Annotated genes were also classified into metabolic and biosynthetic pathways. A total of 7, genes were characterized and assigned to specific pathways. Because many genes figure in multiple biosynthetic or metabolic pathways, a total of 17, enzymatic nodes were described in all pathways. Among the largest metabolic pathways each involving more than 1, genes were purine metabolism and starch and sucrose metabolism.
The 20 largest biosynthetic pathways are summarized in Table 3. Table S6 presents a complete list of genes assigned to specific metabolic and biosynthetic pathways. Each enzyme was also assigned to one of the main enzyme classes Figure 3. In red clover, almost one-half of enzymes belong to transferases Table 3. Twenty largest biosynthetic and metabolic pathways in red clover based on number of genes enzymes involved.
A TBLASTX search was performed to evaluate the distribution of all red clover genes along recently published chromosomes of red clover, chromosomes of the model legume species M. The results were plotted using a window size of kb through genomic sequences Figure 4. Repetitive element and gene densities in each species were distributed along all chromosomes of T. Distribution patterns were similar in both T.
Under the specified criteria, 41, red clover genes were found to be homologs in comparison with M. These are regions with low density of repetitive elements, unlike centromeric regions Torales et al. This can be seen also in the central lines that show the distribution of homologous sequences to M. On the other hand, the gene densities in M. Centromeric regions were also poorer for homologous sequences, for example in chromosome Mt6 and Mt8. Figure 4. Comparison of gene densities and genome structure in legume model species A M.
The 7 T. Gene densities by kb windows are displayed on each chromosome as follows: 2 gene density in T. For those with sufficient flanking sequences, we designed appropriate unique primers to generate PCR product preferentially within — bp. The resulting 4, Because 1, When the total length of coding sequences Especially noteworthy is that no SSR markers were found in the genes belonging to the isoflavonoid biosynthetic pathway, such as 2-dihydroflavonol reductase, chalcone synthase, and isoflavone synthase. Figure 5. Statistics for 4, predicted SSR markers in coding sequences of red clover.
A Basic motif frequencies in coding regions of red clover.
- Related Articles?
- PROC ANOVA: One-Way Layout with Means Comparisons :: SAS/STAT(R) User's Guide, Second Edition;
- 514 (The Tin Train Book 3);
- Secrets and Promises.
B Frequencies of mono-, di-, tri-, tetra-, penta-, and hexamers plus complex SSR motifs in red clover genes containing one to six SSRs per locus. Total number of sequences containing depicted number of SSRs is shown above each column. As expected, the most frequently seen basic motif of microsatellite corresponded to trimeric repeat These motifs were also present mainly in loci with a single SSR marker.
Complex motifs consisted mainly of two—five trimeric motifs, with only 55 Other motifs, such as dimeric and pentameric, were seen much less frequently. Only in 7 SSRs with complex motifs did the complex motif not contain a trimeric repeat. SNPs were identified by aligning reads to predicted coding sequences. Of these SNPs, , SNP markers were also divided between transitions and transversions based on the nature of the A allele.
The majority nearly two-thirds of SNP markers were transitions. In addition, 4, 1. Table 4 presents a complete overview and statistics relating to SNP markers. Table S8 summarizes the complete list of SNP markers, including their positions and additional information. Altogether, 95 SSRs were analyzed for 50 red clover varieties. Figure 6. SSR marker validation in red clover varieties. A Numbers of SSRs amplified in analyzed varieties. B Numbers of varieties amplified for individual SSRs.
C Allele number distribution for SSR markers validated in red clover varieties. Allele number ranged from 1 to 17 Table S1, Figure 6. The similarity between individual varieties of T. Cluster analysis grouped the 50 red clover varieties into two clusters. Sub-cluster IA consisted of the single variety Radegast 4x , developed from landraces that were well adapted locally, from breeding varieties Slovensky podtatransky, Chlumecky, and Horal and a later cross with the variety Weitetra.
Sub-cluster IIA comprised a group of varieties whose genomes were enriched with genotypes of European origin: Dolina 4x , Vulkan 4x , and Sigord 4x were developed by crosses of Czech, Polish, and German varieties; Tabor 2x was developed by mass crosses of selected resistant plants belonging to 49 varieties; and Atlantis 4x is of German origin. Cluster IIB1 consisted of two varieties: Slavoj 2x was developed by the selection of genotypes, and Kvarta 4x; released was developed by polyploidy of landraces and the variety Chlumecky 2x , which itself was a component of the next sub-cluster IIB Chlumecky is the earliest red clover cultivar released developed by individual plant selection from the landrace Cesky.
Sprint 4x was obtained after the polyploidy of four newly bred genotypes of European origin. Figure 7. IIB was a large cluster of diploid and tetraploid Czech, European, and non-European varieties, reflecting that the varieties are often populations developed from genotypes with wide genetic variability, by targeted crosses, polycrosses, and topcrosses suitable for the selection of complex characters, with synteny of the selected genotypes. In addition, Pavo and Astur were released in Switzerland, Nemaro, and Titus in Germany, and Vesna is of Czech origin but developed by crosses of diploid genotypes from Czech, French, Swiss, and German varieties with subsequent polyploidy by colchicine.
Slovak and Swedish red clover genetic material was introgressed into the genome of Tatra, and Blizard was bred using non-Czech genotypes and recurrent phenotypic selection. Start, released in , was used as a component for more recently bred varieties such as Garant, Cyklon, Spur, Trubadur, Dolly, and Tempus. We examined intra-variety genetic heterogeneity using genome-wide SNP genotyping. The majority of those plants analyzed were heterozygous in the tested loci. All amplified fragments corresponded to those predicted. Figure 8. SNP marker validation in red clover individuals by temperature switch method.
The number of annotated genes in this study is higher than the number of genes identified recently by RNA sequencing 34, genes; Yates et al. On the basis of improved annotation of red clover genes, genes were classified into biosynthetic and metabolic pathways and key enzymes were identified. We have found 1, genes involved in purine metabolism, which is the fundamental pathway for plant growth and development Zrenner et al.
Original Research ARTICLE
This pathway is associated with DNA synthesis, energy sources, and synthesis of many primary and secondary metabolic products Stasolla et al. More than 1, genes are also involved in starch and sucrose metabolism, which is one of the most important pathways regarding energy sources in plants.
Moreover, with genes involved, biosynthesis of flavonoids is among the largest metabolic and biosynthetic pathways. These genes are of particular interest to red clover breeders inasmuch as flavonoid biosynthesis is associated with isoflavonoid content in the plant. Because they are known plant estrogen analogs, high levels of isoflavonoids are undesirable in forage varieties Adams, but these are required in varieties used in pharmaceutics Park and Weaver, Based on the distribution of homologous sequences among red clover, M.
In contrast to G. Occasional spikes in density of red clover genes show clusters of numerous gene families, such as genes of resistance. As previously described Kulikova et al. This was in contrast to red clover, where no such decrease was observed. The density of repetitive content along the chromosomes also was inspected and this provided very similar results. Orthologous loci were connected and compared among species visualizing syntenic loci and rearrangements of genome structure. During speciation, red clover clearly underwent complex genome restructuring, possibly associated with reduction of the basic chromosome number from eight to seven.
Results supporting this hypothesis were also found in a comparison among red clover, white clover T. Our results are supported, too, by a recently published paper regarding WGS of red clover and construction of its physical map De Vega et al. A slight discrepancies in locations of homologous sequences are likely the result of different methodology compared to previously published papers Isobe et al. DNA markers have a broad spectrum of uses in both research and practical breeding. They are used in QTL mapping Zhao et al. This is more than twice the number of SSR loci identified from red clover transcriptome sequencing Yates et al.
On the other hand, due to the lower number of genes and shorter length of coding sequence described by RNA sequencing, the average SSR marker frequency is very similar 1 SSR marker per When multiple repeat occurrences are taken into account, In coding sequences, clear domination of trimeric motifs Much lower frequencies were found for other motifs.
Similar results have been obtained also in other species, such as N. This phenomenon is very likely connected to the need to preserve open reading frame within the coding sequences and negative selection pressure against those SSR loci breaking it.
Even in the majority of SSR loci with non-trimeric basic motifs, therefore, the combination of motif length and its repeat number is divisible by three—e. SNPs were also identified in coding sequences of red clover. Nevertheless, when compared to other plant species, it is clear that SNP frequency is influenced by many factors, such as the number of individual plants analyzed in the study, natural variability in the population of the studied species, etc.
In these studies, the SNP number found correlated mainly with the number of individuals analyzed e. Although, just 16 individual plants of the same variety were analyzed in red clover, even higher SNP frequency was obtained, likely due to the outcrossing nature of clovers. Within the identified SNPs, transitions These results are consistent with those in P. A large number of SNPs are now available in red clover for genome-wide association studies and SNP microarray construction, where tens of thousands of markers are required. Outcrossing species populations are exceptionally variable and with a high level of heterozygosity.
The majority of genetic analyses of such species are necessarily carried out in pooled samples in order to collect most of the population variability and also minimize costs. Results obtained from such pooled samples are, however, unsuitable for estimating the copy number of individual alleles, which precludes assessment of exact allele frequencies required to calculate polymorphic information content PIC; Botstein et al.
Recent advances in NGS technologies enable determination of allele frequencies from pooled samples Mullen et al. PIC is commonly used in plant genetics to assess polymorphism level for a marker locus. Raveendar et al. Estimated PIC is usually directly connected with suitability for subsequent utilization, such as in variety identification or selection of suitable material for breeding purposes.
In order to calculate PIC, a precise determination of allelic frequencies in the studied population is required. To overcome the disadvantages of pooled samples, we proposed a modified PIC value termed pooled polymorphic information content pPIC which does not rely on determining allelic frequencies in the selected population. The presented pPIC of the 95 SSR markers analyzed should, however, be taken into account only to evaluate pooled samples similar in size to that of our study. A significant decrease or increase in pooled sample size could shift pPIC and thus degrade the estimation of SSR marker discrimination power.
Our results based on a pooled sample size of 16 plants should nevertheless be optimal for most potential subsequent utilizations. This is particularly important for screening gene bank accessions and large-scale analysis of cultivar identity and seed purity. For red clover, moreover, the optimal bulk size for genetic variation assessment among cultivars has been determined as 20 Kongkiatngam et al. All other SSRs revealed some polymorphism in the analyzed variety populations.
The breeding methods in red clover include procedures suitable for outcrossing crops.
Useful variation in a breeding population can be generated through hybridization and genome introgression, or by chromosome doubling polyploidy by colchicine. Subsequent phenotypic selection of superior individual plants or mass selection must be conducted on the progeny combining the best traits, and successive population breeding is performed. Molecular characterization of the analyzed varieties using SSRs reflects their genetic relationships, and the grouping is shown in Figure 7 and Figure S1.
Tracing the breeding history revealed frequent sharing and exchange of cultivars and newly bred materials among European breeding stations. The possible reasons could be i introgressions from landraces and ii that the varieties were mostly released from to , Chlumecky as early as with the exceptions of Atlantis and Slavoj. Narrowing of the genetic base in the more recent varieties in sub-cluster IIB was also apparent.
The SSR profiling i differentiated varieties with possible introgressions from landraces and ii indicated the existence of diversity at the molecular level among different red clover varieties. The finding of inter-variety heterogeneity has important consequences for breeders who use these varieties. Further progress in red clover breeding can be made by crosses with more distant genotypes as sources of new genetic variability, with new introgressions of important loci for resistance and quality.
The identification of SSR or SNP markers in known-function genes linked to specific traits can facilitate marker-assisted selection. One important task is to develop a platform for red clover genotyping, employing genome-wide distributed SNP markers. The Tatra-derived reference sequence was initially used for the detection of the predicted thousand SNPs. We used a preliminary set of 8, genome-wide distributed SNPs for polymorphism evaluation in individual plants. Arrayit methods provide universal microarray-based platforms for SNP genotyping Schena et al.
NGS methods such as genotyping-by-sequencing and the resequencing of targeted DNA regions from contrasting genotypes appear to be the most essential for SNP discovery and genotyping applications in red clover breeding. Temperature switch PCR can be successfully used in diagnostic applications through single-marker SNP genotyping for targeted coding sequences and for heterozygosity or homozygosity confirmation in validated loci.
Large SNP sets are already available in grain legumes such as soybean Song et al. High-density SNP microarrays can significantly advance breeding applications. JI processed sequencing data, characterized protein-coding genes, and collaboratively with JD performed detailed inspection of gene annotation manually. JI performed gene classification into metabolic and biosynthetic pathways, comparison with other legumes, and generated genome-wide SSR and SNP markers. JR supervised all aspects of the presented analyses. All of the authors contributed to the writing of the manuscript.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The authors thank the Ministry of Agriculture of the Czech Republic grant no.
Is Clover an Alternative Groundcover?
QIA for financial support. Adams, N. Detection of the effects of phytoestrogens on sheep and cattle. Ashrafi, H. BMC Genomics Blanca, J. Botstein, D. Construction of a genetic linkage map in man using restriction fragment length polymorphisms. PubMed Abstract Google Scholar.
Cidade, F. Genetic variation in polyploid forage grass: assessing the molecular genetic variability in the Paspalum genus. BMC Genet. Plant Genomics De Vega, J. Dellaporta, S. Plant Mol. Dice, L. Measures of the amount of ecologic association between species. Ecology 26, Durand, J. Forster, J. Ghamkhar, K. The first genetic maps for subterranean clover Trifolium subterraneum L. Graham, P. Legumes: importance and constraints to greater use. Plant Physiol.
Herrmann, D. Improving persistence in red clover: insights from QTL analysis and comparative phenotypic evaluation. Crop Sci. Isobe, S. Construction of a consensus linkage map for red clover Trifolium pratense L. BMC Plant Biol. Comparative genetic mapping and discovery of linkage disequilibrium across linkage groups in white clover Trifolium repens L. G3 2, — Genome assembly and annotation for red clover Trifolium pratense ; Fabaceae. Jaccard, P. Vaudoise Sci. CrossRef Full Text. Selecting plants with increased total polyphenol oxidases in the genus Trifolium.
Czech J. Plant Breed. Jones, B. Screening legume forages for soluble phenols, polyphenol oxidase and extract browning. Food Agric. Kataoka, R. Integration of linkage and chromosome maps of red clover Trifolium pratense L. Genome Res. Klimenko, I. Mapping candidate QTLs related to plant persistency in red clover. Kongkiatngam, P. Genetic variation among cultivars of red clover Trifolium pratense L. Euphytica 89, — Google Scholar.
Krzywinski, M. Circos: an information aesthetic for comparative genomics. Kulikova, O. Satellite repeats in the functional centromere and pericentromeric heterochromatin of Medicago truncatula. Chromosoma , — Capture of an image to record the placement of the frame and cropping of the initial image accordingly. Due to the procedure of harvesting the plants based on stem placement, a potential mismatch was introduced between the photographed canopy and the analyzed dry matter.
In homogeneous swards this error will approximately even out, due to the similarity of plants being removed and added to the cut. In sparse or inhomogeneous clover-grass swards, however, this uncertainty adds unwanted noise to the gathered sample pairs, especially in the presence of plants with high coverage such as red clover, dandelions, and thistles. Figure 2 shows the variations in the collected sample pairs in terms of clover content in the dry matter composition and the total yield.
The very low clover fraction in multiple samples of the establishment is due to poor plot establishment. Overview of the variation in the sample pairs sorted by the age of the swards. The year axis is discrete, but points are spread out to avoid occlusion of samples. This setup was mounted on a push cart to fix the capture height, minimize movement during shooting, and ease the transportation in the field. The complete camera platform is shown in Figure 3 along with a sample image. The full bit image resolution was exploited to reduce shadow regions in the vegetation using gamma correction and to correct for uneven lighting of the flashes as demonstrated in Figure 3 c,d.
Depending on the vegetation height and camera position, the gathered images had a ground sample distance GSD of 4—6 pixels per mm. Depiction of the camera setup used for the sample acquisition and visualization of the gamma correction used to reduce the impact of shadow regions.
For evaluation of the semantic segmentation performance, a subset consisting of 10 of the sample pairs was selected for hand labeling. Therefore, a single image sample from another dataset was included to maximize the range for semantic segmentation evaluation. Four samples of the subset are shown in Figure 4 in the form of the original framed clover-grass image samples and the corresponding crops.
Visualization of annotated image samples. Top row shows four samples with different clover contents. Bottom row shows crops of the images from the top row to illustrate the image resolution. The black pixels surrounding the image samples mark the border of the harvested samples. When the orientation of the camera does not align perfectly with the harvested patch, the image sample appears to be rotated.
The sample acquisition procedure and equipment follow what was described in Section 2. The field trials were established in and and consisted of ryegrass and white clover mixtures of numerous cultivars. In total, 50 plots were selected by visual inspection to maximize the spread in clover fraction, total yield, time since establishment, stress levels, and phenotypes.
The resulting 50 validation sample pairs were acquired on 20 June , as part of the second cut of the season. Convolutional neural networks require hundreds, or even thousands of labeled images for training [ 20 ]. In the case of semantic segmentation, where each pixel needs to be labeled, achieving this amount of data manually is unfeasible within a reasonable amount of time. This is particularly so in images of species mixtures, where leaves from different species occlude and intertwine with each other and thereby increase the complexity of the labeling task.
As an example, it took an average of 3.
If all primary samples were to be annotated for training, this would correspond to more than h of labor. Dyrmann et al. Inspired by this work, a program which simulates images from clover-grass fields was created. The images are simulated using few hand-segmented samples of clovers, grasses, and weeds, which are placed on top of an image of soil. All the plant samples were acquired from clover-grass field images and selected samples are illustrated in Figure 5.
While generating simulated images, corresponding label-images are automatically generated. A label can take one of the following four classes: clover , grass , soil , and weed. A sample of a simulated images and its corresponding label is shown in Figure 6. Samples of single plants that are used for simulating images. Simulated training data pair.
Grass, clover and soil pixels in the RGB image are denoted by blue, red and black pixels in the label image, respectively. The total number of plant samples was The effect of varying the number of plant samples is discussed in Section 3. To reduce biases in the species distributions and introduce high levels of variability in the simulated images, several field compositions were simulated.
Throughout the field compositions, the clover-grass plant ratio was uniformly distributed. Three different ratios of weeds were used, ranging from none to one-eighth of all the plants in the image. The simulated field density was varied from approximately 10 plants per square meter to complete coverage of the soil. Finally, the clovers were simulated with and without flowers to imitate differences between seasonal cuts.
In total these variations amounted to 46 distinct field compositions with individual minor perturbations per simulated image. To train the neural network, simulated clover-grass images and corresponding label images were generated. To increase the use of the plant samples and extend the variability of the simulated images, each plant sample was randomly augmented with respect to rotation, scale, and saturation before being placed in the simulated image.
Furthermore, the depth was simulated by applying Gaussian shadows to plants when placing them in the image. Applying the large Gaussian kernel to a binary mask of the plant caused thin grasses to cast less shadow than larger clover leaves, leading to a satisfying visual effect. The shadows were added iteratively together with each plant instance, causing underlying soil and plants to become slowly darker.
This effect is clearly seen in Figure 6 a where areas with lower-placed vegetation appear darker. Since the introduction of AlexNet by Krizhevsky et al. These tasks are accomplished by learning abstract features using stacked layers of convolutions and non-linearities by backpropagating thousands of labeled training images through the network and updating the filter kernels accordingly. Contrary to image classification CNNs, the semantic segmentation task of this paper requires a classification output at each spatial position of the input image.
This is traditionally accomplished by replacing fully connected layers with convolutions, to maintain spatial feature maps throughout the network, leading to a fully convolutional neural network. In this study, the fully convolutional network with a output stride of 8 FCN-8s architecture by Long et al.
This model is a modification of the VGG16 architecture [ 22 ], which was made fully convolutional by transforming the fully connected layers to convolutional layers, while preserving the learned parameters. Each of these pooling layers downscales the input by a factor of 2 horizontally and vertically. After the fifth pooling layer, the image is therefore downscaled by a factor of The pooling layers enable the network to learn semantics, at the expense of fine details, which are scarified.
The last convolutional layer and two intermediate layers are followed by deconvolutional layers that up-sample the network output to the size of the input image. In order to compensate for the loss of details and spatial information, shortcuts are made from pooling layers three and four to the deconvolutional layers, whereby finer details can be partly restored. The network architecture is sketched in Figure 7. The fully convolutional network with a output stride of 8 FCN-8s architecture. The network consists of 15 convolutional layers and 5 max pooling layers. The outputs from pooling layers three and four were routed through deconvolution layers to help restore smaller details in the segmented images.
This figure is redrawn from Long et al. The CNNs trained for this research followed the same procedure with regards to data augmentation and hyperparameters as the original architecture and were exclusively trained on simulated clover-grass images and labels. The resolution of the original simulated images is comparable to the gathered clover-grass images of 4—6 pixels per mm.
To increase the robustness of the network towards images captured with various resolutions, at different heights and different plant sizes, the simulated images and corresponding label-images were augmented by scaling the images prior to training. As a result of splitting the task of determining the clover fraction into two problems, the presented results were separated similarly. First, the ability to semantically segment 10 test images was evaluated, followed by evaluation of the dry matter contents on all the gathered sample pairs.
To investigate the requirements of using real plant samples for clover-grass simulations, as shown in Section 2. Therefore, the two trained models only differed in terms of the plant variation in the training images. One model was trained on simulated images from all plant samples. The other model was trained on simulated images from a subset of the plant samples.
The FCN-8s models, trained on simulated data, were evaluated with respect to the quality of the semantic segmentation of real images into clover, grass, and weed pixels. This was tested on the 10 hand-annotated image crops. Following the accuracy metrics of Long et al. Comparison of the approach by Bonesmo et al. IoU: intersection over union. For a fair comparison to the state of the art, the morphological approach of Bonesmo et al.
The test images for this approach were downscaled to match the original GSD stated in the work by Bonesmo et al. Additionally, the parameters were fine-tuned to the images, since the provided parameters led to poor classification results. Along the borders of the image this context information is reduced, leading to a reduced accuracy in the border region. This effect is particularly pronounced when the image size is reduced, as the border region takes up a larger portion of the total image area. To avoid this effect, the test images were semantically segmented in their original sizes, and later cropped, leading to a useful context in the evaluation.
From the segmentation results shown in Table 1 , the improvement from the previous method is clear. The trained convolutional neural network increased the pixelwise accuracy by Qualitative results of the semantic segmentation with the varying clover, grass, and weed densities of Figure 4 are shown in Figure 8. Several observations regarding the FCN-8s models are clearly visible in the four samples:. The majority of the weeds were correctly identified in the first column, while the number of false positives in the third column was reduced. Both FCN-8s models were slightly biased towards detecting clover.
While the classified grass and weed pixels appeared to follow the contours of the plants in the image, the clover classification appeared to be the default choice of the network. This was beneficial in the fourth row, but led to general misclassifications in the second and third rows. Comparison between four ground truth labels of the hand-annotated images, our implementation of the approach by Bonesmo et al.
The areas in blue represent grass, those in red represent clover, those in yellow represent weeds, those in black represent soil, and those in white represent areas that could not be distinguished in the manual annotation. These observations are supported by the quantitative results in Table 1. The networks reached a high pixel-wise accuracy of The increase in mean intersection over union IoU was of 3.
The approach of Bonesmo et al. This was clearly demonstrated in the fourth column.
- GENOME ANNOUNCEMENT.
- Origine du prénom Albert (Oeuvres courtes) (French Edition).
- Sisyphe: Roman (Collection Classique) (French Edition).
- Login using.
The morphological operations were, however, challenged by the varying leaf sizes across the test images. This was exemplified in the second and third column, where wide grass was classified as clover, and small clovers were classified as grass, respectively. The possible use of leaf texture and the surrounding context by the FCN-8s network is believed to be of high importance when segmenting clover grass mixtures. Due to the high level of occlusions in the images, only a fraction of every plant is typically visible, while the leaf texture of the visible part remains.
To verify the usefulness of the trained CNN and validate the coupling between visual clover content in the canopy and clover dry matter fraction, the best-performing FCN-8s CNN was utilized to semantically segment the images from the sample pairs. In contrast to the isolated semantic segmentation case of Section 3. It was better for shadow regions and foreign objects to be ignored than falsely classified, as they influenced the dry matter composition estimation. To apply this, the traditional non-max suppression used for semantic segmentation was substituted by a custom threshold on the individual softmax score maps of the CNN.
By qualitative visual inspection of five images, the softmax thresholds were defined as 0. This corresponds well to the overestimation of clover and underestimation of grass in Figure 8 in Section 3. Three samples of the thresholded segmentation of clover-grass images are shown in Figure 9. Visualization of CNN-driven semantic segmentation of highlighted sample pairs from Figure The first column shows the evaluated image. The second column shows the thresholded identification of grass blue , clover red , and weeds yellow in the image.
Qualitatively, the three samples of semantic segmentation demonstrated an excellent understanding of the images by the convolutional neural network across vegetation densities and clover fractions. Visible weeds, clover, and grasses were detected with high accuracy while withered vegetation and ground were correctly ignored. The coupling between the clover-grass images and dry matter composition was evaluated using the clover fraction metrics defined in Equations 1 and 2 for dry matter samples and image samples, respectively:.
The result of the automated CNN-driven image analysis of visual clover content is shown in Figure The relationship between the CNN-driven analysis and the dry matter clover ratio in the photographed clover-grass patches was eminent and similar to the manually obtained relationship shown by Himstedt et al. The linear regression based on the primary samples pairs led to a clover dry matter fraction prediction with a standard deviation of 7.
This was achieved while covering the spread of total yield, clover fraction, weed levels, fraction of red and white clover, and time in season of the sample pairs. Comparison and linear regression of the visual clover ratio determined by the CNN and the actual clover ratio in the harvested dry matter at the primary field site. The linear relationship between the visual and dry matter fraction of clover in the sample pairs is clear.
The annotations a , b , and c refer to the three image samples shown in Figure 9. Besides the general fitness of the linear regression, multiple outliers appeared in Figure Looking into the underlying image samples and image analysis, two distinct causes were observed. The low-yield outliers along the x-axis all suffered from overestimated clover content. Due to the loss of focus in these samples, blurred vegetation on the ground was often misclassified as clover.
The clover-covered area of the canopy did not always represent the dry matter composition. This was the result of the heavy occlusions and was directly linked to the challenges of using the top-down view. The furthest outlier 0. The trained CNN was also evaluated on the validation field located km from the primary field site, while preserving the threshold parameters. The sample pair relationship of the validation field is shown in Figure When comparing the results of the two field sites, the validation site led to a similar relationship, with an introduction of an offset.
This offset shifts the validation sample pairs towards the upper region of the prediction confidence of the estimator in Figure 10 , leading to a general underestimation of the clover fraction in the validation field. Comparison of the visual clover ratio and the dry matter clover ratio of the validation field site. The linear relationship is more poorly defined, with an introduction of a noticeable offset, possibly due to a less complete segmentation of clovers in the image samples. The ability to train a CNN for image segmentation using solely simulated images has been shown to have good prospects, as it allows one to train a network for tasks for which it would otherwise be unfeasible to achieve the needed amount of data.
This drop is mainly a result of worsened distinction between clover and weeds, caused by the use of only six weeds. This low number of samples does not cover the variation of weeds in the test data in terms of number of neither species nor appearances. This demonstrates that high classification accuracies can be achieved on real images with only few plant samples used for training, and this translates to reaching a working prototype for semantic segmentation within hours with manual labor, as opposed to weeks or months if following a traditional work flow of manual image labeling.
Care should be taken when training the convolutional neural network to span the variations in the test data by training using the corresponding variations in the training data. This was the case for vegetation density and varying ratios of clover, grass, and weeds in the simulated images. More care should be taken when simulating training data to imitate natural and common errors, introduced when collecting images. These include lighting conditions, color temperature, image noise, and blurring. Several of the images with a high ratio of misclassified pixels were blurred.
As the network was trained on sharp images, it is believed that the number of misclassifications can be reduced by introducing blurred simulated images in the training data. From the estimate of the clover content of the dry matter we see that the uncertainty is larger in cases where the clover and grass are mixed compared to cases dominated by one species.
This is because the camera can only see the canopy of the plant cover, and plants of one species that is hidden by the others would thereby make the estimated ratio less accurate.
Is Clover an Alternative Groundcover? | Home Guides | SF Gate
In cases where the sward is dominated by one species, the estimated dry matter-ratio would be less affected by this phenomenon. When evaluating the system on the validation field at a separate location, the automated image-based estimation of the dry matter clover fraction does not translate accurately between the two field sites. The design of the two field trials differs largely, mainly in terms of absence of red clover in the validation field seed mixture, fertilization strategies, organic or conventional farming, and variation of cultivars in the plots.
Through experiments outside of this paper, it has been shown that the system accurately translates to the validation field when lowering the threshold value for detecting clover pixels from 0. This suggests a visual difference between the clovers in the two field sites, leading to partly unclassified clovers. By introducing larger variations of clovers from multiple locations in the image simulation, the CNN should be generalized to better handle natural visual variances between the fields. It is essential to be aware of the accuracy of alternative methods for estimation the botanical dry matter composition.
While separating a forage sample into species by hand does provide the ground truth composition for this paper, this is not feasible for real applications. Fair comparisons include vision-based evaluation by expert consultants in the field and near-infrared reflectance spectroscopy NIRS -based estimation, often integrated in grass sward harvesters used for research.
While the visual estimation accuracy of consultants remains undocumented, this method is time consuming and requires the consultant to inspect the clover-grass throughout the field and map it accordingly. NIRS-based methods, such as in [ 23 ], show comparable estimation accuracy of grass swards of either white clover or red clover, by use of distinct models for each case. The accuracy on grassland swards with mixed red and white clover species, as in this experiment, has not been investigated.
To utilize the botanical composition of clover-grass leys for targeted fertilization, the information must be available at the time of fertilization. The largest amount of fertilizer is typically applied in the spring prior to the start of the growth period. At this stage NIRS for identifying the botanical composition is ruled out, since the crop is too small for harvesting.
This leaves a great potential for the non-destructive approach of monitoring clover-grass mixtures in this paper. Future work includes extending the presented method for clover content estimation to preseason image samples. This extension allows the camera-based system to provide the necessary botanical information for optimizing the fertilizing strategy, for every fertilization. Following the approach presented in this paper, the image analysis of the extension can be operational with a couple of days of labor for gathering relevant image samples and cropping out representative plant samples for data simulation.
Other future work relates to extending the parametric information level delivered by the convolutional neural network to further improve the dry matter composition estimates, or directly predicting the dry matter botanical composition in the neural network.