Automated imaging coupled with AI-powered analysis accelerates the assessment of plant resistance to Tetranychus urticae

doi:10.21203/rs.3.rs-3097958/v1

Download PDF

Research Article

Automated imaging coupled with AI-powered analysis accelerates the assessment of plant resistance to Tetranychus urticae

https://doi.org/10.21203/rs.3.rs-3097958/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 04 Apr, 2024

Read the published version in Scientific Reports →

Version 1

posted

You are reading this latest preprint version

The two-spotted spider mite (TSSM), Tetranychus urticae, is one of the most destructive piercing-sucking herbivores, infesting more than 1,100 plant species. The TSSM has evolved a broad tolerance to different plant xenobiotics, influencing its flexibility to adapt to multiple host plants and pesticides. At the same time, the effective resistance loci in plants are still unknown. To find out more about plant-mite correlation, novel approaches are required allowing the efficient screening of large, genetically diverse populations of two interacting species. Here we propose an analytical pipeline based on high-resolution imaging of infested leaves and an artificial intelligence-based computer program, MITESPOTTER, for analysis of plant susceptibility. Our system precisely identifies and quantifies eggs, feces and damaged areas on phenotypically differentiated leaves. The new method was tested on 14 TSSM-infested Arabidopsis thaliana ecotypes derived from diverse world locations and showing remarkable differences in the listed symptoms. The proposed method also demonstrated the ecotype variation in the mite preference to the age of leaf and egg distribution on the ab/adaxial leaf surface. The presented analytical pipeline can be adapted to different pest and host species facilitating diverse experiments with a high number of specimens such as the screening of a large segregating population of plants leading to the identification of loci for efficient breeding of TSSM-resistant plants.

Tetranychus urticae

high-resolution imaging

artificial intelligence

automatic symptom detection

plant susceptibility variability

• A new protocol for the detection of Tetranychus urticae symptoms was developed, which involves high-resolution imaging and subsequent image analysis using an artificial intelligence-based computer program.

• The measurements revealed significant within-species variation in susceptibility, as well as patterns of symptom distribution on consecutive leaves of the Arabidopsis plant rosette or on the upper and lower sides of a leaf.

• The protocol is ready for optimization to other T. urticae host plant species, as well as for scaling up to monitor hundreds of genetically and phenotypically diversified specimens.

(Automated imaging coupled with AI-powered analysis accelerates the assessment of plant resistance to Tetranychus urticae)

It is well known that herbivorous pest fitness traits are differently affected by the host plants, therefore the classification of plant resistance most often relies on comparison of the development of the pest population, reproductive potential or severity of feeding symptoms. These methods, however, require broad expertise on the given pest biology and behavior. In addition, the difficulty and labor intensity increase the smaller the pest size, which necessitates the use of a microscope equipped with a high-resolution camera. Thus, assessments of host plant resistance to a generalist like the spider mite seem to be a serious bottleneck, especially when multiple pest variants and plant cultivars in diverse environmental conditions are to be compared at one time.

Our main object of interest is the phenomenon of natural host plants’ resistance to the two-spotted spider mite (TSSM; Tetranychus urticae Koch 1836, (Acariformes: Trombidiformes: Tetranychidae)). Its cosmopolitan and generalist nature — it feeds on over 1,100 host plants (van de Vrie et al., 1972; Migeon and Dorkeld, 2018) — together with its short lifespan, high female fecundity and efficient xenobiotic metabolism make TSSM one of the most significant agricultural pests globally. TSSM is a mesophyll cell-content feeder which uses a stylet formed by two cheliceral digits to empty the cell content, causing cell death and the appearance of chlorotic spots (Park and Lee, 2002). Under optimum conditions (temp. 25-30°C and 45-55% relative humidity), TSSM development (from egg to adult stage) can be completed in as little as seven days (Laing, 1969; Shih et al., 1976; Tehri, 2014). Such a short life cycle results in the emergence of up to 25 exponentially growing generations a year (Lucini et al., 2015). A mated female can lay up to 20–30 transparent eggs per day in optimal conditions (more than 200 over its lifespan) mostly on the abaxial side of a leaf (Helle and Sabelis, 1985; Gotoh, 1997; Krips et al., 1998; Brust and Gotoh, 2018). Moreover, TSSM’s high potential of xenobiotic metabolism results in a realistic possibility of developing pesticide resistance (Shih et al., 1976; Van Leeuwen et al., 2010; Rioja et al., 2017).

For all these reasons, new mite-pest resistant crop cultivars are needed. This requires an identification of new loci and the characterization of underlying plant resistance mechanisms. To date, molecular approaches have yielded some promising results but they were mainly focused on candidate gene selection based on their differential expression upon TSSM attack (Li et al., 2002; Santamaria et al., 2018). An alternative solution would be the screening of large polymorphic populations of host plant accessions and mapping chromosomal regions involved in quantitative or qualitative resistance-related traits (Kloth et al., 2012). To apply such an approach, a high-throughput method which quantifies mite activity on the host plant is needed. Such a method should be unbiased (avoiding human-based arbitrary assessment), precise (measuring instead of rating) and efficient (relatively short time of data acquisition). Moreover, there are several measures of plant susceptibility/resistance and they may be biased toward diverse resistance mechanisms, therefore simultaneous collection of data on several related traits is highly desirable (Li et al., 2014; Walter et al., 2015).

In recent years, there has been a lot of interest in the high-throughput phenotyping (HTP), which allows the collection of data on multiple traits of interest at one time by using automated systems of data collection and analysis. Most HTP protocols are based on low-resolution imaging using visible, multispectral, hyperspectral or fluorescence cameras, which allow the recording of traits both visible and invisible to the naked eye (Goggin et al., 2015; Herrmann et al., 2012, 2015; Martin and Latheef, 2018; Nieuwenhuizen et al., 2019; Uygun et al., 2020; Fraulo et al., 2009; Crockett et. al., 2014; Gonzalez-Gonzalez et al., 2021). A big advantage of such data collection is a relatively short recording time and the possibility of storing and reanalyzing data at any time (Goggin et al., 2015). However, they might suffer from lack of specificity i.e. identification of the particular pest species.

Here, we present a semi-automated protocol to monitor T. urticae activity on Arabidopsis thaliana (L.) Heynh (Brassicaceae) plants by detection and analysis of leaf damage, eggs and feces. The central part of this protocol is an artificial intelligence (AI)-powered computer program, MITESPOTTER, which automatically analyzes high resolution images of mite-infested A. thaliana leaves. The measurements revealed a wide variation in overall susceptibility/resistance among 14 Arabidopsis ecotypes as well as showing interesting patterns of symptom distribution among consecutive leaves of a rosette or between the upper and lower side of a leaf, confirming the great potential of this protocol and AI in future research.

TSSM age‑synchronized colony

A TSSM age‑synchronized population was obtained according to Barczak-Brzyżek et al. (2017) with some modifications. Briefly, to synchronize the age of females, freshly detached leaves of three-week-old plants of the common bean(Phaseolus vulgaris L. cv. Ferrari, PNOS, Ożarów Mazowiecki, Poland) were placed on wet cotton in Petri dishes (150 mm in diameter) with abaxial (lower) side up and ten young females from the stock lab colony maintained on common bean plants (for more than 150 generations) were transferred to each of 10 detached leaves. The Petri dishes were placed in a growth chamber (SANYO Plant Growth Chambers MLR-350H) under controlled conditions (photoperiod — 16/8h day/night; light intensity: 150 µmol photons m^-2s^-1, relative humidity, RH - 65±10% and temp. — 23±1°C) to let the females lay eggs for 24 hours. Then the females were removed and when the larvae hatched they were moved to two-week-old common bean plants growing in the growth chamber to allow the females to develop synchronously.

TSSM infestation of A. thaliana ecotypes

Fourteen A.thaliana L. accessions, originally derived from ecotypes found and collected in various world locations, and referred to in this paper as ecotypes, were acquired from the Nottingham Arabidopsis Stock Centre (NASC; Nottingham, UK): Car-1 (NASC id: 76742), Cvi-0 (76789), Ms-0 (76555), Col-0 (76778), Ts-1 (76615), Got-7 (76495), Bur-0 (N6643), Rrs-7 (76593), Dor-10 (76806), Kondara (76532), Ler-0 (77020), Tamm-2 (76610), Can-0 (38909), Stp-0 (77283) (Table S1). The plants were sown on peat pellets (Jiffy, Netherlands), stratified for two days at 4°C and grown for three weeks in controlled conditions (16/8h (day/night) photoperiod; 150 µmol photons m^-2s^-1 light intensity, 70% RH and temp. 22°C).

Ten five-day-old mite females originating from the age-synchronized colony were transferred using a fine brush (diameter 1.8 mm) to the youngest leaves of the rosette of each A. thaliana experimental plant. Females had free choice to feed and lay eggs on each rosette leaf for 72 hours. Mite-infested A. thaliana plants were placed in the growth chamber in the same controlled conditions as described above.

Imaging of mite-infested A.thaliana leaves

All A. thaliana rosette leaves from a single mite-infested plant were cut using a scalpel, and arranged on a square Petri dish lid (120 x 120 mm) covered with double-sided transparent adhesive tape starting from the youngest (4 mm long leaf) to the oldest leaf excluding cotyledons.

The imaging of leaves differing in size and shape was performed using a specially configured Leica Thunder Imager System (Leica Microsystems Ltd., Switzerland) consisting of an M205 FA stereo microscope with a DFC7000 T camera and fully motorized XY-Scanning Stage (Märzhäuser Wetzlar GmbH & Co. KG, Germany). The acquisition software — Leica Application Suite X (LasX; Leica Microsystems CMS GmbH, Germany) was operated on a PC (HP Z4 G4 Workstation, Hewlett-Packard Company‎, Palo Alto CA, USA with a NIVIDA GeForce RTX 2080 Ti graphics card, Nvidia Corporation, Santa Clara CA, USA). The image acquisition was performed using reflected light (LED5000 High Diffuse Dome Illuminator; Leica Microsystems Ltd., Switzerland) under 25x magnification. Using a navigator tool from LasX software, the image acquisition area and focus were set manually. A full leaf scan included approx. 400 images of neighboring and partially overlapping pictures which were merged automatically after scanning with the imaging system software. The images of abaxially and adaxially scanned leaf surfaces were exported to TIFF format and stored on a QNAP TS431 disc array (QNAP Systems Inc., New Taipei City, Taiwan).

Image analysis

The analysis of high-resolution images of TSSM-infested A. thaliana leaves was performed by a computer program, MITESPOTTER (developed specifically for this project), detecting and determining the number of eggs laid by females, the area of feeding damage, and the amount (number and area) of fecal pellets. All graphical data managing and specific analyses including image cataloging, leaf measurement and indexing, object detection and quantification were done using MITESPOTTER. The program includes optional manual correction of the results and data export for downstream analyses.

Neural networks

The detection of eggs, leaf damage, and fecal pellets by the MITESPOTTER program was based on neural networks and required prior network training with manually marked items.

The identification of eggs and feces was treated as an object detection problem. In deep learning, object detection is finding instances within an image which fulfil the criteria of the given neural network. The bounding box that surrounds the items found is rectangular. Detected objects are counted and described by position, dimensions, and area. These tasks were done using the Detectron2 library created by Facebook AI Research (Wu et al., 2019).

To detect the feces, we used common architecture in object detection problems — Faster RCNN. The Faster RCNN is based on a two-stage algorithm. These types of models are built from two algorithms. The first of them finds candidate objects in the image. The second stage is the verification of every candidate and generating features like a bounding box for each candidate. In the Faster RCNN (Ren et al., 2015), in the first stage, we used Resnet50 (He et al., 2016) to generate the proposal regions. The rest of the network architecture is standard for this model. The eggs were detected similarly using the Faster RCNN approach.

The identification of feeding damage was treated as a semantic segmentation task. Image segmentation is the task of classifying each pixel. This means assigning each pixel to one category from a defining set. Every pixel on our image may be classified as a symptom area or not. This allows thedetermination of the size, shape, and area of the objects of interest. This problem was solved by Feature Pyramid Network (FPN) (Lin et al., 2017). This network represents the encoder-decoder type of architecture. In these models, the encoder network generates an abstract representation of an image. In the second step, the decoder network from this representation creates a fine output. As an encoder network, we used NFNet L1 with Efficient Channel Attention (Brock et al., 2021).

To train the neural network to find areas of feeding damage we used a dataset containing 20 images of leaves. Each of the images was divided into patches of size 128 x 128 pixels. Before becoming an input to a model, all images underwent standardization and a five-fold dilation intended to expand the masks. The dataset was divided into train, validation, and test sets with an 80-10-10 ratio.

Statistical Analysis

To evaluate the significance of differences (at p < 0.05) in the number of eggs laid by a female in three days, the abundance of feeding symptoms and feces between 14 A. thaliana ecotypes, a one-way ANOVA test was performed using R language and R Studio (R Core Team, 2021).

Procedure description

To streamline the process of discovering natural resistance of A. thaliana ecotypes to TSSM, we developed a semi-automated protocol which collects and measures unbiased data on female oviposition rate, the area of feeding symptoms and the area of fecal pellets. Our procedure consists of microscope scanning, producing high-resolution images of the leaf surface and image analysis with the specially developed AI-powered computer program MITESPOTTER. The protocol consists of manual and automated steps optimized to handle a task to screen approx. 1,000 mite-infested A. thaliana plants within one vegetation season engaging two skilled technicians.

The procedure is outlined in Fig. 1 and presented in the 10 steps described below. Since the objects to be identified and measured should be recorded in a resolution sufficient to distinguish as many details as possible, we decided to use microscope scanning giving a resolution of 8500 dpi. As a result, the initial image of max. 10 x 10 cm could be zoomed in to clearly distinguish the required details – eggs, irregular feeding symptoms, feces (Fig. 2) and many others. The high-resolution microscopic scans were done with the same settings for the whole experiment and successfully subjected to automatic analysis with the specially-developed computer program – MITESPOTTER. The MITESPOTTER consists of the following modules: (1) image manager – cataloging full-size microscopic scans and decoding leaf side and accession, (2) leaf identifier – cataloging, numbering measuring surface and checking the abaxial/adaxial overlap of individual leaves, (3) object detection – segmenting leaf images, detection of eggs, feeding symptoms and feces by neural-network-based algorithms and marking them on unsegmented images, (4) manual verification, and (5) data export for downstream analyses.

The main steps of the procedure

A. Preparing the material for imaging

Step 1. Stick two 10 cm long (5 cm wide) pieces of double-sided transparent adhesive tape in parallel to the lid of a square 12 cm plastic Petri dish creating approx. 10 cm x 10 cm area for placing the leaves from one rosette (Fig. 3A). IMPORTANT! Use a plastic squeegee for better attachment of the adhesive tape and reduction of air bubbles because many of them may resemble mite eggs in obtained images of the adaxial side of the leaf.

Step 2. Cut off the leaves from the rosette starting from the youngest having a length of minimum 4 mm (smaller leaves are difficult to handle and are not subject of TSSM attack due to very dense trichome coating). Follow the natural phyllotaxy of a given rosette and end with the oldest (first leaf appearing above cotyledons; Fig. 3B). Place the leaves on the tape with abaxial side up using bare fingers (latex or nitrile examination gloves and other hydrophobic intermediate layers for leaf contact caused detachment of numerous eggs from the abaxial leaf surface; Fig. 3C). IMPORTANT! Leaves should be prepared for scanning shortly before the imaging to prevent the tape from misting up with humidity. It's crucial to keep a similar arrangement of leaves on the plates for each rosette: start from the left of the upper row and continue from the left of the lower row leaving a few mm clear spacing between leaves to allow automatic leaf detection and numbering during analysis.

B. Imaging

Step 3. Place the Petri dish with arranged leaves on the microscope stage with the leaf’s abaxial side up. Use a navigator (LasX program) to define the area that specifies the plate scan field, manually set the focus and start imaging by choosing the START button. When the scanning finishes, the merging process starts automatically (this option should be activated in the settings). Then turn over the Petri dish to record image of the adaxial side of the leaf and run scanning similarly. IMPORTANT! To minimize uncontrolled Petri dish displacements due to inertia during stage XY movements, a specific plate adaptor was made which additionally keeps the focus plane on the leaf when the Petri dish is turned over for imaging of the adaxial side of the leaf (Fig. 3D).

Step 4. Export the obtained image to a TIFF file by choosing the EXPORT IMAGE option in the LasX program menu. A new window will open where the file name and destination folder should be specified. IMPORTANT! The initial MITESPOTTER program configuration requires the file name to be created in such a way that it can identify the A. thaliana ecotype. The file name should contain information about: abaxial/adaxial side of leaf scanned, a given ecotype consecutive number and biological replicate, total number of leaves on the plate (e.g. D_A_1_I_2_K1_12). The obtained image can also be saved in the default LIF file format (Leica file) which can be opened with the LasX program and exported later to other graphical file formats. The imaging of one side of the plate produces files 3-4 GB in size therefore sufficient data storage capacity and PC performance should be ensured.

C. Image analysis

Step 5. Open the MITESPOTTER program and choose the RELOAD button in the File list box (Fig. 4A) to reload the file list (the file list is refreshed on demand or once a day due to image database size). IMPORTANT! The default automatic update of the file list is every day at 12 am.

Step 6. When the update is done (in our case 1 hour), choose image for analysis and click the RUN button in Analysis box (Fig. 4E) to perform the analysis of an unanalyzed pair of images of one plant (scans of abaxial and adaxial sides of leaf ).

Step 7. After the analysis is done, select the CSV button in the File list box (Fig. 4A) to export the data to a csv file. If verification of image analysis is needed, skip this step and go to section D (we recommend brief visual inspection of all scans after automatic analysis; specific cases when manual correction might be needed were described in the IMPORTANT note in section D). IMPORTANT! The csv file contains columns with: the name of the ecotype, the subsequent number of the plant (biological replicate), leaf side, and results for each side of a leaf separately, leaf area, area of feeding symptoms, number of eggs, number and area of feces.

D. Analysis verification (optional)

Step 8. After the MITESPOTTER analysis is done, a manual correction can be made. Select an image from the “File list” and press the VIEW button in “Analysis box” (Fig. 4E). A new window will open (Fig. 5A-E). IMPORTANT! Visual inspection of MITESPOTTER results could be helpful when images were analyzed incorrectly in the following cases: extremely diverse leaf phenotype of a given ecotype or mutant, local loss of focus due to leaf curvature, significant egg shedding and dispersal outside the leaf area, significant misinterpretation of air bubbles as eggs or omission of eggs, the presence of chlorosis/necrosis not resulting from mite activity, etc. During verification, the significance threshold level can be changed, for feces — within a range from 0.5 to 0.95, and for eggs — from 0.8 to 0.95 (Fig. 5B). The EXPANDED VIEW button next to a given leaf miniature icon (Fig. 5E) leads to the final high-resolution image of a leaf (adaxial or abaxial view) with detected objects visible on demand by pressing the buttons SHOW DAMAGES, SHOW EGGS or SHOW FECES (Fig.6A-D).

Step 9. To remove misdiagnosed eggs or feces from all detected ones, double click on the square marking an object and the program will ask if this is an object to be removed. To add unrecognized eggs draw a square with the mouse select tool. The changes should be confirmed by pressing the SAVE button (Fig. 6B). Use arrow buttons to navigate between high-resolution images of leaves and their sides of a given plant (Fig. 6C).

Step. 10. To export the data, go back using the BACK button in the right upper corner (Fig. 6A and Fig. 5A) to the start screen and follow step 7.

Efficiency of MITESPOTTER detection

The effectiveness of object detection can be measured as a percentage of correctly detected objects from all objects existing in the image. The simplest method of measuring the correctness of semantic segmentation was ECA NFNet L2, which counts the pixel accuracy (the percentage of well-matched pixels). The percentage of correct pixels for traits in our program was: 99.33% for the abaxial side of the leaf and 99.87% for the adaxial side of the leaf.

A. thaliana susceptibility to TSSM measured by MITESPOTTER

In this study, the susceptibility of 14 A. thaliana ecotypes to TSSM was determined based on three traits – the area of mite feeding symptoms, female oviposition rate and feces area measured by the MITESPOTTER program. With this program we could have avoided splitting the large number of experimental plants into groups to be analyzed at one time. However, such a scenario (splitting a large tested population) is still realistic if over 1,000 plants are to be analyzed. Then seasonal changes may influence the results even though the initial rearing of the T. urticeae population, age-synchronization of females and infestation were done in stable temperature, light and humidity. Therefore, the presented measurements were made in three repetitions at six-week intervals and a reference ecotype Col-0 was added to every group of plants tested.

The results were presented as absolute and relative values. Among the examined ecotypes, the least susceptible was Car-1 irrespective of whether absolute or relative values were considered in each of three susceptibility determinants. When the absolute and relative trait values were considered, the most susceptible accessions were Tamm-2 and Stp-0, respectively (Fig. 7A-D). This was true for the area of feeding symptoms and oviposition rate, whereas in the case of feces area the comparison of absolute values failed to indicate with statistical significance a single most susceptible ecotype (Fig. 7E-F). The detailed lineup of tested accessions also varied slightly depending on whether absolute or relative values were considered and susceptibility determinants used. In summary, susceptibility based on the criterion of area of relative feeding symptoms varied more than 18-fold between outermost A. thaliana ecotypes. However, for the oviposition rate, the dynamic range was around 8-fold, whereas for the feces area more than 3-fold (Table S2).

Distribution of TSSM symptoms on A. thaliana leaves

The experimental design used for this study and the MITESPOTTER computer program allow an examination of the A. thaliana ecotype’s susceptibility to TSSM based on the distribution of mite feeding symptoms, female oviposition rate and feces area on consecutive rosette leaves (Fig. 8) as well as showing the proportion of eggs and feces on abaxial and adaxial leaf surfaces (Fig. 9). Analysis of the distribution of these traits on consecutive leaves sometimes showed repetitive large-scale symptoms of all three on some and very little incidence on others (Fig. 8). Such a pattern was not conserved for all individual plants of the same ecotype but was visible for approximately 50% of plants where repetitive peaks of mite activity were seen (Fig. S1). For all tested plants and ecotypes there were no TSSM infestation symptoms on the youngest leaves, which may be a result of dense trichomes forming a physical barrier for females. The mature, fully grown and expanded leaves did not form trichomes de novo and their density even on the adaxial side of the leaf did not interfere with female activity.

The ratio of eggs on the adaxial/abaxial side of leaves was mainly around 1:5, which corresponds with TSSM’s nature of staying mostly on the abaxial side of the leaf. However, in the case of two ecotypes (Dor-10 and Ms-0), around 60% of eggs were laid on the adaxial side and in the case of two other ecotypes(Col-0 and Bur-0) around 40% of eggs were found on the adaxial side (Fig. 9A). The percentage of feces detected on the adaxial side ranged from 41% to 73% (Fig. 9B).

Approaches to monitoring plant-spider mite interaction

Studying the interactions between herbivorous pests and host plants may be a starting point of research focusing on the herbivore side (e.g. adaptation) or the host plant (e.g. defense mechanisms). Such research often requires a tedious inspection and classification of pest population structure and dynamics as well as the damage symptoms on hundreds of plants creating a bottleneck in the availability of qualified experts. In particular, sustainable agriculture in the era of climate warming requires an acceleration of resistance breeding based on well-described molecular mechanisms, which is difficult to accomplish without precise and reproducible assessment of genetically and phenotypically polymorphic populations of a given plant species allowing individual classification as susceptible or resistant (Barrett et al., 2009).

Plant resistance-related traits can be defined on the basis of pest performance characteristics, for example: number of individuals (or its DNA content), population structure, oviposition rate, developmental time (kri98) or by the quantification of their effects on host plants e.g. chlorophyll loss (Grime, 1998), marker gene expression, content of plant’s secondary metabolites (Goggin et al., 2015; Sharma, 2009), biomass loss, proportion of tissue lesions or necrosis and other feeding symptoms.

In the case of plant-spider mite interactions, modern methods to assess infestation are often based on analysis of visible or multispectral images recorded by rather low-resolution cameras and subsequent calculation of vegetation indices (e.g. Herrman et al., 2012; Uygun et al., 2020). Such methods allow correlation of observed symptoms to spider mite infestation and can sometimes distinguish the effects of the activity of two different pest species and plant malnutrition symptoms (Gonzalez-Gonzalez et al., 2021). Such methods are ideal for the monitoring of a given species on a given plant variety for the optimization of crop protection, however still lack robustness in terms of natural variability of pest races and developmental stages or host genotype (cultivars, breeding lines, segregating progenies, ecotypes, mutants). New developments in neural network models provide some hope for improving this especially for images recorded in variable light conditions (Crocket et al., 2014). Obviously, these methods are still lacking the versatility of a human expert and are rather limited to a particular task and budget. An interesting and smart tradeoff of cost, throughput and expert engagement without sophisticated equipment was shown by Cazaux et al. (2014), who developed a method to assess the susceptibility of A. thaliana to TSSM by measuring damage by a semi-automated selection of the altered area on a fine square grid over leaf images generated by an office scanner. The protocol provided here fully automatizes this process, requiring only 1,000 examples of the trait’s images for neural network training.

Variation of feeding symptoms

The presented protocol, which includes microscope scanning of TSSM-infested leaves and AI-aided image analysis using the MITESPOTTER program, allows large amounts of data to be obtained regarding plants’ resistance to TSSM. It accelerates the assessment, standardizes the criteria, measures precisely, and reduces the engagement time of experienced specialists. Moreover, the data can be stored and reanalyzed and the adaptation of the protocol to other pest-host plant models is not complicated.

During the optimization of the protocol, several problems were solved. For example, the recorded leaf damage, eggs and fecal pellets, differed between adaxial and abaxial scans. This was due to the fact that the abaxial side was photographed directly whereas the imaging of the adaxial side passed through the plastic plate and adhesive tape causing different light scattering. Moreover, the adhesive tape added noise disturbing the detection of eggs due to small air bubbles present in the tape glue, which sometimes could be mistaken for eggs. Some additional noise may also originate from leaves stuck unevenly to the tape forming closed spaces which could quickly steam up by transpiration disturbing subsequent imaging and detection.

Consequently, for neural network training separate, abaxial and adaxial training sets were prepared for eggs and leaf damage. In the case of eggs, another image-diversifying factor was the egg age, which could vary from 0 to 3 days. So the training set should contain transparent, straw yellow, and orange eggs. In the case of mite feeding symptoms, their shape and area also varied depending on the side of the leaf and overlapped only partially. This could depend on which mesophyll cells were damaged (sponge, palisade or both). Therefore, in the MITESPOTTER configuration, we decided to conduct damage detection of the two sides separately. Then the area was summarized and plotted on both sides in the same shape (the overlapping portion was not duplicated). Another problem concerning mainly the adaxial side of the leaves was caused by trichomes, which often obscured the objects to be detected. Admittedly, the trichomes interfered with object detection on the youngest leaves with the highest trichome density. Fortunately, the trichome density was also an efficient physical barrier causing these leaves to be unwillingly chosen by mites to feed and lay eggs on. On the older, more expanded leaves, the chosen neural network model efficiently ignored the trichomes. Also, the fecal pellets varied in appearance. Both small, black, single pellets and larger clusters were efficiently detected by MITESPOTTER, however, the detection of white guanine feces would require more optimization and therefore they were excluded from this study.

MITESPOTTER manual adjustments

The above listed possible problems with egg color variation and air bubbles necessitated the use of much larger training sets for neural networks and prompted us to include a function to remove manually misidentified eggs during the optional verification step in the MITESPOTTER program (Fig. 6). This option could also be used in the case of misdetection of feces due to dirt particles or residues of dead adults (usually they were excluded automatically) or unusual leaf color variation and trichome density. However, the possibility of leaf color variation would better be compensated for by adjusting the significance threshold levels for the detection of eggs and feces. In our opinion, the default and optimal levels were 0.8 and 0.9 for eggs and feces, respectively. The feeding symptoms were recognized at the unchangeable significance level of 0.5. These settings allow for efficient analysis even of morphologically differentiated A. thaliana ecotypes, but extending the application of MITESPOTTER to other plant species may require more manual optimization (Fig. 10).

Improving the procedure throughput

The presented protocol allows for the assessment of the activity of TSSM on hundreds of plants. There is the potential to optimize it for other plant-pest interactions (Fig. 10). We are also aware that several bottlenecks exist and further optimization would increase the protocol’s performance. Below are a few issues where we see a realistic chance for improvements.

1. Infestation – in our protocol we arduously applied 10 young females from a synchronized population per plant, but there are a few alternative options (the trade-off is that they give a new source of variability within the experiment):

Massive infestation without synchronization or selection – all stages of larvae and adults are transferred to plants e.g. by brush (for a smaller number of mites per plant) or by pump for more than 20 females per plant (Cazaux et al., 2014; Zhurov et al., 2014). Using this type of infestation, the measured resistance/susceptibility may be more variable because (a) not every female will be at the peak of its fecundity, (b) larvae, nymphs and males do not lay eggs and (c) the structure of the population used may vary.
Free migration – “free choice” scenario – plants are put into the stock colony of TSSM (where mites of all stages of their life cycle are present) or overpopulated leaves from the stock colony are placed on top of experimental plants. Mites have the freedom to migrate, which depends on the maternal population host species/ecotype and the colony density, as well as physical proximity/contact of leaves. After the experiment, the symptoms can be measured as in the presented protocol or the number of mites can be assessed by brushing out the mites to double-sided adhesive tape and scanning it using a similar imaging system as above.

2. Fewer leaves to be analyzed – in the case of A. thaliana, plants a few days younger could allow all leaves to fit in one row, reducing the time required for specimen preparation and scanning. Different growth intensities of A. thaliana ecotypes should be considered as well as the possible susceptibility discrepancy between younger and older seedlings. Such optimization would not be possible with other species having larger leaf sizes and numbers where specific sampling protocols should be elaborated to provide statistically significant results.

3. Image quality – the full 10 x 10 cm area scan is composed of approx. 400 merged camera shots. A higher resolution of the camera and a larger field of view of the microscope optics would allow for a reduction in the number of pictures taken from each specimen.

4. Image size – the uncompressed image exceeds 3 GB, implying quite high requirements with respect to storage capacity and computer performance for downstream analyses. Considering progress in computer hardware engineering, we expect gradual improvement – reduction of computing time.

Example study – trait variability among A. thaliana ecotypes

Trait variability among the 14 ecotypes tested here showed the dynamic range of 18.4, 7.9 and 3.2 for relative values of the area of feeding symptoms, oviposition rate and feces area, respectively. The values obtained suggest polygenic determination of these traits similarly to the results obtained by Zhurov et al. (2014) where susceptibility was assessed by measurement of the damaged leaf area. Besides a different method of measurements in our experiments and the cited report (MITESPOTTER vs. manual area selection; three days after infestation vs. four; three-to-five-day old unadapted and synchronized females vs. unadapted females manually selected from unsynchronized population) the most and least susceptible A. thaliana accession show similar dynamic ranges of the damaged area. The three characteristics presented here, describing the susceptibility of 14 A. thaliana ecotypes cumulated over three days, result from very complex biological processes suggesting that a diverse genetic repertoire may be responsible for its measured value. Moreover, we used synchronized but unadapted females. Despite these factors, the area of feeding symptoms, female fecundity and feces area indicate the same ecotypes as the most and least susceptible (the detailed rank of other ecotypes differ slightly depending on the trait measured) indicating the existence of a general antixenosis/antibiosis mechanism in the T. urticae–A. thaliana interaction. Deciphering the molecular background of this variation is of the highest importance both for basic research and plant breeding. Another highly desired follow-up research would be addressing the question to what extent the presented ecotype variation would be reduced if the host adaptation step is used (Magalhaes et al., 2007; Grbic et al., 2011; Dermauw et al., 2013).

Example study – Distribution of TSSM feeding symptoms on leaves

The described method allows measurement of the symptoms on each leaf individually, sometimes showing peaks and valleys on the chart (Fig. 8). In natural conditions the mite population tends to colonize plants with an aggregated distribution pattern. This may be due to unevenly spread resources, varying environment, less species-specific competition and negative antagonism (Kumaran, 2011). In our experiments, these factors are rather negligible due to the low density of synchronized females as the only herbivore on the tested plants, however we see great potential for using our method in research on mite population dispersion within and between plants. On the other hand, the observed patterns of symptom distribution within one plant may be associated with parastichy-dependant distribution of systemic acquired resistance (SAR) signals (Roberts et al., 2007; Ferrieri et al. 2015; Heyer et al., 2018). Parastichy in Arabidopsis describes direct or indirect vascular connections of leaves formed during development and depends on the orientation of phyllotaxy and the rosette size and can follow the n+3, n+5 or n+8 pattern (Dengler, 2006). The best results in the cited literature, however, were obtained when only one leaf was infected/infested and SAR distribution was monitored by marker gene expression. In the case of our experiments, the mite migration and activity did not produce convincing parastichy patterns, suggesting that rosette topology (mainly the shoot apex where young females were applied) and random leaf contacts within the rosette caused unequal distribution of females and symptoms. This does not exclude the SAR influence on representatives of the same or different species (Kielkiewicz et al., 2019) but the experimental setup applied made it impossible to observe it. The individual measurements of leaf symptoms by the MITESOPTTER program may also be useful in other experiments where e.g. plant canopy distribution of pests or pathogens is important.

An interesting part of our results was also the observation that there is variability in the distribution of traits depending on the leaf side (number of eggs and feces area; the damaged area detected largely overlaps on the abaxial and adaxial side). A couple of factors may contribute to this variation but they are usually related to the environment or species community cohabiting the same plant. Assuming there was no environmental variability (growth chamber) the innate physical, chemical or molecular factors of the host plant contributed to this variation and this is another fascinating area to explore.

Downstream applications of collected data

Forward genetics, i.e. linking a specific phenotypic trait to single or multiple genes, is crucial for plant breeding and basic research. The most precise approaches require screening of a trait variation in extensively polymorphic and large populations. The advent of relatively low cost Next Generation Sequencing methods has provided an incredible source of genetic polymorphism data, however, the phenotypic trait assessment and quantification are still difficult to simplify and accelerate. The presented protocol is designed to collect data on phenotype variation with a specific focus on traits related to TSSM susceptibility of A. thaliana ecotypes where hundreds of them are fully sequenced. Further, such data can be used in association mapping to identify plant genes or genetic markers to be used in TSSM resistance breeding. Such experiments, however, would require screening of mite susceptibility of many more than 14 ecotypes or segregating progeny (100-200) (Ristova et.al., 2018).

We developed a new AI-powered computer program – MITESPOTTER – which detects and measures TSSM feeding symptoms, the number of eggs and area of feces on all rosette leaves of 14 A. thaliana ecotypes/accessions at one time. We implemented this program into the plant susceptibility/resistance assessment protocol based on high-resolution microscope imaging of both leaf surfaces infested by TSSM. This protocol can be adapted to other small pest-host plant interactions. Initial results show a great potential for our method in studying the distribution of mite-pest both among potential host plants and/or within the plant canopy. Specifically, the screening of large genetically diverse plant populations with the analytical pipeline described here can accelerate the identification of genes contributing to increased pest resistance.

Author Contribution Statement

MF and MK formulated the hypothesis and conceived the research plan. MF, MK and KM supervised the experiments. EZ, AW, MK, KS, PD and ABB performed the experiments. EZ, AW and MF analyzed the results. EZ and AW prepared the figures. MF wrote the manuscript with the contribution of all authors. All authors read and approved the manuscript.

Acknowledgments

We would like to acknowledge Hanna Załęska for maintaining the TSSM stock culture with meticulous care, and Ewa Siedlecka for her dedication in tending to the plants used in this research. We are also grateful to Stefan Smoleński for his perseverance in maintaining the phytotron facility despite numerous breakdowns. Additionally, we extend our gratitude to the team at KAWA.SKA company for providing an individually assembled Thunder Imager demo, which proved essential in collecting preliminary results for our funding efforts. We sincerely appreciate the support of the Polish National Science Centre for providing the funding (grant no. 2019/33/B/NZ9/01305) that supported this research.

Barczak-Brzyżek, A., Kiełkiewicz, M., Górecka, M., Kot, K., Karpińska, B., & Filipecki, M. (2017). Abscisic Acid Insensitive 4 transcription factor is an important player in the response of Arabidopsis thaliana to two-spotted spider mite (Tetranychus urticae) feeding. Experimental and Applied Acarology, 73(3), 317-326.
Barrett, L.G., Kniskern, J.M., Bodenhausen, N., Zhang, W. and Bergelson, J. (2009), Continua of specificity and virulence in plant host–pathogen interactions: causes and consequences. New Phytologist, 183: 513-529.
Brock, A., De, S., Smith, S. L., & Simonyan, K. (2021). High-performance large-scale image recognition without normalization. In International Conference on Machine Learning (pp. 1059-1071). PMLR.
Brust, G. E., & Gotoh, T. (2018). Mites: biology, ecology, and management. In Sustainable Management of Arthropod Pests of Tomato (pp. 111-130). Academic Press.
Cazaux, M., Navarro, M., Bruinsma, K. A., Zhurov, V., Negrave, T., Van Leeuwen, T., Grbic, V. & Grbic, M. (2014). Application of two-spotted spider mite Tetranychus urticae for plant-pest interaction studies. Journal of visualized experiments: JoVE, (89).
Crockett, C., Liburd, O., & Abd-Elrahman, A. (2014). Development of image based detection methods for two-spotted spider mite, Tetranychus urticae Koch, on strawberries. In Proceedings of the Entomological Society of America Annual Meeting.
Dengler, N. G. (2006). The shoot apical meristem and development of vascular architecture. Botany, 84(11), 1660-1671.
Dermauw, W., Wybouw, N., Rombauts, S., Menten, B., Vontas, J., Grbic, M., Clark, R.M., Feyereisen, R. & Van Leeuwen, T. (2013). A link between host plant adaptation and pesticide resistance in the polyphagous spider mite Tetranychus urticae. Proceedings of the National Academy of Sciences, 110(2), E113-E122.
Ferrieri, A. P., Appel, H. M., & Schultz, J. C. (2015). Plant vascular architecture determines the pattern of herbivore-induced systemic responses in Arabidopsis thaliana. PLoS One, 10(4), e0123899.
Fraulo, A. B., Cohen, M., & Liburd, O. E. (2009). Visible/near infrared reflectance (VNIR) spectroscopy for detecting two-spotted spider mite (Acari: Tetranychidae) damage in strawberries. Environmental entomology, 38(1), 137-142.
Goggin, F. L., Lorence, A., & Topp, C. N. (2015). Applying high-throughput phenotyping to plant–insect interactions: picturing more resistant crops. Current Opinion in Insect Science, 9, 69-76.
Gonzalez-Gonzalez, M. G., Blasco, J., Cubero, S., & Chueca, P. (2021). Automated Detection of Tetranychus urticae Koch in Citrus Leaves Based on Colour and VIS/NIR Hyperspectral Imaging. Agronomy, 11(5), 1002.
Gotoh, T., & Kubota, M. (1997). Population dynamics of the citrus red mite, Panonychus citri (McGregor)(Acari: Tetranychidae) in Japanese pear orchards. Experimental & Applied Acarology, 21, 343-356.
Grbic, M., Van Leeuwen, T., Clark, R. M., Rombauts, S., Rouzé, P., Grbic, V., ... & Van de Peer, Y. (2011). The genome of Tetranychus urticae reveals herbivorous pest adaptations. Nature, 479(7374), 487-492.
Grime, J. P. (1998). Benefits of plant diversity to ecosystems: immediate, filter and founder effects. Journal of Ecology, 86(6), 902-910.
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
Helle, W., & Sabelis, M. W. (Eds.). (1985). Spider mites: their biology, natural enemies and control (Vol. 1, pp. 141-160). Amsterdam: Elsevier.
Herrmann, I., Berenstein, M., Sade, A., Karnieli, A., Bonfil, D. J., & Weintraub, P. G. (2012). Spectral monitoring of two-spotted spider mite damage to pepper leaves. Remote Sensing Letters, 3(4), 277-283.
Herrmann, I., Berenstein, M., Paz-Kagan, T., Sade, A., & Karnieli, A. (2015). Early detection of two-spotted spider mite damage to pepper leaves by spectral means. In Precision agriculture'15 (pp. 64-70). Wageningen Academic Publishers.
Heyer, M., Reichelt, M., & Mithöfer, A. (2018). A holistic approach to analyze systemic jasmonate accumulation in individual leaves of Arabidopsis rosettes upon wounding. Frontiers in Plant Science, 9, 1569.
Kiełkiewicz, M., Barczak-Brzyżek, A., Karpińska, B., & Filipecki, M. (2019). Unravelling the complexity of plant defense induced by a simultaneous and sequential mite and aphid Infestation. International journal of molecular sciences, 20(4), 806.
Kloth, K. J., Thoen, M. P., Bouwmeester, H. J., Jongsma, M. A., & Dicke, M. (2012). Association mapping of plant resistance to insects. Trends in plant science, 17(5), 311-319.
Krips, O. E., Witul, A., Willems, P. E. L., & Dicke, M. (1998). Intrinsic rate of population increase of the spider mite Tetranychus urticae on the ornamental crop gerbera: intraspecific variation in host plant and herbivore. Entomologia experimentalis et Applicata, 89(2), 159-168.
Kumaran, N. (2011). Within-plant and within-leaf dispersion pattern of two-spotted spider mite, Tetranychus urticae Koch (Acari: Tetranychidae) on okra. Archives of Phytopathology and Plant Protection, 44(20), 1949-1957.
Laing, J. E. (1969). Life history and life table of Tetranychus urticae. Acarologia, 11(1), 32-42.
Li, C., Williams, M. M., Loh, Y. T., Lee, G. I., & Howe, G. A. (2002). Resistance of cultivated tomato to cell content-feeding herbivores is regulated by the octadecanoid-signaling pathway. Plant Physiology, 130(1), 494-503.
Li, L., Zhang, Q., & Huang, D. (2014). A review of imaging techniques for plant phenotyping. Sensors, 14(11), 20078-20111.
Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., & Belongie, S. (2017). Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2117-2125).
Lucini, T., Faria, M. V., Rohde, C., Resende, J. T. V., & de Oliveira, J. R. F. (2015). Acylsugar and the role of trichomes in tomato genotypes resistance to Tetranychus urticae. Arthropod-Plant Interactions, 9, 45-53.
Magalhaes, S., Fayard, J., Janssen, A., Carbonell, D., & Olivieri, I. (2007). Adaptation in a spider mite population after long‐term evolution on a single host plant. Journal of evolutionary biology, 20(5), 2016-2027.
Martin, D. E., & Latheef, M. A. (2018). Active optical sensor assessment of spider mite damage on greenhouse beans and cotton. Experimental and Applied Acarology, 74, 147-158.
Migeon, A., & Dorkeld, F. (2018). Spider Mites Web: A comprehensive database for the Tetranychidae
Nieuwenhuizen, A. T., Kool, J., Suh, H. K., & Hemming, J. (2019). Automated spider mite damage detection on tomato leaves in greenhouses. In XI International Symposium on Protected Cultivation in Mild Winter Climates and I International Symposium on Nettings and 1268 (pp. 165-172).
Park, Y. L., & Lee, J. H. (2002). Leaf cell and tissue damage of cucumber caused by two-spotted spider mite (Acari: Tetranychidae). Journal of Economic Entomology, 95(5), 952-957.
R Core Team (2021). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/.
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28.
Rioja, C., Zhurov, V., Bruinsma, K., Grbic, M., & Grbic, V. (2017). Plant-herbivore interactions: a case of an extreme generalist, the two-spotted spider mite Tetranychus urticae. Molecular Plant-Microbe Interactions, 30(12), 935-945.
Ristova, D., Giovannetti, M., Metesch, K., & Busch, W. (2018). Natural genetic variation shapes root system responses to phytohormones in Arabidopsis. The Plant Journal, 96(2), 468-481.
Roberts, K., Love, A. J., Laval, V., Laird, J., Tomos, A. D., Hooks, M. A., & Milner, J. J. (2007). Long‐distance movement of Cauliflower mosaic virus and host defence responses in Arabidopsis follow a predictable pattern that is determined by the leaf orthostichy. New Phytologist, 175(4), 707-717.
Santamaría, M. E., Arnaiz, A., Velasco-Arroyo, B., Grbic, V., Diaz, I., & Martinez, M. (2018). Arabidopsis response to the spider mite Tetranychus urticae depends on the regulation of reactive oxygen species homeostasis. Scientific reports, 8(1), 1-13.
Sharma, A., Chandraker, S., Patel, V. K., & Ramteke, P. (2009). Antibacterial activity of medicinal plants against pathogens causing complicated urinary tract infections. Indian journal of pharmaceutical sciences, 71(2), 136.
Shih, C. I. T., Poe, S. L., & Cromroy, H. L. (1976). Biology, life table, and intrinsic rate of increase of Tetranychus urticae. Annals of the Entomological Society of America, 69(2), 362-364.
Tehri, K. (2014). A review on reproductive strategies in two spotted spider mite, Tetranychus urticae Koch 1836 (Acari: Tetranychidae). Journal of Entomology and Zoology Studies, 2(5), 35-39.
Uygun, T., Ozguven, M. M., & Yanar, D. (2020). A new approach to monitor and assess the damage caused by two-spotted spider mite. Experimental and Applied Acarology, 82, 335-346.
Van de Vrie, M., McMurtry, J., & Huffaker, C. (1972). Ecology of tetranychid mites and their natural enemies: A review: III. Biology, ecology, and pest status, and host-plant relations of tetranychids. Hilgardia, 41(13), 343-432.
Van Leeuwen, T., Vontas, J., Tsagkarakou, A., Dermauw, W., & Tirry, L. (2010). Acaricide resistance mechanisms in the two-spotted spider mite Tetranychus urticae and other important Acari: a review. Insect biochemistry and molecular biology, 40(8), 563-572.
Walter, A., Liebisch, F., & Hund, A. (2015). Plant phenotyping: from bean weighing to image analysis. Plant methods, 11(1), 1-11.
Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., Girshick, R. (2019) Detectron2. https://github.com/facebookresearch/detectron2
Zhurov, V., Navarro, M., Bruinsma, K. A., Arbona, V., Santamaria, M. E., Cazaux, M., ... & Grbic, V. (2014). Reciprocal responses in the interaction between Arabidopsis and the cell-content-feeding chelicerate herbivore spider mite. Plant physiology, 164(1), 384-399.

No competing interests reported.

JPSsupplement.docx

Download PDF

Journal Publication

published 04 Apr, 2024

Read the published version in Scientific Reports →

Version 1

posted

You are reading this latest preprint version

Automated imaging coupled with AI-powered analysis accelerates the assessment of plant resistance to Tetranychus urticae

Status:

Journal Publication

Version 1

Abstract

Figures

Key Message

Introduction

Materials And Methods

Results

Discussion

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1