Research Article  Open Access
Sara Mardanisamani, Tewodros W. Ayalew, Minhajul Arifin Badhon, Nazifa Azam Khan, Gazi Hasnat, Hema Duddu, Steve Shirtliffe, Sally Vail, Ian Stavness, Mark Eramian, "Automatic Microplot Localization Using UAV Images and a Hierarchical ImageBased Optimization Method", Plant Phenomics, vol. 2021, Article ID 9764514, 16 pages, 2021. https://doi.org/10.34133/2021/9764514
Automatic Microplot Localization Using UAV Images and a Hierarchical ImageBased Optimization Method
Abstract
To develop new crop varieties and monitor plant growth, health, and traits, automated analysis of aerial crop images is an attractive alternative to timeconsuming manual inspection. To perform permicroplot phenotypic analysis, localizing and detecting individual microplots in an orthomosaic image of a field are major steps. Our algorithm uses an automatic initialization of the known field layout over the orthomosaic images in roughly the right position. Since the orthomosaic images are stitched from a large number of smaller images, there can be distortion causing microplot rows not to be entirely straight and the automatic initialization to not correctly position every microplot. To overcome this, we have developed a threelevel hierarchical optimization method. First, the initial bounding box position is optimized using an objective function that maximizes the level of vegetation inside the area. Then, columns of microplots are repositioned, constrained by their expected spacing. Finally, the position of microplots is adjusted individually using an objective function that simultaneously maximizes the area of the microplot overlapping vegetation, minimizes spacing variance between microplots, and maximizes each microplot’s alignment relative to other microplots in the same row and column. The orthomosaics used in this study were obtained from multiple dates of canola and wheat breeding trials. The algorithm was able to detect 99.7% of microplots for canola and 99% for wheat. The automatically segmented microplots were compared to ground truth segmentations, resulting in an average DSC of 91.2% and 89.6% across all microplots and orthomosaics in the canola and wheat datasets.
1. Introduction
Changing climate, increasing population, and demand for highquality food will challenge producers to grow more food with greater efficiency. A major pillar of the efforts to address this challenge is the selective breeding of higheryield crop varieties with greater tolerance to extreme environments. Breeding routinely involves assessment of plants from candidate genetic lines to quantify desirable phenotypes. Current methods for assessing plant phenotypes are timeconsuming, relying on laborious manual qualitative assessment, and often only consider a subsampling of the plants grown. For infield trials, where a field area is subdivided into many microplots, within which are grown many plants from a single genetic line, breeders must physically visit each microplot and visually assign scores based on qualitative and quantitative traits of plants and rank the plants at multiple stages of development. A new technique known as imagebased phenotyping has potential to improve this subjective process [1, 2].
Imagebased phenotyping uses machine learning and image processing techniques to obtain quantitative measurements of the structural and functional properties of plants. For infield trials, growth, health, and physical traits can be quantified from overhead images acquired from cameras mounted on tractors or unmanned aerial vehicles (UAV). Aerial imaging has potential for faster and more accurate assessment of a field [1, 3]. A cameraequipped UAV typically collects many highresolution, overlapping aerial images of small sections of the field. Using software, aerial images are stitched together to generate an orthomosaic image, providing a complete view of all the microplots in the field in a single image. Most subsequent analysis relies on knowing the precise locations and identities of each microplot in the orthomosaic image [4, 5]. It is this problem that we study in this paper since localizing microplots in the orthomosaic image is a fundamental first step for any kind of imagebased microplotlevel phenotype analysis.
Microplots may or may not be regularly shaped or aligned due to inaccuracy during planting/seeding. Even if they are aligned, for example, in a gridlike arrangement (which is common), the orthomosaic may be warped due to small, cumulative errors in the stitching process or inaccuracies in geolocation information, which makes microplot localization more challenging. For example, such warping can make it difficult or impossible to rely on an expectation that microplot boundaries are perfectly straight lines.
Microplots can theoretically be extracted from UAV orthomosaics based on GPS data alone. Usually, this means combining GPS data derived from the orthomosaic with preestablished GPSbased microplot shapefiles. However, both GPS sources are not always present. Or depending on the precision of GPS systems, the two GPS sources could disagree, sometimes in a nonlinear way, by 10 meters or more [6]. Finally, as the growing season progresses, microplots can grow beyond their preestablished shapes in unpredictable ways. Taken together, these factors make it difficult to propagate known plot locations from one orthomosaic image to another of the same field taken later or earlier in the season, necessitating an approach to microplot extraction based on image analysis rather than on GPS alone.
Automated detection and segmentation of microplots would enable automatic monitoring and quantification of microplot phenotypes, allowing a faster selection process that requires much fewer personhours than manual assessment.
There are three general approaches used to extract microplots from orthomosaic images such as manual, semiautomatic, and automatic methods.
The first method is the manual method that localizes a microplot using the geolocation information on the drone to generate bounding boxes generated geometrically. Haghighattalab et al. implemented a simple gridbased method to extract microplots [7]. As this method was based on the assumption of a fixed microplot boundary size, the polygon boundary may overlap part of the adjacent microplots. This method is easy to implement and fast, but it did not account for gaps between microplots and gaps between each range of microplots.
The semiautomatic approach partially automates plot localization but still requires some manual intervention. Hearst and Cherkauer proposed a framework to extract microplots based on their map coordinates [8]. This approach used Ground Control Points (GCP) and depends on geolocation accuracy of imagery. Their results might also be affected if there is little spacing between microplots or when spaces between crops are obscured by plants in the later stages of growth. Haghighattalab et al. extracted microplots from wheat breeding nursery images [7]. Their semiautomatic algorithm localizes microplots by classifying pixels of images into “vegetation” and “soil” areas. Similarly, Recio et al. proposed a microplotbased image processing method for automated extraction of trees [9]. They used the means classification algorithm to classify image pixels into “tree” and “nontree” groups. The limitations of both approaches are that they assume ideal separation between plots and sometimes merge multiple microplots into a single microplot. Khan and Miklavic proposed a semiautomated approach to extract microplots from wheat field orthomosaic images [10]. They used particle swarm optimization to find the optimal alignment of a bounding box grid. This method requires a software tool to manually create a cellular grid based on the number of rows and columns laid over the image. Moreover, the method has been tested with only one type of crop and 60 very clearly separated microplots. For many crops and field trial layouts, monitoring the complete growing season requires a robust method to capture images with little to no visible ground between microplots. Tresch et al. developed a semiautomatic approach to obtain microplot information for drone imagery of whole fields using the microplot extraction method known as EasyMPE (microplot extraction) [11]. They were able to identify crop rows and columns in soybean and sugarbeet crops on six different field datasets. However, their method was designed for midseason images in which there is a visible gap between microplots. Their technique might not work for laterseason images when plants are larger and such gaps may not exist. Matias et al. developed an R package called FIELDimageR software to analyze many microplots in orthomosaic images [12]. The software was able to crop microplot area, count the number of plants per microplot, and measure canopy cover percentage, vegetation indices, and plant height. Although the software was userfriendly and able to analyze orthomosaic images, the first stage of the software to extract the microplots was semiautomatic. Robb et al. proposed an image processing method to segment crop microplots from aerial imagery [13]. The proposed method was based on identifying the division between microplots to demarcate them. This paper provided a consistently highperforming approach for delineating crop microplots and used minimal input from the user. In general, existing semiautomatic methods have several drawbacks: they require a tool to create a cellular grid based on the number of rows and columns laid over the image, the grid’s position and attributes are only obtained using external software, and it is only used with one kind of plant and a limited number of clearly delineated microplots. Furthermore, the semiautomatic methods often assume a rectangular grid and are not easily generalizable to nonrectangular arrangements of microplots.
The third approach is automatic methods, which are fully automatic without any manual intervention required. Parraga et al. proposed a segmentation methodology based on image processing techniques using UAV imagery. The proposed segmentation method had four steps: preprocessing, filtering, ROI map, and validation. This method can automatically segment the wheat microplots during the whole growth stage. However, it needs to tune several variables to work for all images during the whole wheat cycle, which results in a long time run. Ahmed et al. proposed an algorithm to detect and segment lentil microplots from multispectral aerial images [14]. One of the advantages of this approach is that the proposed method is fully automatic. However, this method does not correctly work in lateseason images because neighbouring microplots grow into each other. Also, it is unable to distinguish lentil microplots from the guard microplots. Chen and Zhang developed a Python package, GRID (GReenfield Image Decoder), to overcome the limitation of identifying pixel of interest (POI) and microplot segmentation [15]. GRID was able to detect different field layouts such as microplots arranged in rhombus or grid and obtained better performance compared to other software programs with higher computing time. This tool had some limitations. First, the proposed tool is unable to distinguish POI regions if weeds are nearly identical to the vegetation of interest. Second, GRID uses visible channels RGB to control noise and shade and is invalidated for other multispectral images. Despite advances over previous automatic approaches, these methods have some drawbacks. Multiple microplots may be seen as a single microplot, and ideal separation is assumed between microplots. The proposed methods used midseason images that have wellspaced microplots and needed tuning several variables to work for all images.
To the best of our knowledge, all of the published works on microplot extraction have been developed using learningbased and imagebased analysis approaches. Although some models are automatic, all models have been designed and tested with orthomosaics having a relatively small number of microplots and used midseason orthomosaic images where there are no canopy closure and relatively uniform appearance of vegetation within microplots and clear gaps in between microplots. Some methods require a tool to manually create a cellular grid [10]. Among the aforementioned approaches, developing an automated method to test with a wider range of crops and growth stages and combine a priori knowledge of the trial layout (initial shapefile) with an optimization approach to adjust the microplot boundaries based on image analysis of the vegetation and soil is needed. Therefore, developing an automatic method where a map of the known microplot layout is overlaid and positioned using learningbased or imagebased methods is necessary. An approach to automatically identify microplots from orthomosaic images in a broader range of growth stages is needed. Overlaying a map of the known microplot layout is a good starting point, but the precise boundaries of each microplot need to be refined due to the aforementioned problems of warping and geolocation errors. Clicking four corner points of a rectangular area to initialize an algorithm assumes a rectangular grid layout, even though some trials may have more unequal number of plots in rows or columns due to irregular field shapes. Our main objective is to develop a robust method for the accurate extraction of microplots that can be used in highthroughput imagebased plant phenotyping pipelines from aerial droneacquired images of fields with a greater diversity of plot layouts.
In this paper, we propose a threestep imagebased optimization approach where a map of the known microplot layout is overlaid and positioned as optimally as possible. In the next step, the position of individual microplots is optimized within the constraints of the overall layout. Compared to existing methods, ours adds robustness to challenging situations where there are a lack of visible microplot spacing and a lack of no canopy closure. Our approach also only requires a toscale microplot layout map as input. The threestage optimization procedure proceeds in a coursetofine manner from overall field blocks to groups of microplot columns and finally to individual microplots to mitigate the effects of orthomosaic warping.
The rest of the paper is organized as follows. In Section 2, we describe the datasets for wheat and canola, present the microplot localization algorithm, and evaluate algorithm performance. Section 4 discusses the detection and segmentation results of the proposed algorithm. Finally, Section 5 ends the paper with a short summary and conclusion.
2. Material and Methods
2.1. Dataset Description and Image Acquisition
The datasets used in this paper were obtained from two wheat and canola breeding trials. To capture canola images, a Draganfly X4P quadcopter (Draganfly Innovations Inc., Saskatoon, SK, Canada) carrying a MicaSense RedEdge camera (Micasense Inc., Seattle, WA, USA) was used to take images in the summer of 2017. This camera captures images with five spectral channels: red, blue, green, nearinfrared, and rededge. Images were taken at a height of 15 meters or 20 meters depending on the day for canola datasets. The canola orthomosaic images are about pixels in size. The canola trial has 39 rows and 6 columns of microplots in three aligned blocks of two columns. The microplot dimensions were ft with an intermicroplot gap of ft between microplots in the same block. A Sentera double 4K NDVI camera was used to take images of wheat at a height of about 19 meters in the spring of 2018. This camera captures NDVI images. The wheat orthomosaic images are about pixels and contain 47 rows and 12 columns of microplots in three aligned blocks of 15, 15, and 17 rows. The wheat microplot dimensions are ft with an intermicroplot gap of ft between microplots in the same block. A highresolution orthomosaic image of each field was produced by stitching images with Agisoft Metashape (Agisoft LLC, St. Petersburg, Russia). Figure 1 shows the RGB orthomosaic from canola and the NDVI image from wheat breeding trial fields in the summer of 2017 and spring of 2018, respectively.
2.2. Ground Truth
Microplot ground truth bounding boxes were manually annotated in the orthomosaic images by a lab technician. A grid of equalsize rectangles corresponding to the known microplot layout was overlaid on the orthomosaic image so that each microplot is near its correct position. A lab technician used LabelImg to create the rectangles around the microplots. The bounding boxes were created by clicking two opposite corner points. Once done, the annotations were saved as xml in pascal voc format.
2.3. Microplot Localization Algorithm
2.3.1. Algorithm Inputs
Our method requires the following inputs: (i)The orthomosaic image in which microplots are to be localized(ii)The horizontal and vertical scale of the orthomosaic image in pixels per meter(iii)A toscale microplot map of the microplot layout in the field in the form of a list of lists of coordinates of corner points of the microplot bounding boxes in units of meters relative to an arbitrary origin
The scale of the orthomosaic image is determined by the commercial software used to stitch the orthomosaics. The microplot map is easy to produce from the known field layout. It is usually convenient to use the topleftmost microplot corner as . The list of lists of microplot coordinates is ordered so that microplots can be easily associated with their microplot IDs in the field trial design. In the case of our regulargrid field layouts, the microplots are ordered in a columnmajor fashion.
2.3.2. Preprocessing and Map Scaling
RGB orthomosaic image inputs are converted to a grayscale vegetation index image using an index that produces larger values for vegetation than soil. There are a few vegetation indices that satisfy this requirement such as NDVI, Excess Green (ExG), or Excess Green minus Excess Red (ExGExR) used by [16, 17]. Indeed, any vegetation index that behaves similarly could be used. We used ExG for our canola orthomosaic images. Grayscale orthomosaic image inputs are assumed to already be a suitable vegetation index image. In the case of our wheat images, the camera used captured NDVI images directly, which we used as input. It is assumed that the orthomosaic is oriented such that the long edges of microplots are parallel to the horizontal image coordinate axes.
The input scales of the orthomosaic image are used to scale the microplot map so that it is the correct size within the orthomosaic image’s coordinate system by multiplying map’s horizontal and vertical coordinates by the image’s horizontal and vertical scale in pixels per meter, respectively.
2.3.3. PerBlock Optimization
The next step of the algorithm is to find a translation of the scaled microplot map that maximizes the amount of vegetation within the microplot areas. Since the convex hulls of our microplot maps are rectangular, we searched for a translation that maximizes the sum of the vegetation index within the entire microplot map’s bounding rectangle. For more irregularly shaped microplot maps or orthomosaics where stitchinginduced warping would make successful alignment of the entire microplot map difficult, one could divide the microplot map into a small number of blocks and find a translation for the initial alignment of each block independently. In general, let be a set of minimum bounding boxes around blocks of microplots, be the translation of some by , and be the sum of the vegetation index image within a rectangle . The objective function to be maximized is where is the rectangle translated by and is the number of pixels enclosed by all rectangles in . As long as the vegetation index ranges between 0.0 and 1.0, will has a value in the same range.
The that optimizes this objective function is found using differential evolution (DE) [18]. It is an evolutionary algorithm, which is able to find the minimum of a function without requiring its derivative. Since this objective function and objective functions in subsequent algorithm steps are composed of nondifferentiable functions, DE was an ideal optimization algorithm for our task. Figure 2 shows the “optimal” translation of the microplot maps.
2.3.4. PerColumn Optimization
As can be seen in the magnified areas of the orthomosaics in Figure 2, the initial placement of the blocks does not always align every microplot rectangle perfectly with microplots in the orthomosaic. Such misalignment is caused by warping of the orthomosaic which occurs during the stitching process. Some misalignment caused by warping can be reduced by dividing the microplots into more than one block and performing the previously described initialization step independently on each block. In any case, the second phase of our proposed algorithm repositions each column independently within constraints of the expected microplot map to mitigate these effects. This is done across all blocks rather than for each block independently to take advantage of the overall gridlike layout of microplots.
For each column, we let be the bounding box for the entire column and optimize equation (1) using DE to find a translation vector so that encloses a greater sum of excess green. Then, each individual rectangle in the column is translated by . This works for most orthomosaics, but we observed that this can reposition columns such that they are overlapping. This occurs in some lateseason images where there is canopy closure between neighbouring columns with no visible ground. This is corrected in a second pass through the columns where for each column (except for the rightmost column), if the bounding box of overlaps that of the column to its right, , then is translated to the left so that it is positioned pixels to the left of where is the number of pixels between and in the initial scaled microplot map. The percolumn optimization process is described in Algorithm 1.

For microplot maps which are not organized into equal length columns, this step could be generalized and be applied to any set of subsets of microplots and their bounding boxes or convex hulls with varying degrees of ease depending on the complexity of the spatial relationships between the subsets of microplots. For subsets with rectangular bounding boxes in a gridlike arrangement, the generalization is fairly easy. Moreover, this step could equivalently be performed as a perrow optimization.
2.3.5. PerMicroplot Optimization
While the percolumn optimization step somewhat mitigates the orthomosaic warping problem, individual microplots in a column can still be positioned poorly especially if the warping causes vertical displacement.
The permicroplot optimization method requires three steps. The complete steps are summarized in Algorithm 2.

(1) Update Initial Points. Individual microplot positions are optimized in a columnmajor order, starting from the topleft microplot, then working down the column, and starting at the top of the next column to the right when the current column has been completed. The differential evolution (DE) method is a populationbased stochastic method that needs two important parameters for optimization, initial points and bounding research. In our project, the initial points are the upperleft corner and the lowerright corner of each bounding box. To update the initial points, the location of the optimized microplots from the previous column and the location of the previous microplot in the same column are needed. Let and be the upperleft corner and the lowerright corner of all microplot rectangles, respectively.
Let be the height of microplot and the shows the difference between two microplots in the same row but in a different column that can show the amount of warping from the previous column. By observing the warping between two neighbouring microplots in the same row, the amount of displacement due to warping is equal to or less than . Then, initial points are updated when . As can be shown in Figure 3, updated initial points for are and . If , it is more likely that the optimized microplot from the previous column was aligned more closely or exactly to its neighbouring microplot. Therefore, the initial points are not updated. In addition, the initial points of microplots in the first column and the first row of microplots in the microplot map are not updated because there is no information from their previous microplots.
(2) Optimize the Location of Microplot. In this stage of our algorithm, the position of individual microplots in the microplot map is adjusted by optimizing equation (1) to achieve the best possible overlap with the corresponding microplot in the orthomosaic. For this phase, we formulated an objective function consisting of five components using the underlying microplot layout and image characteristics while retaining the general principle of maximizing overlap with plant areas: where is a translation vector, is the bounding box of the microplot whose position is to be optimized, are weighting terms ( and are embedded in , equation (9)), and the remaining parameters to the individual components are functions of and are described in the following paragraphs.
Individual microplot positions are optimized in a columnmajor order, starting from the topleft microplot, then working down the column, and starting at the top of the next column to the right when the current column has been completed. The microplot position is optimized using DE to find the that minimizes this function. The search space of the optimization is bounded such that is between and , and is between and . In our implementations, we use bounds dependent on , the known gap size in pixels between adjacent microplots in the initial scaled mircoplot map.
The first term in the cost function is (equation (1)) which is minimized when the sum of the vegetation index within the microplot is maximized (for our datasets, the index is ExG for canola and NDVI for wheat).
The next term is the even spacing cost (EC). It is minimized when the vertical gap between a microplot and its previously optimized neighbour equals the expected gap based on the initial field map. For microplots at the top of the column that have no previously processed neighbour above them, the EC term is not used. Let be the current gap between the current microplot and the previously optimized microplot above it, where is the initial gap after the optimization of ’s position, but before the commencement of the optimization of ’s position. Let be the normalized vertical gap between the microplot being optimized and microplot above it in the column:
The even spacing cost (EC) is then defined as
The search space for is bounded by the expected vertical gap between microplots so that the optimizer can move the bounding box at most pixels up or down. Therefore, we normalized based on the difference between the maximum possible gaps between the positions of two estimated bounding boxes. We apply an exponential function to this to get . An exponential cost was selected instead of a linear cost to obtain a higher cost value for small normalized gaps (errors) in spacing.
The third term is the alignment cost (AC). It is minimized when the centers of the microplots in a column are horizontally aligned. Let the starting center of the current microplot rectangle be and () be the center of the previously processed microplot rectangle above it. Then, the alignment cost (AC) is
A linear cost here yielded a better result than a nonlinear cost because gaps between columns, especially in the canola images, are larger than the gap between microplots within a column. Thus, the penalty for smaller deviations needed to be lessened in comparison to the evenness cost.
The fourth term is the translation cost (TC) which encourages translation vectors with shorter length. It is minimized when . This is desirable since we can be reasonably certain at this phase of the algorithm that microplots in the microplot map are already quite close to their correct position. Translation cost is defined as the sum of the normalized Manhattan distance of the shift.
The last term in the objective function is the edge cost (EdC). This term encourages microplots to be nudged away from positions where the very edge of the microplot area overlaps nonmicroplot areas in the orthomosaic. It is illustrated in Figure 4. Let be the subimage of the vegetation index orthomosaic enclosed by the microplot’s bounding box with width pixels and height pixels. The edge cost is minimized when the sum of the smallest row sum and smallest column sum of is maximized. Let be the vegetation index of the pixel at the th row and th column of the subimage . The smallest row and column sums are where and are normalizing factors. In the example in Figure 4, the bottommost row and rightmost column have the smallest vegetation index because the microplot is misaligned so the bottom row and right column overlap the ground rather than vegetation. Thus, and are the minimum column and row sums, respectively. The edge cost to be minimized is where and are the remaining weighting terms omitted from equation (3).
(3) Find Overlapped Microplot and Omit It 1.2 cm. In this stage of permicroplot optimization algorithm, the overlapped microplot in the field is found by using information from the previous optimized microplot. The reason for adding this stage to the algorithm is that initial points for the first column and first row of individual microplots in the microplot map are not updated. Since the initial points and research areas play a vital role in optimizing the location of microplots, there is a high possibility of overlapping between microplots in the first column. Two kinds of overlapping may occur after optimizing the location of microplot. First, the current optimized microplot’s location is overlapped with the previous microplot; second, the previously optimized microplot is overlapped with the currently optimized microplot. Figure 5 shows two examples of overlapped microplots in the field.
is defined as the height of the overlapped area between two microplots. Microplot and microplot are two previously optimized neighbours. If the distance is equal or less than expected gap based on the initial map, it means that the current microplot overlapped with its previous neighbour. Then, to omit the overlapping problem, and . When the distance is more than the expected gap , it means that the previously optimized microplot overlapped with the current microplot. Then, to omit the overlapping problem, and .
The complete algorithm is summarized in Algorithm 3.

2.4. Hyperparameter Tuning
We performed a hyperparameter tuning step that optimizes to accommodate seasonal variations in the microplot orthomosaic. We used DE for the hyperparameter search. For each dataset, we randomly selected one early, one mid, and one lateseason image on which to perform the hyperparameter search. The objective function for this DE optimization was the mean DSC of the three images after segmentation with our algorithm, and are the variables being optimized. For the DE algorithm, the minimum and maximum bounds on the variables were set to 0.0 and 1.0.
The values of the hyperparameters used for evaluation are reported in Section 2.5.3.
2.5. Algorithm Performance Evaluation
We assessed the algorithm’s ability to detect microplots and its ability to segment microplots accurately.
2.5.1. Detection Metrics
A microplot is considered detected if the overlap between the microplot and its ground truth is acceptably high. We quantified this overlap using the Dice Similarity Coefficient (DSC) [19]. If is the set of pixels in a ground truth microplot and is the set of pixels in the corresponding automatically segmented microplot, then the DSC is
Thus, a microplot can be considered detected if its DSC exceeds some threshold . DSC ranges between 0.0 (no overlap at all) and 1.0 (perfect overlap).
2.5.2. Segmentation Metrics
To evaluate the accuracy of the segmentation, we determined the DSC for each microplot (including microplots not deemed to be detected). We also determined the displacement error for each microplot, defined as the Euclidean distance between the center point of the segmented rectangle and its corresponding ground truth bounding box. This metric is used by [10].
In order to facilitate comparison with future studies, we also quantified the segmentation accuracy in terms of sensitivityspecificity and precisionrecall. Sensitivity, specificity, precision, and recall are quantities that are computed from the following counts: (i)True Positive Pixels (TP): number of pixels correctly labeled as belonging to the correct microplot(ii)True Negative Pixels (TN): number of pixels incorrectly labeled as belonging to the correct microplot(iii)False Positive Pixels (FP): number of pixels correctly labeled as background (not belonging to the correct microplot)(iv)False Negative Pixels (FN): number of pixels incorrectly labeled as background
Sensitivity and specificity are complementary measures. Sensitivity is the proportion of pixels that are part of a microplot that are correctly labeled as such:
Specificity is the proportion of pixels that are not part of a microplot that are correctly labeled as such:
The segmentation accuracy is higher when both sensitivity and specificity are close to 1.0. When sensitivity is higher, it means that more of the foreground pixels are being included in the segmentation. When specificity is higher, it means fewer background pixels are being included in the foreground. Both conditions are required for excellent segmentation.
Precision and recall are similar complementary measures in that better segmentations are characterized by both measures being closer to 1.0. Recall is defined in the same way as sensitivity (see above). Precision is the proportion of pixels labeled as microplot that actually do belong to microplots:
The main difference between sensitivityspecificity and precisionrecall is that the latter does not take into account TN pixels in the segmentation. Precisionrecall therefore does not account for correctly labeled nonmicroplot pixels. However, precisionrecall is a good alternative to sensitivityspecificity when the number of background (nonmicroplot) pixels is large relative to the number of foreground (microplot) pixels in an image, which can inflate the specificity score because TN pixels vastly outnumber FP pixels. Thus, we report both pairs of measures.
2.5.3. Evaluation Procedure
Our algorithm was evaluated on the ten canola and seven wheat orthomosaic images described in Section 2.1. The algorithm was implemented in Python on a MacBook Pro (15inch, 2017) equipped with 2.8 GHz QuadCore Intel Core i7 CPU (16 G memory).
For the canola dataset, the perblock optimization step was performed using a single block for the entire field. For wheat, the warping was severe enough to require dividing the microplots into three blocks for this step. The blocks were formed from the three obvious subregions in the field with the larger spaces between them as seen in Figure 1(b). Results of the perblock optimization step on the wheat dataset in a single block and three blocks are shown in Figures 6(a) and 6(b), respectively. To assess the benefits of using multiple blocks, we also computed results for the algorithm where all microplots in the wheat orthomosaics are processed as a single block.
The search bounds on for the DE optimizer in the perblock optimization step were set to , , and , where is the height of orthomosaic image minus the height of all microplots and the expected distance between them and is the width of the orthomosaic image minus the width of all microplots and the expected distance between columns taken from the initial microplot map for both canola and wheat orthomosaics.
The search bounds on for the DE optimizer in the percolumn optimization step were set to and , where is the expected distance between immediately adjacent microplots and is the expected distance between immediately adjacent columns taken from the initial microplot map for both canola and wheat orthomosaics.
The search bounds on for the DE optimizer in the permicroplot optimization step were set to and for canola orthomosaics. Due to severe warping in the wheat dataset, two search bounds were set. and for microplots in the first row and the first column of microplots in the microplot map in the orthomosaics and and for the rest of microplots in the microplot map in the orthomosaics where is the expected distance between immediately adjacent microplots taken from the initial microplot map.
Hyperparameters used for both wheat and canola datasets were and , respectively (rounded to three significant digits).
Because the differential evolution (DE) method is a populationbased stochastic method that uses random initializations, we ran our algorithm 5 times for each orthomosaic image and retained the result with the smallest sum of over all microplots.
3. Results and Discussion
3.1. Detection Results
We considered a microplot to be detected with an acceptable overlap if its DSC is 0.5 or greater. It must be understood that this DSC threshold is not used to measure whether a plot is segmented acceptably well, but rather as a measure of the detection rate, that is, the proportion of microplots for which we are able to establish a correspondence with one of the bounding boxes and associate it with a microplot ID. The position of such rectangles is adjusted in a later phase of the algorithm before assessing the quality of the segmentation of the microplot. A large proportion of microplots was detected using this criterion in both the wheat and canola datasets for early, mid, and lateseason images. Table 1(b) presents the number of unidentified microplots for each canola orthomosaic image. On average, 99.74% of microplots in an orthomosaic images were detected, an average of 0.6 undetected microplots per orthomosaic.
(a) The results of DSC and median of displacement error after using perblock, percolumn, and permicroplot optimization steps for the canola dataset  
 
(b) The results of DSC and correctly identified microplots for the canola dataset  

Table 2(b) reports the same data for the wheat orthomosaics using one block and three blocks for the perblock optimization step. Because of the aforementioned warping, the algorithm detected only 88.55% of microplots in an orthomosaic on average or an average of 64.57 unidentified microplots per image (using the criterion). However, the use of three blocks, which processes the initial microplot map in smaller sections, achieves much improved performance. Table 2(b) shows the results of using three blocks where the average percentage of identified microplots in an orthomosaic improves to 98.99% or an average of 5.7 unidentified microplots per orthomosaic.
(a) The results of DSC and median of displacement error after using perblock, percolumn, and permicroplot optimization steps for the wheat dataset using a single block and three blocks in the perblock optimization step  
 
(b) The results of DSC and correctly identified microplots for the wheat dataset using a single block and three blocks in the perblock optimization step  

3.2. Segmentation Results
Tables 1(a) and 2(a) show the mean microplot DSC and median of displacement error after perblock, percolumn, and permicroplot optimization steps for our ten canola and seven wheat orthomosaic images. The algorithm achieved an average DSC of 86.58%, 88.34%, and 91.15% across all microplots and orthomosaics tested for the canola dataset after perblock, percolumn, and permicroplot optimization steps, respectively. Furthermore, the algorithm achieved an average DSC of 59.04%, 67.67%, and 89.56% across all microplots and orthomosaics tested for the wheat dataset after perblock, percolumn, and permicroplot optimization steps, respectively. We present the displacement errors for each of the ten different canola trials in Table 1(a), and the median errors are 0.86, 0.72, and 0.51 ft across all microplots and orthomosaics tested for the canola dataset after perblock, percolumn, and permicroplot optimization steps, respectively. In addition, the median errors are 3.76, 2.72, and 0.42 ft across all microplots and orthomosaics tested for the wheat dataset after perblock, percolumn, and permicroplot optimization steps, respectively. The mean microplot DSC is significantly increased when three optimization steps are performed for the wheat and canola datasets. In addition, the median of displacement error is reduced compared to performing only one or two optimization steps. The results show the potential advantages of performing a threelevel hierarchical optimization method in the presence of low to high severe orthomosaic warping.
Table 1(b) shows the mean microplot DSC (including undetected microplots) for our ten canola orthomosaic images. The algorithm achieved an average DSC of 91.15% across all microplots and orthomosaics tested for the canola dataset. The August 16 image is a lateseason image where the gaps between microplots were obscured by vegetation. Although the median of displacement errors and mean of DSC are higher than those of other images, the number of unidentified microplots remained low (Table 1(a)). It is the only canola image where the number of unidentified microplots was greater than 1.
Table 2(b) shows mean DSC and the displacement errors for the seven different wheat orthomosaics when a single block and three blocks are used. The algorithm achieved an average DSC of 88.55% and 98.99% across all microplots and orthomosaics tested for the wheat dataset using single block and three blocks, respectively. The median of displacement error when a single block used is 0.52 ft overall. The warping results in microplot map rectangles being closer to a neighbouring microplot in the orthomosaic than the correct microplot in the orthomosaic. Then, the permicroplot optimization step aligns them more closely to the neighbouring microplot. The spacing between microplots is 2 ft and the width of each microplot is 3.5 ft; thus, a microplot aligning to its neighbouring microplot would cause a displacement error of around 5.5 ft. Table 2(b) shows the median of displacement error when three blocks used are 0.42 ft overall. The number of the undetected microplot is reduced compared to a single block, showing the potential advantages of processing the microplot map in blocks in the presence of orthomosaic warping.
Figure 7 shows a plot of precisionrecall and ROC curves for both canola and wheat datasets, respectively. In Figure 7(a), at recall 0.8, most images have a precision of 0.9 or higher except one image. The August 16 image is a lateseason image that the gaps between microplots were covered by vegetation, and the estimated bounding boxes were more likely to be out of position as a result. Figure 7(b) shows that at a specificity of 0.93, the sensitivity of the segmentation is at least 0.8. Again the exception is the August 16 image.
In Figure 7(c), for the wheat images (using 3 blocks), at recall 0.8, most images have a precision of 0.88 or higher except one image. The May 8 image is a lateseason image that the gaps between microplots were covered by vegetation. Although the image had a lower precision at recall 0.8, the percentage of the identified microplots for the image was 96.63% from Table 2(b) and 545 microplots were detected () out of 564 microplots. Also, in Figure 7(d), we see that at specificity 0.8, all but one image has a sensitivity of 0.72 or higher. Again the exception is the May 8 lateseason images.
Figures 7(e) and 7(f) illustrate how selection of the DSC threshold for microplot detection affects the number of unidentified microplots. In Figure 7(e), at , the number of unidentified microplots for the canola dataset is almost zero and does not begin to increase rapidly until the DSC threshold is 0.7 or more. Similar behaviour is observed for the wheat dataset.
Figure 8 shows sample segmentations for both crops used in this study including examples of early, mid, and lateseason orthomosaics. Red rectangles illustrate the final positions of the microplot map rectangles after completion of the algorithm.
4. Discussion
We have developed a fully automatic method to extract microplots in orthomosaic images. Our proposed method can be applied to rectangularshaped microplot layouts. These layouts include rows and columns arranged perpendicularly. Our method works for NDVI and ExG images and allows users to derive the different vegetation indices as input. The most important advantage of this method can be generalized to other crop types as long as the field layout map is provided and plants have a consistent size within a trial. The main difference of our approach with previous methods is combining prior knowledge (shapefile) with image analysis within a multiobjective optimization and evaluating the approach with a wide range of crop type and growth stage. We obtained a better result for the canola dataset compared to the wheat dataset. As shown in Figure 1, there is more misalignment in the wheat dataset caused by the warping of the orthomosaic, which occurs during the stitching process. Furthermore, the wheat dataset has some missing microplots that caused an error while extracting the exact location of microplots.
As shown in Tables 1(a) and 2(a), the mean microplot DSC is significantly increased, and the median of displacement error is reduced after performing each optimization step. It means that each step of optimization is necessary for the presence of low to high severe orthomosaic warping for the wheat and canola datasets.
In this project, initial points and research bounds are essential parameters for the optimization to get the best results. Due to severe warping in the microplot map, some initial points are closer to a neighbouring microplot in the orthomosaic than to the correct microplot in the orthomosaic. Then, the permicroplot optimization step aligns them more closely to the neighbouring microplot. Updating initial points could help to solve an overlap problem and decrease the number of unidentified microplots. To update initial points, the algorithm requires the information of previously optimized microplots. Because of this, the initial points in the first row and column of microplots in the microplot map in the orthomosaic could not be updated. Then, the stage of finding an overlapped microplot was added to the permicroplot optimization step.
The method for extracting and localizing microplots in the field described in this study would benefit any applications that use a remotely sensed field trial, whether for lodging prediction, herbicide tolerance estimation, detecting crop rows, detecting plants, estimating height, or estimating yield [20–22]. For any of these tasks, we must begin by establishing the precise location and area and perimeter of the field’s microplots. Although imagebased phenotyping performs better than previous manual techniques, there is still substantial labour involved in identifying and segmenting microplots using manual, semiautomatic, and automatic methods. The automatic registration of microplot locations provided by the optimization algorithm proposed here can simplify the tedious preanalysis task of identifying field microplots in an orthomosaic image of a plant research field trial. Our research, by obtaining precise location from microplots, provides a useful and scalable method for highthroughput plant phenotyping scenarios. The outcome of this research will help to increase the throughput of aerial imagebased phenotyping.
One limitation of permicroplot optimization used for microplot identification is that our algorithm used NDVI images or ExG indices that provided higher intensity value of vegetation than that of soil. A limitation of finding hyperparameters is that hyperparameter tuning on the whole dataset is quite timeconsuming. As such, the hyperparameter tuning method used in this study has used only three images from the dataset to obtain an efficient weight vector. Using a technique such as fold cross validation on the whole dataset may result in hyperparameter values that better generalize unseen datasets. This could take the form of random splits of a dataset into “training” (for optimizing hyperparameters) and “testing” (for performance evaluation) sets.
In future work, we aim to develop an algorithm that accounts for the different physical properties of various crop types and different microplot sizes. This will be achieved by finding general optimization techniques and making modifications to the objective function that are well suited to manipulate images and vegetation properties. In addition, since the search bounds are important when optimizing the position of blocks, columns, and individual microplots, we suggest including the search bounds in the hyperparameter tuning model. Furthermore, since height information might be helpful for lateseason imagery, we could add the height information as another channel to our algorithm. Also, EasyMPE is a semiautomatic approach to obtain microplot information for drone imagery of whole fields [11]. The main part of EasyMPE is the microplot extraction part that is related to our work. This method was designed for midseason images in which there is a visible gap between microplots. In this method, the field is partitioned based on the sum of white pixels in the columns in a segmented binary image. Given the knowledge that we have on lateseason and earlyseason images, it is more likely that this method will not work on these kinds of images. In future work, we could compare our algorithm with the EasyMPE method and apply EasyMPE in lateseason and earlyseason images to ensure our claim.
In some cases, plant senescence or disease may cause microplots to no longer be green, in which case it will be difficult for our algorithm to detect it. Indeed, some examples of this can be seen in the lower block of microplots in Figure 8. We might be able to detect and repair such errors by comparing overlapping microplot rectangles to all four of their expected neighbours in the grid layout. This additional context could be either added as a postprocessing step or incorporated into the objective function of the permicroplot optimization step.
5. Conclusion
We have developed an algorithm based on a novel application of imagebased optimization techniques to extract and segment microplots within a field. Our proposed method for segmenting wheat and canola trials performs automatic initialization of the known field layout over the orthomosaic images in roughly the right position. The algorithm not only relieves the processing time bottleneck of identifying and segmenting microplots in highthroughput imagebased phenotyping pipeline but also can simplify the tedious preanalysis task of identifying microplots in an orthomosaic image. Being able to segment and extract the exact location of microplots for highthroughput plant phenotyping scenarios, such pipelines will be critical in the future search for the higher quality of harvested crop needed to feed a growing population. As the threestage optimization procedure proceeds in a coursetofine manner from overall field blocks, it has the potential to be adapted for more specific purposes, such as extracting the different physical properties of various crop types and different microplot sizes.
Data Availability
The data is available upon request.
Conflicts of Interest
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.
Authors’ Contributions
S. Mardanisamani developed, improved, and implemented the main idea and expanded and modeled it to an automated hierarchical algorithm to detect/segment all microplots individually in both canola and wheat datasets, prepared related figures and tables throughout the paper, assessed the method performance using detection and segmentation metrics, and wrote the paper. T. W. Ayalew modeled the main problem as an optimization algorithm, contributed to the objective function design and implementation, implemented and tuned the hyperparameter algorithm, performed initial evaluation, and contributed to the writing of the permicroplot optimization section. M. A. Badhon contributed to the development and implementation of the algorithm, tuned initial hyperparameters, and performed preliminary evaluations. N. A. Khan contributed to the writing of a preliminary version of the abstract and introduction, in implementation of some initial methodologies, experimental results, and revisions. G. Hasnat helped in creating orthomosaic ground truth, writing the methodology skeleton, and drawing the image illustrations. H. Duddu and S. Shirtliffe provided the imagery for the canola dataset. S. Vail was the breeder for the canola microplots. I. Stavness helped formulate the main idea and helped revise the manuscript. M. Eramian contributed to the conceptual idea of positioning the initial microplot map and then refining the position of the individual microplots and formulating the positioning as an optimization problem, supervised the daytoday activities of the first author, and was a major contributor to the writing and presentation of the paper.
Acknowledgments
The authors would like to acknowledge William van der Kamp for his curation of the datasets, helping us find the imagery we needed, and providing the georeferencing information needed to compute image scales. We also thank BioConsortia, Inc. (Davis, California, USA) for providing imagery for the wheat experiments which are used with their generous permission. ExThis research was undertaken thanks in part to funding from the Canada First Research Excellence Fund.
References
 J. L. Araus and J. E. Cairns, “Field highthroughput phenotyping: the new crop breeding frontier,” Trends in Plant Science, vol. 19, no. 1, pp. 52–61, 2014. View at: Publisher Site  Google Scholar
 A. Chawade, J. van Ham, H. Blomquist, O. Bagge, E. Alexandersson, and R. Ortiz, “Highthroughput fieldphenotyping tools for plant breeding and precision agriculture,” Agronomy, vol. 9, no. 5, p. 258, 2019. View at: Publisher Site  Google Scholar
 R. Xu, C. Li, and A. H. Paterson, “Multispectral imaging and unmanned aerial systems for cotton plant phenotyping,” PLoS One, vol. 14, no. 2, article e0205083, 2019. View at: Publisher Site  Google Scholar
 W. Guo, B. Zheng, A. B. Potgieter et al., “Aerial imagery analysisquantifying appearance and number of sorghum heads for applications in breeding and agronomy,” Frontiers in plant science, vol. 9, p. 1544, 2018. View at: Publisher Site  Google Scholar
 T. Duan, B. Zheng, W. Guo, S. Ninomiya, Y. Guo, and S. C. Chapman, “Comparison of ground cover estimates from experiment plots in cotton, sorghum and sugarcane based on images and orthomosaics captured by UAV,” Functional Plant Biology, vol. 44, no. 1, pp. 169–183, 2017. View at: Publisher Site  Google Scholar
 P. Barry and R. Coakley, “FIELD Accuracy TEST of RPAS photogrammetry,” International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XL1/W2, pp. 27–31, 2013. View at: Publisher Site  Google Scholar
 A. Haghighattalab, L. G. Perez, S. Mondal et al., “Application of unmanned aerial systems for high throughput phenotyping of large wheat breeding nurseries,” Plant Methods, vol. 12, no. 1, p. 35, 2016. View at: Google Scholar
 A. A. Hearst and K. A. Cherkauer, “Research Article: Extraction of small spatial plots from georegistered UAS imagery of crop fields,” Environmental Practice, vol. 17, no. 3, pp. 178–187, 2015. View at: Publisher Site  Google Scholar
 J. Recio, T. Hermosilla, L. Ruiz, and J. Palomar, “Automated extraction of tree and plotbased parameters in citrus orchards from aerial images,” Computers and Electronics in Agriculture, vol. 90, pp. 24–34, 2013. View at: Publisher Site  Google Scholar
 Z. Khan and S. J. Miklavcic, “An automatic field plot extraction method from aerial orthomosaic images,” Frontiers in plant science, vol. 10, p. 683, 2019. View at: Publisher Site  Google Scholar
 L. Tresch, Y. Mu, A. Itoh et al., “Easy MPE: extraction of quality microplot images for UAVbased highthroughput field phenotyping,” Plant Phenomics, vol. 2019, article 2591849, pp. 1–9, 2019. View at: Publisher Site  Google Scholar
 F. I. Matias, M. V. CarazaHarter, and J. B. Endelman, “FieldimageR: an R package to analyze orthomosaic images from agricultural field trials,” The Plant Phenome Journal, vol. 3, no. 1, 2020. View at: Publisher Site  Google Scholar
 C. Robb, A. Hardy, J. H. Doonan, and J. Brook, “Semiautomated field plot segmentation from UAS imagery for experimental agriculture,” Frontiers in Plant Science, vol. 11, 2020. View at: Publisher Site  Google Scholar
 I. Ahmed, M. Eramian, I. Ovsyannikov et al., “Automatic detection and segmentation of lentil crop breeding plots from multispectral images captured by UAVmounted camera,” in 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1673–1681, Waikoloa, HI, USA, 2019. View at: Publisher Site  Google Scholar
 C. J. Chen and Z. Zhang, “Grid: a python package for field plot phenotyping using aerial images,” Remote Sensing, vol. 12, no. 11, p. 1697, 2020. View at: Publisher Site  Google Scholar
 G. E. Meyer and J. C. Neto, “Verification of color vegetation indices for automated crop imaging applications,” Computers and Electronics in Agriculture, vol. 63, no. 2, pp. 282–293, 2008. View at: Publisher Site  Google Scholar
 D. M. Woebbecke, G. E. Meyer, K. von Bargen, and D. A. Mortensen, “Color indices for weed identification under various soil, residue, and lighting conditions,” Transactions of the ASAE, vol. 38, no. 1, pp. 259–269, 1995. View at: Publisher Site  Google Scholar
 R. Storn and K. Price, “Differential evolution  a simple and efficient heuristic for global optimization over continuous spaces,” Journal of Global Optimization, vol. 11, no. 4, pp. 341–359, 1997. View at: Publisher Site  Google Scholar
 L. R. Dice, “Measures of the amount of ecologic association between species,” Ecology, vol. 26, no. 3, pp. 297–302, 1945. View at: Publisher Site  Google Scholar
 S. Mardanisamani, F. Maleki, S. Hosseinzadeh Kassani et al., “Crop lodging prediction from UAVacquired images of wheat and canola using a DCNN augmented with handcrafted texture features,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Long Beach, California, USA, 2019. View at: Google Scholar
 P. Lottes, J. Behley, N. Chebrolu, A. Milioto, and C. Stachniss, “Robust joint stem detection and cropweed classification using image sequences for plantspecific treatment in precision farming,” Journal of Field Robotics, vol. 37, no. 1, pp. 20–34, 2020. View at: Publisher Site  Google Scholar
 M. N. Reza, I. S. Na, S. W. Baek, and K.H. Lee, “Rice yield estimation based on Kmeans clustering with graphcut segmentation using lowaltitude UAV images,” Biosystems Engineering, vol. 177, pp. 109–121, 2019. View at: Publisher Site  Google Scholar
Copyright
Copyright © 2021 Sara Mardanisamani et al. Exclusive Licensee Nanjing Agricultural University. Distributed under a Creative Commons Attribution License (CC BY 4.0).