101 also link a topology analyses of the root system architecture. 102 More recently, the GROWSCREEN-Root software [9, 10] became a relevant phe- 103 notyping tool for roots grown in Petri dishes with (semi-)transparent nutrient gel. Using 104 infra-red imaging, their image processing algorithm included nonlinear filtering and 105 image smoothing to compensate varying background illumination. Root lines were de- 106 tected by means of local features based on steerable filters, and subsequently tracked 5 107 from top to tip using a root following algorithm that took into account a correcting fac- 108 tor based on root orientation. The detection of crossing and branches was based on the 109 detection of T -shaped junctions and the recursive execution of the root tracking pro- 110 cedure described earlier. Finally, the authors incorporated plausibility checks and/or 111 heuristics to include temporal information to perform phenotyping of the same root 112 along time. Based on this data, the authors extracted quantitative information of the 113 root system architecture such as root thickness, length, orientation angle, and branch 114 or node order. 115 The EZ-Rhizo software is a semiautomated tool based for measuring the length, 116 angle and number of lateral roots, as well as the position, length and angle of lateral 117 roots of any order [11]. It is a heavily user driven system, as mainly consists of a user- 118 controlled threshold-based binarization process followed by a series of noise removal 119 operations that end up in the detection, skeletonization and manual retouching of the 120 roots, upon which measurements are automatically made. 121 Focusing on tracing Arabidopsis seedling roots grown on agar plates, RootTrace 122 [12] performs root tracing from a user-defined start location to the tip of the root. To 123 that end, it applies the Condensation particle filter using a simple weighted color model 124 [13]. Following the root trace and updating the appearance model of the root as it 125 goes, it applies a hysteresis thresholding to detect the root tip, which makes it robust 126 against slight changes in root appearance and variations in background appearance 127 across images. After detecting the center line of the root, RootTrace can measure the 128 curvature of the root by calculating local angles at each point along the root trace. 129 Recently, RootTrace capabilities were expanded to count lateral roots and the tracking 130 model extended to recover strongly curved and agravitropic roots [14] and put to the 131 test in [15]. 132 In [16], Yazdanbakhsh and Fisahn introduced PlaRoM, a hardware platform and 133 imaging system to monitor the growth of Arabidopsis roots grown in agar plates, which 134 are scanned using a camera-microscope unit mounted on a robotic arm. Their image 135 processing algorithm allows root tip detection [17], as well as measuring growth ve- 136 locity profiles on time lapse records. 137 LeBot et al. developed DART [18], a software tool that relied heavily on manual 6 138 labour from the user to set root markers that define root system architecture, so it should 139 be regarded as a technical aid to visualize, organize and store the information contained 140 in root images. 141 The work by Iyer-Pascuzzi et al. presented an imaging system based on gel-based 142 growth cylinders, which were photographed at different rotation angles [19]. The im- 143 age processing mainly consisted of adaptive thresholding segmentation to obtain a bi- 144 nary mask of the root, from which several quantitative root traits were computed (e.g. 145 perimeter, convex area, average root radius, total length, etc.). These parameters were 146 compacted into a feature vector that was used to train a Support Vector Machine clas- 147 sifier to perform classification into two rice genotypes. 148 SmartRoot is a software toolbox developed by Lobet et al. that allowed semiau- 149 tomated image analysis for the quantitative analysis of root growth and architecture 150 of complex root systems [20]. Using a vectorial representation of roots, SmartRoot 151 implements a root tracing algorithm that requires the user to click anywhere along the 152 root. The root system is treated as a collection of roots (possibly connected) that are 153 individually represented as sets of connected segments. 154 GiA Roots is a semiautomated image analysis based tool that allows the calcula- 155 tion of up to 19 root system architecture traits [21]. To that end, the image is first 156 subjected to segmentation (either via global, adaptive or double adaptive thresholding) 157 to separate the root from the background. Next, the resulting binary image is optionally 158 skeletonized, and root traits are automatically computed upon it. 159 The RootReader2D software [22] operated on images captured by a standard cam- 160 era using a imaging setup with two cross-polarized filters to enhance the contrast of 161 the root systems from the background, which simplified subsequent image analysis. 162 The characteristics of the captured images made it possible to use plain thresholding to 163 segment the root, which was then skeletonized. A connectivity analysis of the pixels of 164 the root skeleton facilitated the detection of the root endpoints and divide the root into 165 segments, transforming the root into a graph suitable for finding the shortest connected 166 path between endpoints, using individual segment lengths as the graph weights. 167 The work by Benoit et al. [23] constitutes an interesting contribution to one of the 168 recurrent problems encountered when conducting root phenotyping: multiple cross- 7 169 ing seedlings. The authors proposed using a variant of the Perona-Malik anisotropic 170 diffusion equation to solve the segmentation at the crossings locations along the main 171 orientations of the objects. 172 In [24], the RootNav software was presented as a tool for the semiautomated mea- 173 surement of two-dimensional root architectures. Using a top-down approach to detect 174 roots based on EM clustering and A* search, the authors fitted a model path from root 175 apices to the seed point, thus identifying root branches. Furthermore, the user is given 176 the chance to manually modify the automatically traced paths. 177 A different approach to root phenotyping was the landmark-based root system ar- 178 chitecture analysis presented by Ristova et al. in [25]. Using a model of 20 landmarks 179 placed manually on recognizable developmental landmarks on the root, the authors 180 used Principal Component Analysis to distinguish between plants under different hor- 181 monal treatment conditions based on root allometry variations. 182 Slovak et al. introduced the BRAT software for Arabidopsis root phenotyping [26]. 183 Operating on images captured through a cluster of flatbed scanners, BRAT uses SIFT 184 features to perform alignment between the images of the same plate taken at different 185 time instants. To detect plants, the algorithm first performs shoot detection based on 186 color segmentation and then uses the HSV color model to detect the roots, applying a 187 median filter for noise reduction. To detect the pixels belonging to the root, a twofold 188 criterion that combines Sobel edge detection and background illumination variations 189 removal is employed. Finally, roots are skeletonized, which allows computing several 190 traits such as root length, tortuosity, growth rate, etc. BRAT can operate on a more 191 automated or a semi-automated mode, and it needs guidance as regards the placement 192 of the seeds. 193 Focusing on the specific problem of mapping nodulation patterns in legume roots, 194 Remmler et al. designed an semi-automated algorithm to trace the primary and lateral 195 roots of peas, although it requires intensive intervention by the user, clicking on the 196 start and end points of the root [27]. 197 The work by Kumar et al. presented an imaging system combining image pre- 198 processing for feature enhancement, feature extraction, and supervised statistical learn- 199 ing to perform root tip detection [28]. The proposed system first segmented the root 8 200 from the background, and then applied a thinning transformation on the root. This was 201 followed by corner detection for local feature extraction based on Zernike moments, 202 which were fed into a Gaussian mixture model for training a statistical classifier. This 203 type of classifier performed a hierarchical classification of the detected corners into 204 root tip (distinguishing between primary and lateral root tips) and non-root tips. 205 RootGraph was a root analysis software structured as a four-step pipeline compris- 206 ing segmentation, distance transform, skeletonization and graph computation [29]. At 207 its latter stage, RootGraph uses graph optimization algorithms to automatically sepa- 208 rate lateral and primary roots, estimating several phenotypic root traits such as length 209 and diameter. 210 More recently, the PlantRoot imaging system was presented in the context of the 211 RhizoChamber-Monitor robotic platform for root growth patterns grown in rhizoboxes 212 [30]. That software allowed to analyze the spatiotemporal dynamics of root growth 213 from time-course images of multiple plants, quantifying growth traits of primary and 214 lateral roots (length and diameter) in a dynamic manner. 215 What are the main advantages of MyROOT 2.0. when compared to other existing 216 root length measurement software tools? First, MyROOT 2.0 operates on photographs 217 of agar plates imaged from a top perspective with a standard digital camera or even 218 a good quality cell phone, with no need for special nor expensive imaging equipment 219 (such as camera-microscope units mounted on a robotic arm [16] or an array of flatbed 220 scanners1 [26]). Second, MyROOT 2.0 requires no manual effort from its users to 221 complete root length measurements, with no need for pinpointing individual features 222 of each and every one of the roots under analysis (e.g. as in [12, 27]). Third, MyROOT 223 2.0 is capable of accurately estimating image scale from the analysis of a piece of 224 measuring tape stuck on the plate lid. Furthermore, MyROOT 2.0 ensures accurate 225 root length measurement by detecting the root start and end points, while other systems 226 merely detect the plant shoot to define the root start point, which is incorrect from a 1 It is fair to mention that these software tools are also able to process images taken by other means. However, their image analysis algorithms are highly adapted to the extremely stable visual characteristics of flatbed scanner images, offering little adaptability to process different types of images satisfactorily. 9 Measuring tape detection and analysis pixels/mm Measuring tape bbox Measuring tape factor detection analysis (if image scale is known) Root segmentation Root length measurement Root ROI Root mask Root Hypocotyl Root detection generation tracking detection measurement (a) MyROOT 2.0 pipeline block diagram Root segmentation Root length measurement Measuring tape detection and analysis (b) MyROOT 2.0 key processing regions Figure 2: Block diagram and key processing regions of MyROOT 2.0. 227 biological point of view [11, 26]. 228 3. Description of MyROOT 2.0 229 MyROOT 2.0 is designed as a three-stage image processing pipeline comprising i) 230 the detection and analysis of the measuring tape to infer image scale, and the roots ii) 231 segmentation and iii) measurement. Figure 2(a) depicts the schematic block diagram 232 of MyROOT 2.0 pipeline, while Figure 2(b) represents the image regions in which the 233 processing of each pipeline stage takes place. 234 The following paragraphs summarize the operations performed on the input image. 235 First, MyROOT 2.0 automatically detects the measuring tape regardless of its position 236 on the plate. Next, it analyzes the region corresponding to the tape, detecting the 10 Measuring tape detection 0° bbox Radon Median Crop transform filtering 90° Hough Bottom-hat Otsu transform filtering segmentation q Rotation Radon Local maxima pixels/mm -q transform detection factor Measuring tape analysis Figure 3: Block diagram of the measuring tape detection and analysis module. 237 millimeter lines and performing rotation corrections if necessary, extracting the pixel- 238 to-millimeters equivalence corresponding to the image. 239 Next, the analysis focuses on the roots. MyROOT 2.0 first detects the seedlings 240 rows in the plate, finding the root region of interest (root ROI) that contains each row 241 of roots to be measured. Then, the roots enclosed in these ROI are segmented from 242 the background. Subsequently, each root is tracked upwards from its tip to its starting 243 point, the position of which is defined by the detection of each root’s hypocotyl. As a 244 result, roots can be traced and their length measured. 245 Finally, the measurement results are output as an image indicating the length in mil- 246 limeters of each root. Furthermore, these measurements are also transferred to RSML, 247 Excel and text files to facilitate subsequent analysis. 248 The following sections present a thorough description of the methods and algo- 249 rithms that implement the processes just outlined. 250 3.1. Measuring tape detection and analysis 251 The first stage of the pipeline aims at automatically detecting the measuring tape 252 placed on the plate lid for determining the conversion factor between pixels and mil- 253 limeters for the given input image, which is a key issue for reporting accurate root 254 measurements. Figure 3 shows the schematic diagram of this module. 255 Taking into account the luminance contrast between the measuring tape and the 11 256 background, we search for a white patch in the image. To do so, we compute the Radon 257 transform of the image at 0 and 90 degrees to obtain its 1D projections onto the vertical 258 and horizontal axes. However, the edges of the plate often show high luminance in 259 the image, which may lead to spurious peaks in these projections. For this reason, we 260 apply median filtering on them to soften such peaks and thus, avoid confusing these 261 regions with the measuring tape. 262 Next, finding the widest regions with maximum value in both projections allows 263 finding the bounding box of the measuring tape. The inner side of this bounding box is 264 the region of interest which is processed to ultimately obtain the pixels-to-millimeters 265 equivalence. 266 To that end, we search for the centimeters marks on the tape by first segmenting 267 the region of interest using Otsu’s method. Then, we apply the morphological bottom- 268 hat operator to remove the millimeter marks. To compensate for slight slants occurred 269 at the time of sticking the measuring tape on the agar plate lid, we apply the Hough 270 transform for line detection to determine the predominant orientation of the centimeter 271 marks, and subsequently rotate the image by the negative of that angle to ensure the 272 vertical orientation of the centimeter marks. Next, we compute the 1D Radon trans- 273 form at 90 degrees to obtain a horizontal projection of the marks on the measuring 274 tape. On this projection, we apply a local maxima finding algorithm to find the me- 275 dian separation between consecutive centimeter marks, which is used to compute the 276 pixels-to-millimeters conversion factor. 277 3.2. Root segmentation 278 The root segmentation module aims at separating pixels that belong to the roots 279 from pixels in the background. The processes followed to achieve this goal are repre- 280 sented in the block diagram depicted in Figure 4. Root segmentation Root ROI ROI Color Ridge Binarization + roots mask normalization mask detection detection blob removal Figure 4: Block diagram of the root segmentation module. 12 20 15 10 5 0 0 200 400 600 800 1000 1200 2020 1515 1010 200 200 400 400 600 600 800 800 1000 1000 1200 1200 20 15 2020 10 1515 1010 5 0 0 200 200 200 400 400 400 600600 600 800 800 800 1000 1000 1000 1200 1200 1200 (a) Leaves detection (b) Longest roots detection (c) Detected root ROI Figure 5: Root ROI detection steps. 281 3.2.1. Root ROI detection 282 The root segmentation process starts by detecting the seedling rows in the plate. 283 The roots of each seedling row are considered to belong to a root region of interest 284 (ROI). Thus, the detection of root ROI consists of two main steps: i) detection of 285 seedling rows, and ii) determination of each root ROI bounding box. 286 To detect seedling rows, the inner plate region is first automatically cropped by 287 detecting plate borders using adaptive thresholding. Then, the resulting image is seg- 288 mented in the RGB space to detect the green leaves of the seedlings, considering as 289 leaf pixels all those in which the green channel takes the highest value (see Figure 290 5(a)). After applying morphological opening to remove noisy detections, we compute 291 the horizontal projection of the binary image corresponding to the leaves. On this pro- 292 jection, we detect local maxima to determine the number of seedling rows. 293 Moreover, from this projection we define the upper edge of the bounding box (or 294 boxes) enclosing the root ROI in the plate. Also, the vertical projection of the same 295 binary image enables us to define the right and left edges of the root ROI bounding 296 boxes. 297 Finally, using the aforementioned edges to obtain subimages corresponding to each 13 298 root ROI, we apply Otsu’s method to perform a preliminary detection of the roots. The 299 goal of this step is to find the longest roots in the ROI to define the bottom edge of 300 the bounding box. This is accomplished by applying a triple noise reduction process: 301 first, morphological opening is applied to remove noisy isolated pixels corresponding 302 to plate lid scratches or droplets. Then, connected components with non-vertical orien- 303 tation are also removed. Then, the horizontal projection of the resulting binary image is 304 subjected to a uniform quantization process to eliminate noisy detections. Finally, the 305 first zero after the largest peak of the quantized projection is used to define the vertical 306 coordinate of the bottom edge of the bounding box (see Figure 5(b)). 307 As a result, MyROOT 2.0 presents the input image with the detected bounding 308 boxes superimposed (see Figure 5(c)). The vertices of the bounding boxes are drag- 309 gable, so the user can modify the detected root ROI. Moreover, it is also possible to 310 add a new root ROI in an entirely manual manner if necessary. 311 3.2.2. Root detection and segmentation 312 From this point, the goal is creating the root mask, i.e. a binary mask for each root 313 ROI in which root pixels are represented in white and background pixels, in black. 314 In an attempt to make MyROOT 2.0 useful under a wide range of image acquisition 315 conditions (e.g. different illuminant types, or changes in camera settings), we aim at 316 making the root segmentation process robust to these variations by first conducting 317 a color normalization step. Taking into account that roots and leaves are typically 318 whitish and greenish, respectively, we apply a transformation that makes these colors 319 contrast with the dark background to simplify the subsequent segmentation steps. This 320 transformation consists in converting the image to the HSV color model and obtaining 321 a hue-independent new version of the image by combining the saturation and value 322 channels as (1 − S). ∗V . 323 Next, we perform root detection which not only considers the intensity of root 324 pixels, but also considers the topological structure of the roots with respect to their 325 surroundings. Indeed, as roots have a tubular structure (e.g. similar to blood vessels), 326 we use ridge detection for determining which pixels correspond to the roots. To that 327 end, we compute the local blockwise standard deviation of the color normalized image 14 Figure 6: Root segmentation intermediate steps: color normalization, ridge detection and binarization. 328 at a 5x5 pixels scale. 329 Finally, we must take into account that false ridge detections may occur caused by 330 scratches, root reflections or condensation droplets on the plate lid. For this reason, we 331 finally apply a binarization on the image obtained from the ridge detection, and remove 332 small connected components to get rid of noisy detections. As a result, a binary mask 333 corresponding to the roots is obtained. 334 Figure 6 presents the results of the color normalization, ridge detection and bina- 335 rization processes just described. 336 3.3. Root measurement 337 This module takes the mask generated by the roots segmentation block described 338 in the previous section and conducts the individual measurement of each root. 339 To measure root length, three sequential processes are conducted: root tracking, 340 hypocotyl detection and root measurement. 341 3.3.1. Root tracking 342 First, root tracking aims at providing a complete list of pixels belonging to each 343 root. In this work, we propose a bottom-up tracking approach that searches for white 15 (a) (b) Figure 7: Root tracking intermediate steps: (a) bottom-up tracking, and (b) root identification (right). 344 blobs upwards row by row, starting from the root tip and finishing at the line traced 345 after the hypocotyl detection process (described in Section 3.3.2). 346 The tracking process assumes constant velocity (i.e. soft constant changes in po- 347 sition). Using variable t to refer to rows, the position of a track in the next state (i.e. 348 row) is predicted as χ̄t = χt +Vχ , with velocity Vχ = χt+1 − χt . 349 The tracking process starts at the lowest row of the image where at least one root is 350 detected. We refer to this detection as dkt , where t refers to the row and k refers to the 351 root. These detections are depicted as green circles in the lowest part of Figure 7(a). 352 Upon each new detection, a new root (χkt ) is generated, and it is assigned a zero initial 353 velocity (Vχk = 0), and a predicted position equal to its original position (χ̄kt = χkt ). 354 At this point, all the created roots –namely n– are collected creating the root track set 355 (R = {[χ1t ,Vχ1 , χ̄1t ]; . . . ; [χnt ,Vχn , χ̄nt ]}). 356 The next step of the tracking consists in detecting white blobs in the next row 357 (d t+1 ). Then, using a matching process based on distance we pair detections with 358 predictions (d t+1 =⇒ χ̄ t ), which are represented by dashed red circles and arrows in 359 Figure 7(a). As a result of this matching process, three main different situations may 360 be encountered: 361 • Correct match: the detection is matched to a existing root (indicated by gray 362 circles in Figure 7(a)). In this case, the values of the current state of the root 16 363 (velocity and predicted position) are updated (χ = d, Vχ = χt+1 − χt , χ̄ t+1 = 364 χ t +Vχ ). 365 • Missed match: this means that one of the existing roots (χk ) does not match with 366 any detection (depicted as blue circles in Figure 7(a)). In this case, the prediction 367 is assigned as current state (χk = χ¯k ), while velocity and prediction are updated 368 ¯ = χ t +V ). (V = χ t+1 − χ t , χ t+1 χk k k k k χk 369 • Missed detection: this happens when the current detection (dk ) does not match 370 with none of the existing roots (shown as green circles in Figure 7(a)). This 371 means that a new root has been found, so it is added to our root set (R = 372 [R; [χ t = dk ,Vχ = 0, χ̄ t = dk ]]). 373 During the root tracking process, two more special cases may occur. First, when a 374 current root matches with more than one detection, which happens when a split occurs 375 (purple circle in Figure 7(a)). In this case, we create a new root sharing the same his- 376 torical record as the old one. And second, when two roots converge and, consequently, 377 they both match with a single detection (yellow circle in Figure 7(a)). In this situation, 378 the shortest root is eliminated from the root set and is added as a sub-root of the longest 379 one. 380 Finally, using the detections history we can determine which pixels belong to each 381 root, thus completing the root identification process (Figure 7(b)). 382 3.3.2. Hypocotyl detection 383 As mentioned earlier, the hypocotyl is the stem of a germinating seedling, found 384 below the cotyledons (seed leaves) and above the radicle (root). During the first days of 385 germination, the hypocotyl is clearly visible as a brownish oval in the junction between 386 the leaves and the root. In later stages of plant development, the growth of the leaves 387 hides the hypocotyl, although its position in space is invariant along time. Thus, its 388 detection allows setting the start point of the roots. 389 However, the shape and appearance of hypocotyls may vary from one seedling to 390 another. For this reason, as in this work we assume that seedlings are placed in rows, 391 the detection of a few hypocotyls will enables us to trace a regression curve to estimate 17 392 the position of the non-detected ones. The intersection between the regression curve 393 and each root will be used as the point at which the root will be cut (in terms of length 394 measurement), as illustrated in Figure 8. 395 To detect the hypocotyls, we follow a classic object detection approach based on 396 learning the visual characteristics of the hypocotyls by training a supervised classifier 397 on a labeled set of images containing positive and negative examples. In particular, 398 we propose extracting visual image features to learn color and appearance models – 399 using Histograms of Oriented Gradients (HOG) [31] as the appearance descriptor, and 400 color histograms as the color descriptor–, plus using a linear Support Vector Machine 401 (linSVM) as the classifier. 402 To train the hypocotyl detector, we use a manually annotated hypocotyl positive 403 and negative examples to learn the color and appearance models (see Figure 9 for a 404 few samples of the positive examples. 405 To compute appearance and color descriptors, a 48 × 48 pixels sample window is 406 divided in regular not overlapping 8 × 8 pixels cells. The HOG descriptor is obtained 407 by concatenating and normalizing histograms of gradient orientations computed inside 408 each cell, while the color descriptor consists in concatenated color histograms (one per 409 channel) computed using the same cells distribution. The color and appearance descrip- 410 tors are used to train the linSVM classifier, which learns the max-margin hyperplane 411 that better splits these samples in the descriptor space. 412 Once the detector is trained, we can proceed to detect hypocotyls on the input Figure 8: Regression curve (in blue) traced after the detection of hypocotyls (represented by black circles). 18 Crop Ruler Figure 9: A few positive examples used for training the hypocotyl detector. 413 image. Using a sliding window approach, an exhaustive search of hypocotyls is con- 414 ducted. Finally, while maintaining the highest scored windows as true detections, we 415 define a hypocotyl detection curve using polynomial regression. 416 3.3.3. Root measurement 417 Once the root tracking process is completed and the hypocotyl detection regression 418 curve is traced, each root is measured based on its tracked previous positions. 419 We first compute the root length in pixels by adding the distances between consec- 420 utive positions, and then obtain the actual root length by multiplying the pixel length 421 by the pixels-to-millimeters conversion factor. 422 Next, we apply a two-stage refinement process: in the first step, those roots that 423 have been tracked beyond the hypocotyl regression curve are cut at the point where the 424 curve intersects the root, updating the length. 425 And in the second step, we remove unwanted roots. This can be done in two ways: 426 i) we can remove noisy roots, i.e. those roots which are a percentage shorter than the 427 longest measured root, as usually roots on a plate tend to have similar lengths, and 428 ii) we can remove any root at will if, for some reason, the user considers it should be 429 excluded from the measurement. 19 (a) (b) (c) Figure 10: Root length measurement refinement steps: (a) original image, (b) root cut at hypocotyl regression curve, and (c) root removal. 430 Figure 10 illustrates the different stages of the root length measurement refinement 431 process just described. Starting with the measurements presented in Figure 10(a), roots 432 are first cut at the hypocotyl regression curve length (see roots #4, #5 and #6 in Figure 433 10(b)). Then, the user can select removing any root from the analysis, such as root #3 434 in Figure 10(c). 435 Finally, as a unique identifier is given to each root, and the individual root length 436 measurement results are stored in RSML, Excel and plain text files for subsequent 437 analysis. Moreover, a labeled version of the input image presenting the root identifiers 438 and the length measurements is also generated and saved (see Figure 11). 20 Figure 11: Results of the individual roots measurements. 439 4. Experiments 440 This section presents a series of experiments oriented to evaluate different aspects 441 of the algorithms that implement MyROOT 2.0. First, we describe the set of images 442 of agar plates with Arabidopsis seedlings used for conducting the experiments. Sec- 443 ond, we present an experiment in which the precision of the automatic detection and 444 analysis of the measuring tape is tested. Third, we evaluate the root ROI detection pro- 445 cess. Next, we analyze the accuracy of the hypocotyl detection algorithm by training 446 the detector upon different color and appearance models built using wild-type and mu- 447 tant hypocotyls. And finally, we present an experiment that compares the root length 448 measurements obtained by MyROOT 2.0 with the measurements obtained by two state- 449 of-the-art software tools: BRAT and EZ-Rhizo. 450 4.1. Image data and experimental setup 451 In order to validate the proposed framework, a set of 120 images with 4928×3264 452 resolution was acquired by the plant scientists in our team using a D7000 Nikon cam- 453 era. To take the pictures, the agar plates were laid on a dark surface and a measuring 454 tape at least 1 cm long was placed on them. The images were taken on different days 455 by different users, always from a top perspective and under slightly different conditions 456 as regards distance and illumination. The images shown in Figure 12 represent a small 457 sample of such pictures. We have selected several subsets of these images to conduct 458 the experiments described in the forthcoming sections. 21 459 As for the setup of the agar plates, the Arabidopsis seedlings have been arranged 460 in one or two rows of around 15 seeds each. The seedlings have been grown in lab 461 conditions (controlled illumination, temperature and humidity) during a period of time 462 ranging from three to eight days, which allows not only conducting instant root length 463 measurement experiments, but also performing continuous root length measurements 464 of growing roots during successive days. 465 Finally, the manual measurements conducted by expert plant scientists, they have 466 been completed using the ImageJ open source image processing program [32]. 467 4.2. Automatic image scale from measuring tape detection and analysis 468 This first experiment evaluates the accuracy of MyROOT 2.0 to automatically de- 469 termine the image scale from the detection and analysis of the measuring tape. To that 470 end, a set of 20 images was processed both by means of MyROOT 2.0 and manually by 471 at least three expert plant scientists to obtain the equivalence between pixels and one 472 centimeter. 473 To obtain the equivalence manually with ImageJ, a line was drawn over one cen- 474 timeter of the measuring tape in the image. Then, using the Analyze → Set scale menu 475 option of ImageJ, the distance of that line was set as the equivalent to 10 mm. 476 To test the robustness of the measuring tape detection and analysis algorithms 477 against changes in image characteristics, images with different scales and illumina- 478 tion conditions were selected, and also with different placement, size and orientation 479 of the measuring tape (see Figure 12). 480 Table 1 presents the pixels-to-centimeter equivalences obtained by manual mea- 481 surements (average of values obtained by the plant scientists who did the experiment) 482 and by means of MyROOT 2.0. It can be observed that the absolute difference be- 483 tween measurements is smaller than 1 pixel in 13 of the 20 plates, and only in 3 of 484 the 20 plates this difference is larger than 2 pixels. The correlation coefficient between 485 manual and MyROOT 2.0 measurements is 0.999. 486 On average, the average absolute difference between manual and MyROOT 2.0 487 measurements is 0.93 pixels. This means that MyROOT 2.0 attains subpixel precision 22 Figure 12: Samples of the set of images taken by the plant scientists in our team used for evaluating MyROOT 2.0. 488 at determining the pixels-to-centimeters equivalence from the analysis of the measuring 489 tape placed on the plate lid. 490 It is interesting to note that, in order to obtain satisfactory results, the Ruler Thresh- 491 old (RT ) parameter had to be tuned for a few of the images in the testbed, depending 492 on the illumination conditions. For most of them, the default RT = 0.7 value was ad- 493 equate. However, this parameter had to be increased (typically to RT = 0.9) for more 494 brightly illuminated images, or decreased to RT = 0.5 for darker images. 495 4.3. Root ROI Detection 496 In this experiment, we evaluate the automatic detection of the roots ROI. To that 497 end, we selected a set of 45 images in which seedlings were arranged in one or two 498 rows, which are the most common way plant scientists place seedlings on Petri dishes 499 for their experiments. In particular, 20 images contained one row of seedlings, while 500 the 25 remaining images contained two rows, which amounts to a total 70 root ROI. 501 Moreover, these images were taken by different plant scientists at different points in 502 time, so they vary in terms of illumination conditions and distance to the camera. This 23 Table 1: Average manual vs. MyROOT 2.0 pixels-to-centimeter equivalence (in pixels). Plate Manual MyROOT 2.0 Absolute difference 1 226.32 226.67 0.35 2 244.00 243.33 0.67 3 237.00 235.00 2.00 4 254.00 253.33 0.67 5 232.00 231.67 0.33 6 209.19 206.67 2.52 7 233.00 231.67 1.33 8 229.50 230.00 0.50 9 283.57 281.67 1.90 10 217.00 216.67 0.33 11 268.09 265.83 2.26 12 269.33 269.17 0.16 13 255.71 253.33 2.38 14 276.31 276.67 0.36 15 241.43 240.00 1.43 16 248.50 248.33 0.17 17 250.13 250.00 0.13 18 215.26 215.00 0.26 19 215.65 215.83 0.18 20 215.59 215.00 0.59 503 enabled us to test the algorithm in a wide range of situations. 504 We first evaluate the accuracy of the algorithm to determine the correct number of 505 seedling rows in each image. In 44 of the 45 images, the correct number of seedling 506 rows is detected. The only error occurred in an image containing two rows, but only 507 one was detected because it contained less than half the seedlings of the other and did 508 not surpass the predefined threshold to be considered as a row. However, as mentioned 509 earlier, MyROOT 2.0 allows the user to define root ROI manually after the automatic 510 detection, so this infrequent problem can be circumvented. 511 Next, we evaluate if the root ROI bounding boxes automatically generated by My- 24 512 ROOT 2.0 contain all the complete roots of the corresponding seedling row. This is 513 an important metric as regards the ability of MyROOT 2.0 to provide an automatic 514 segmentation of the regions of interest of the image without user supervision. 515 The obtained results show that 98.6% of the detected root ROI contained all the 516 roots in the corresponding seedling row. Moreover, 92.9% of the root ROI contained 517 the complete roots, and only in a few cases did the longest roots extend beyond the 518 limits of the bounding boxes. 519 These results highlight that the root ROI detection module is capable of providing a 520 highly accurate segmentation of the regions of interest in the image in a fully automatic 521 manner. However, thanks to the interactive design of MyROOT 2.0, a user that was not 522 satisfied with the detected root ROI could easily add additional ROI or modify the 523 shape of the detected ones, and proceed with the analysis. 524 4.4. Hypocotyl Detection 525 This experiment evaluates the algorithm designed for setting, in an automatic man- 526 ner, the starting point of each root by means of hypocotyl detection. In particular, we 527 evaluate the performance of the detector depending on the morphological differences 528 of hypocotyls caused by different gene expression patterns, a typical situation encoun- 529 tered in plant science. 530 To that end, we selected a set of 47 images containing hypocotyls with varied mor- 531 phologies: wild-type and mutant (long, short and dwarf). These images were manually 532 annotated to define the bounding boxes of the hypocotyls to generate positive exam- 533 ples to train the hypocotyl detector and also to test its performance. In total, 1737 534 hypocotyls were annotated (907 wild-type and 830 mutants), which were also flipped 535 to obtain a total 3474 positive examples. 536 These examples were divided into a training and a test set comprising 958 hypocotyls 537 (530 wild-type and 428 mutant) and 779 hypocotyls (377 wild-type and 402 mutants), 538 respectively. 539 The training examples were used to build detection models of i) only mutant, ii) 540 only wild-type, and iii) all (mutant plus wild-type), which were tested to detect only 541 mutant, only wild-type, and mutant plus wild-type hypocotyls. By doing so, we eval- 25 Test: WT Test: Mutants 1 1 0.490 - Model All 0.526 - Model All 0.397 - Model WT 0.474 - Model WT 0.500 - Model Mut 0.454 - Model Mut 0.8 0.8 0.6 0.6 Precision Precision 0.4 0.4 0.2 0.2 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Recall Recall (a) Only wild-type (b) Only mutant Test: All 1 0.452 - Model All 0.414 - Model WT 0.385 - Model Mut 0.8 0.6 Precision 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 Recall (c) All Figure 13: Precision-Recall curves of the hypocotyl detection experiments. 542 uate how hypocotyl appearance variability caused by genetic mutations affects the de- 543 tection process. 544 To evaluate the hypocotyl detection process we compute the Precision-Recall (PR) 545 curve in the three detection scenarios. Figures 13(a), 13(b) and 13(c) present the 546 PR curves obtained for the only wild-type, only mutant, and mutant plus wild-type 547 hypocotyl detection experiments. The legends of each figure show the average preci- 548 sion obtained by each trained model. 549 It can be observed that the benefit obtained from using detection models specifi- 550 cally based on a particular type of hypocotyls is variable. On the one hand, the highest 551 average precision obtained when detecting mutant hypocotyls is obtained when the de- 552 tector is trained on only-mutant positive examples (see Figure 13(b)). On the other 553 hand, the best performance attained when detecting wild-type hypocotyls is given by 26 554 the wild-type plus mutant detector (see Figure 13(a)). This result is probably caused by 555 the inherent appearance variability of wild-type hypocotyls, while their mutant coun- 556 terparts show more uniform morphological traits. 557 However, it is not unusual to have mixed types of hypocotyls on a single plate. 558 Thus, it makes sense to conduct an experiment that aims to detect hypocotyls of any 559 class. The PR curves corresponding to this experiment are presented in Figure 13(c). 560 In this case, using a detector trained on both mutant and wild-type hypocotyls yields 561 the highest average precision, outperforming specific models by more than 4%. 562 4.5. Root Measurement 563 To finally validate the proposed framework, this experiment compares the perfor- 564 mance of MyROOT 2.0 with other available software tools. For this reason, we com- 565 pare manual measurements to those obtained by MyROOT 2.0 and two state-of-the-art 566 root length measurement software tools: BRAT and EZ-Rhizo. 567 Manual measurements were made using the Segmented line option in ImageJ. Each 568 root was tracked by clicking several times from the starting point of the root to the root 569 tip. Then, the length of the segmented line was measured and ImageJ obtained the root 570 length in millimeters using the scale set as described in Section 4.2. 571 In [7], this comparison was made on a set of images taken by our team of plant 572 scientists. In contrast, to make this comparison totally unbiased and more challenging, 573 in this study we used three images: one taken by our team of plant scientists (Figure 574 14(a), containing 28 roots), and the example images provided with the BRAT and EZ- 575 Rhizo software tools (Figures 14(b) and 14(c)), containing 24 and 4 roots, respectively. 576 Thus, we ensure that the testbed includes images deemed as ideal for test by the authors 577 of the mentioned software tools. 578 First, we compare the ability of the three software tools to correctly detect roots. 579 Table 2 shows the number and percentage of roots correctly detected by each software 580 on each plate. It can be observed that MyROOT 2.0 and EZ-Rhizo are capable of 581 detecting most roots (94.6% and 89.3%, respectively), while BRAT shows a really low 582 root detection accuracy (69.6%) when applied on images the illumination conditions of 583 which differ from those considered ideal for its configuration. Moreover, BRAT offers 27 (a) Plate MyROOT 2.0 (b) Plate BRAT (c) Plate EZ-Rhizo Figure 14: The three images used for making the comparison between MyROOT 2.0, BRAT and EZ-Rhizo. Table 2: Comparison between MyROOT 2.0, BRAT and EZ-Rhizo in terms of root detection accuracy (number and percentage of detected roots). MyROOT 2.0 BRAT EZ-Rhizo Plate MyROOT 2.0 26/28 (92.9%) 15/28 (53.6%) 22/28 (78.6%) Plate BRAT 24/24 (100%) 24/24 (100%) 24/24 (100%) Plate EZ-Rhizo 3/4 (75%) 0/4 (0%) 4/4 (100%) 584 no controls (unlike MyROOT 2.0 and EZ-Rhizo) to tune the parameters that control 585 root detection, so little can be done if it fails at detecting roots. 586 And second, we compare the accuracy of the root length measurements offered 587 by the three softwares. As a ground truth, we used the manual measurements by an 588 expert plant scientist in our team. Table 3 presents the average absolute differences 589 between the manual measurements of all the roots in each plate and the measurements 590 yielded by each software tool. It can be observed that MyROOT 2.0 offers by far the 591 highest accuracy on the MyROOT 2.0 and EZ-Rhizo plates, while achieving nearly 592 equal accurate measurements as BRAT on the BRAT plate. If the average absolute 593 error is computed over all the set of roots in the test bed, MyROOT 2.0 attains 0.62 mm 594 error, while this error is as high as 3.43 mm and 3.05 mm for BRAT and EZ-Rhizo, 595 respectively. For a detailed view of the individual root length measurements, please 596 refer to Appendix A. 597 These results indicate that MyROOT 2.0 is a truly useful to obtain accurate primary 28 Table 3: Comparison between MyROOT 2.0, BRAT and EZ-Rhizo in terms of root length measurement accuracy with respect to manual measurements (average absolute measurement error in millimeters). MyROOT 2.0 BRAT EZ-Rhizo Plate MyROOT 2.0 0.38 7.98 2.66 Plate BRAT 0.79 0.59 3.30 Plate EZ-Rhizo 1.35 - 3.78 598 root length measurements in an automatic manner. While it is fair to say that BRAT 599 is really accurate at measuring roots, it is also true that it has a high root detection 600 error rate. On the other hand, EZ-Rhizo provides little accurate measurements when 601 compared to MyROOT 2.0. This is due to the fact that the hypocotyl detection module 602 based on visual appearance models of MyROOT 2.0 allows to adjust the root measure- 603 ment to its true limits. In contrast, EZ-Rhizo includes a very simple hypocotyl detection 604 step based on color, which does not improve the accuracy of its measurements. 605 5. Discussion 606 Plant phenotyping is current topic of debate with a wide range of agricultural ap- 607 plications. Among all plant organs, the root is essential for overall plant growth and 608 development. Plant scientists have long used the primary root of Arabidopsis as a 609 developmental model due to its simple and stereotyped cell type organization, as its 610 length is a phenotypic trait related to key aspects of plant growth and viability with 611 valuable implications in agriculture. 612 To reduce the human effort required for measuring the length of the primary root 613 of Arabidopsis seedlings grown on agar plates in laboratory conditions, this work has 614 introduced MyROOT 2.0, an automatic tool designed jointly by computer vision and 615 plant biology experts. 616 When evaluating its ability to infer image scale, MyROOT 2.0 attains subpixel 617 precision to determine the correspondence between pixels and centimeters, analyzing 618 to that end a segment of measuring tape placed on the plate lid. This avoids setting strict 619 requirements in the image acquisition process. Moreover, it is also capable of detecting 620 the regions of the image where the roots are arranged without user supervision. 29 621 To increase the accuracy of root length measurement, MyROOT 2.0 includes an ob- 622 ject detection module trained to find the starting points of the roots, i.e. the hypocotyls. 623 Due to the morphologic differences that the expression of different genes may cause 624 on the appearance of hypocotyls, different color and appearance models have been em- 625 ployed and tested to detect the hypocotyls. The conducted experiments indicate that 626 training the detector with both mutant and wild-type hypocotyls yields the highest av- 627 erage detection precision, thus accommodating appearance variations due to genetic 628 causes. 629 When comparing the manual root length measurements made by expert plant sci- 630 entists and those obtained automatically via MyROOT 2.0, average differences below 1 631 millimeter are found over nearly 300 individual root measurements, which is a satisfac- 632 tory result considering the huge savings in manual labor derived from using MyROOT 633 2.0. 634 Finally, the comparison between MyROOT 2.0 and two well-established state-of- 635 the-art root length measurement software tools (BRAT and EZ-Rhizo) reveals that our 636 proposal outperforms its competitors in terms of root detection and length measurement 637 accuracies, achieving error measurements four to five times smaller. 638 All these features, coupled with its ability to process batches of images, makes 639 MyROOT 2.0 a very appealing tool to perform high throughput automatic root pheno- 640 typing. 641 It is important to highlight that the involvement of plant scientists in the develop- 642 ment loop of MyROOT 2.0 has made its usability a priority. For this reason, we have 643 minimized the number of parameters adjustable by the user, leaving only those directly 644 related with the image appearance (lighting condition, contrast) or the post processing 645 of detected roots (e.g. root length filtering) available for the user to be set manually. 646 All the many other parameters involved in the image processing pipeline (measuring 647 tape detection, ROIs detection, root mask generation and root tracking, hypocotyl de-648 tection) have been set for a robust and optimal performance. By doing so, we have 649 obtained a software tool that is highly usable for a wide range of images with different 650 acquisition conditions. 651 Further research will be oriented towards the development of a day-by-day root 30 652 growth monitoring algorithm that, starting from the detection of the hypocotyl, allows 653 to detect abnormal root growth patterns. Our future plan also contemplates the in- 654 corporation of additional root architecture traits beyond primary root length, such as 655 detecting branching points and measuring secondary roots. This will be important for 656 the high throughput analysis of overall root system architecture traits in both plant 657 biology and agricultural phenotyping. 658 Software download: 659 MyROOT 2.0 is available for download at https://bit.ly/2NtQDZB. 660 Author Contributions: 661 A.G., M.F. and X.S. developed the algorithms for the method. A.G., X.S., I.B- 662 P. and D.B-E. performed the validation experiments. I.B-P. and D.B-E. acquired the 663 dataset. X.S. and A.I.C-D. designed and supervised the study. X.S. and A.G. wrote the 664 manuscript, and X.S., A.I.C-D. and I.B-P. edited the manuscript until publication. 665 Funding 666 A.I.C-D. is a recipient of a BIO2016-78955 grants from the Spanish Ministry of 667 Economy and Competitiveness and a European Research Council, ERC Consolidator 668 Grant (ERC-2015-CoG 683163). I.B-P. is funded by the FPU15/02822 grant from 669 the Spanish Ministry of Education, Culture and Sport. D.B-E. is contracted with the 670 BIO2016-78955 grant in the A.I.C-D laboratory. CRAG is funded by Severo Ochoa 671 Programme from Centers of Excellence in R&D 2016-2019 (SEV-2015-485 0533). 672 Conflicts of Interest: 673 The authors declare no conflict of interest. The funders had no role in the design 674 of the study; in the collection, analyses, or interpretation of data; in the writing of the 675 manuscript, or in the decision to publish the results. 31 676 Appendix A. Manual vs. MyROOT 2.0, BRAT and EZ-Rhizo root length mea- 677 surements 678 This section presents the individual root length measurements corresponding to the 679 experiment described in Section 4.5. In particular, Table A.4 presents the root length 680 measurements obtained manually, via MyROOT 2.0, BRAT and EZ-Rhizo. Each row 681 corresponds to an individual root, which is identified by an ID comprising two capital 682 letters (MR, BR or EZ, to indicate if it belongs to the MyROOT 2.0, BRAT or EZ-Rhizo 683 plate) and a number. 684 References 685 [1] K. Wetterstrand, DNA Sequencing Costs: Data from the NHGRI Genome 686 Sequencing Program (GSP), http://www.genome.gov/sequencingcosts (last ac- 687 cessed on September 2017) (2014). 688 [2] R. C. Kuijken, F. A. van Eeuwijk, L. F. Marcelis, H. J. Bouwmeester, Root phe- 689 notyping: from component trait in the lab to breeding, Journal of Experimental 690 Botany 66 (18) (2015) 5389. 691 [3] A. Wasson, R. Richards, R. Chatrath, S. Misra, S. Prasad, G. Rebetzke, 692 J. Kirkegaard, J. Christopher, M. Watt, Traits and selection strategies to improve 693 root systems and water uptake in water-limited wheat crops., Journal of Experi- 694 mental Botany 63 (2012) 3485–3498. 695 [4] D. Dresboll, K. Thorup-Kristensen, B. McKenzie, L. Dupuy, A. Bengough, Time- 696 lapse scanning reveals spatial variation in tomato (solanum lycopersicum l.) root 697 elongation rates during partial waterlogging., Plant and Soil 369 (2013) 467–477. 698 [5] N. Ytting, S. Andersen, K. Thorup-Kristensen, Using tube rhizotrons to measure 699 variation in depth penetration rate among modern north-european winter wheat 700 (triticum aestivum l.) cultivars., Euphytica 199 (2014) 233–245. 32 701 [6] N. Fàbregas, F. Lozano-Elena, D. Blasco-Escámez, T. Tohge, C. Martı́nez- 702 Andújar, A. Albacete, S. Osorio, M. Bustamante, J. L. Riechmann, T. No- 703 mura, et al., Overexpression of the vascular brassinosteroid receptor brl3 confers 704 drought resistance without penalizing plant growth, Nature communications 9 (1) 705 (2018) 4680. 706 [7] I. Betegon-Putze, A. González, X. Sevillano, D. Blasco-Escámez, A. I. Cao- 707 Delgado, Myroot: A method and software for the semiautomatic measurement of 708 primary root length in arabidopsis seedlings, The Plant Journal (in press) (2019). 709 [8] J. Arsenault, S. Poulcur, C. Messier, R. Guay, Winrhlzo: A root measuring system 710 with a unique overlap correction method., HortScience 30 (4) (1995) 906. 711 [9] M. Mühlich, D. Truhn, N. K.A., A. Walter, S. H., A. T., Measuring plant root 712 growth., in: Proc. 30th DAGM Symposium, 2008, pp. 497–506. 713 [10] K. Nagel, B. Kastenholz, S. Jahnke, D. van Dusschoten, T. Aach, M. Mühlich, 714 D. Truhn, H. Scharr, S. Terjung, A. Walter, U. Schurr, Temperature responses of 715 roots: impact on growth, root system architecture and implications for phenotyp- 716 ing., Functional Plant Biology 36 (2009) 947–959. 717 [11] P. Armengaud, K. Zambaux, A. Hills, R. Sulpice, R. Pattison, M. Blatt, A. Amt- 718 mann, Ez-rhizo: Integrated software for the fast and accurate measurement of 719 root system architecture., Plant Journal 57 (2009) 945–956. 720 [12] A. French, S. Ubeda-Tomás, T. Holman, M. Bennett, T. Pridmore, High- 721 throughput quantification of root growth using a novel image-analysis tool., Plant 722 Physiology 150 (2009) 1784–1795. 723 [13] M. Isard, A. Blake, Condensation: conditional density propagation for visual 724 tracking., IJCV 29 (1998) 5–28. 725 [14] A. Naeem, A. P. French, D. M. Wells, T. P. Pridmore, High-throughput feature 726 counting and measurement of roots, Bioinformatics 27 (9) (2011) 1337. 33 727 [15] D. M. Wells, A. P. French, A. Naeem, O. Ishaq, R. Traini, H. Hijazi, M. J. Bennett, 728 T. P. Pridmore, Recovering the dynamics of root growth and development using 729 novel image acquisition and analysis methods, Philosophical Transactions of the 730 Royal Society B: Biological Sciences 367 (1595) (2012) 1517–1524. 731 [16] N. Yazdanbakhsh, J. Fisahn, High throughput phenotyping of root growth dynam- 732 ics, lateral root formation, root architecture and root hair development enabled by 733 PlaRoM., Functional Plant Biology 36 (2009) 938–946. 734 [17] N. Yazdanbakhsh, J. Fisahn, Development of a robot-based platform applied to 735 simultaneous root growth profiling of seedlings growing in a petri dish., in: Proc. 736 8th WSEAS Int’l Conference on Mathematics and Computers in Biology and 737 Chemistry, 2007, pp. 69–73. 738 [18] J. Le Bot, V. Serra, J. Fabre, X. Draye, S. Adamowicz, L. Pagès, Dart: a software 739 to analyse root system architecture and development from captured images, Plant 740 and Soil 326 (1) (2010) 261–273. 741 [19] A. Iyer-Pascuzzi, O. Symonova, Y. Mileyko, Y. Hao, H. Belcher, J. Harer, 742 J. Weitz, P. Benfey, Imaging and analysis platform for automatic phenotyping and 743 trait ranking of plant root systems., Plant Physiology 152 (1) (2010) 1148–1157. 744 [20] G. Lobet, L. Pagès, X. Draye, A novel image-analysis toolbox enabling quantita- 745 tive analysis of root system architecture, Plant physiology 157 (1) (2011) 29–39. 746 [21] T. Galkovskyi, Y. Mileyko, A. Bucksch, B. Moore, O. Symonova, C. Price, 747 C. Topp, A. Iyer-Pascuzzi, P. Zurek, S. Fang, J. Harer, P. Benfey, J. Weitz, Gia 748 roots: software for the high throughput analysis of plant root system architecture., 749 BMC Plant Biology 12 (1) (2012) 116. 750 [22] R. Clark, A. Famoso, K. Zhao, J. Shaff, E. Craft, C. Bustamante, S. McCouch, 751 D. Aneshansley, L. Kochian, High-throughput two-dimensional root system phe- 752 notyping platform facilitates genetic analysis of root growth and development., 753 Plant, Cell & Environment 36 (2013) 454–466. 34 754 [23] L. Benoit, D. Rousseau, E. Belin, D. Demilly, S. Ducournau, F. Chapeau- 755 Blondeau, C. Dürr, Locally oriented anisotropic image diffusion: application to 756 phenotyping of seedlings., in: Proc. 8th International Joint Conference on Com- 757 puter Vision, Imaging and Computer Graphics Theory and Applications, 2013. 758 [24] M. P. Pound, A. P. French, J. A. Atkinson, D. M. Wells, M. J. Bennett, T. Prid- 759 more, Rootnav: Navigating images of complex root architectures, Plant Physiol- 760 ogy 162 (4) (2013) 1802–1814. 761 [25] D. Ristova, U. Rosas, G. Krouk, S. Ruffel, K. D. Birnbaum, G. M. Coruzzi, 762 Rootscape: A landmark-based system for rapid screening of root architecture in 763 arabidopsis, Plant Physiology 161 (3) (2013) 1086–1096. 764 [26] R. Slovak, C. Goschl, X. Su, K. Shimotani, T. Shiina, W. Busch, A scalable open- 765 source pipeline for large-scale root phenotyping of Arabidopsis., Plant Cell 26 766 (2014) 2390–2403. 767 [27] L. Remmler, L. Clairmont, A. Rolland-Lagan, F. Guinel, Standardized mapping 768 of nodulation patterns in legume roots., New Phytologist 202 (2014) 1083–1094. 769 [28] P. Kumar, C. Huang, J. Cai, , S. Miklavcic, Root phenotyping by root tip detection 770 and classification through statistical learning., Plant and Soil 380 (1) (2014) 193– 771 209. 772 [29] J. Cai, Z. Zeng, J. N. Connor, C. Y. Huang, V. Melino, P. Kumar, S. J. Miklavcic, 773 Rootgraph: a graphic optimization tool for automated image analysis of plant 774 roots, Journal of Experimental Botany 66 (21) (2015) 6551–6562. 775 [30] J. Wu, Q. Wu, L. Pagès, Y. Yuan, X. Zhang, M. Du, X. Tian, Z. Li, Rhizochamber- 776 monitor: a robotic platform and software enabling characterization of root 777 growth, Plant Methods 14 (1) (2018) 44. 778 [31] N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in: 779 CVPR, San Diego, CA, USA, 2005. 35 780 [32] J. Schindelin, C. Rueden, M. Hiner, K. Elicein, The ImageJ ecosystem: An open 781 platform for biomedical image analysis., Molecular Reproduction and Develop- 782 ment 82(7-8) (2015) 518–529. 36 Table A.4: Manual vs. MyROOT 2.0, BRAT and EZ-Rhizo root length measurements (all values in millime- ters). ID Manual MyROOT 2.0 BRAT EZ-Rhizo MR1 24.182 23.630 15.913 24.680 MR2 22.519 22.896 0.139 22.460 MR3 25.546 25.886 - - MR4 20.188 20.615 10.800 23.630 MR5 24.077 24.500 - - MR6 19.801 19.709 - 23.100 MR7 24.183 24.240 - 27.730 MR8 24.550 24.172 - 26.600 MR9 26.399 25.627 0.112 28.560 MR10 22.051 21.829 - - MR11 21.212 21.082 - 27.160 MR12 24.734 24.842 - 33.160 MR13 20.613 20.864 15.227 22.820 MR14 19.293 - - - MR15 18.665 19.469 - 19.840 MR16 19.772 20.690 2.236 21.080 MR17 21.525 21.757 22.942 23.530 MR18 21.154 21.282 - 23.850 MR19 16.187 17.109 - 15.890 MR20 24.471 25.328 9.394 27.610 MR21 12.127 12.455 12.341 - MR22 9.412 9.608 8.798 7.390 MR23 11.384 11.477 10.396 13.690 MR24 10.594 10.332 10.000 - MR25 18.445 18.417 - 22.120 MR26 21.682 21.068 10.945 24.540 MR27 1.565 - 1.469 3.980 MR28 19.425 19.708 20.158 22.310 BR1 7.54 7.855 6.379 10.80 BR2 3.58 3.607 3.397 6.21 BR3 6.93 7.011 6.339 9.71 BR4 3.94 5.693 3.103 7.06 BR5 2.51 3.952 1.549 4.03 BR6 5.29 6.450 4.788 8.01 BR7 8.12 8.783 8.389 12.64 BR8 5.28 6.924 5.099 10.51 BR9 6.17 7.283 6.285 9.68 BR10 9.6 9.761 10.143 13.83 BR11 7.89 8.816 8.184 10.50 BR12 9.81 10.380 10.403 13.51 BR13 9.55 9.601 8.365 12.62 BR14 4.72 3.687 3.903 7.90 BR15 6.24 6.671 5.867 9.82 BR16 8.39 9.728 8.107 11.61 BR17 5.23 7.233 4.804 8.42 BR18 4.54 5.907 6.292 7.66 BR19 9.28 10.095 8.446 13.03 BR20 9.26 9.643 8.540 11.99 BR21 6.98 6.202 7.229 10.32 BR22 8.85 9.193 9.381 11.93 BR23 8.92 9.282 8.404 12.40 BR24 10.09 9.904 10.409 13.73 EZ1 69.43 70.944 - 73.87 EZ2 63.77 63.87 - 65.90 EZ3 66.29 68.738 - 70.42 EZ4 63.58 - - 67.98 37