Delineation of Crop Field Areas and Boundaries from UAS Imagery Using PBIA and GEOBIA with Random Forest Classiﬁcation

: Unmanned aircraft systems (UAS) have been proven cost- and time-e ﬀ ective remote-sensing platforms for precision agriculture applications. This study presents a method for automatic delineation of ﬁeld areas and boundaries that uses UAS multispectral orthomosaics acquired over 7 vegetated ﬁelds having a variety of crops in Prince Edward Island (PEI). This information is needed by crop insurance agencies and growers for an accurate determination of crop insurance premiums. The ﬁeld areas and boundaries were delineated by applying both a pixel-based and an object-based supervised random forest (RF) classiﬁer applied to reﬂectance and vegetation index images, followed by a vectorization pipeline. Both methodologies performed exceptionally well, resulting in a mean area goodness of ﬁt (AGoF) for the ﬁeld areas greater than 98% and a mean boundary mean positional error (BMPE) lower than 0.8 m for the seven surveyed ﬁelds.


Introduction
Today, under the scope of climate change, climatic hazards have a higher probability of occurrence and crop insurance has become more critical to growers. To have proper insurance premiums, both the growers and the crop insurance agencies need to have precise measurements of the cropped field area and boundaries. Field areas and boundaries can be determined by Global Navigation Satellite System (GNSS)-based in situ surveys but they are costly in time and effort. Space-borne optical images have also been tested [1][2][3], but they lack in spatial and temporal resolution. Moreover, this imagery is prone to atmospheric interference and it can be costly when acquired by commercial satellites. An alternative is to use images acquired by unmanned aircraft systems (UASs) which have the advantages of being portable, flexible, and cost and time-effective [4]. UAS imagery is big data, but efficient machine-learning (ML) algorithms with today's computational capacities have also made the big data process much easier than in the recent past [5].
There are only a few studies about using UAS imagery for field areas and boundaries extraction, mainly for cadastral applications. Red, green and blue (RGB) UAS imagery was used for extracting cadastral boundary features using a deep learning algorithm [6] or using the ENVI (Exelis Visual Information Solutions, Boulder, CO, USA) segmentation and object generation software [7]. In both cases, the fields were either bare soil or vegetated. Bare soil field boundaries were delineated from blue

Unmanned Aircraft System (UAS) Data Collection
The UAS image data were acquired during summer 2018 through multiple surveying campaigns under clear sky conditions (Table 1). The unmanned aerial vehicle (UAV) was a DJI Matrice 100 and the multispectral payload was the MicaSense RedEdge 3 camera (MicaSense Inc., U.S.A.). The RedEdge camera has five sensors capturing the radiation reflected in the blue, green, red, red-edge, and NIR bands of the electromagnetic spectrum ( Table 2). The spatial resolution or ground sample distance (GSD in cm/pixel) is common for all five sensors and depends on the UAV flight altitude, the sensors field of view (horizontal: 47.9°, vertical: 36.9°, diagonal: 58.1°), and the focal length (5.4 mm). With a flying altitude of 100 m, the corresponding GSD is about 8 cm/pixel.

Unmanned Aircraft System (UAS) Data Collection
The UAS image data were acquired during summer 2018 through multiple surveying campaigns under clear sky conditions (Table 1). The unmanned aerial vehicle (UAV) was a DJI Matrice 100 and the multispectral payload was the MicaSense RedEdge 3 camera (MicaSense Inc., USA). The RedEdge camera has five sensors capturing the radiation reflected in the blue, green, red, red-edge, and NIR bands of the electromagnetic spectrum ( Table 2). The spatial resolution or ground sample distance (GSD in cm/pixel) is common for all five sensors and depends on the UAV flight altitude, the sensors field of view (horizontal: 47.9 • , vertical: 36.9 • , diagonal: 58.1 • ), and the focal length (5.4 mm). With a flying altitude of 100 m, the corresponding GSD is about 8 cm/pixel. The camera performs radiometric calibration of the acquired imagery with data from the downwelling light sensor positioned on the top of the UAV, while the sensors are looking at the nadir position with respect to the UAV flight path. The Micasense RedEdge system also incorporates a GNSS receiver for the accurate geolocation of the captured imagery. To allow the creation of reflectance orthomosaics, a MicaSense RedEdge reflectance calibration panel is employed before every survey to measure the incoming radiation. The UAV flights were conducted at an altitude of 100 m, at noontime for minimizing shadowing effects and under clear sky conditions to avoid cloud shadowing. The flight paths were planned so that there is a minimum of 75% overlapping and side-lapping for neighboring images.

Data Processing
For this study, the visualization, maps generation, and general Geographic Information Systems (GIS) procedures were handled through the QGIS software [23]. Two image processing methods were tested and assessed for their value in the delineation of field areas and boundaries that was performed with two different levels of information: firstly, on the pixel-level information (PBIA methodology) and secondly on the object-level information (GEOBIA methodology). The steps and procedures used for delineating field areas and boundaries are described in Figure 2 for the PBIA methodology and in Figure 3 for the GEOBIA methodology [15,24].
In both PBIA and GEOBIA methodologies, the first step was to generate georeferenced reflectance orthomosaics from the blue, green, red, red-edge, and NIR images from each flight campaign through the photogrammetric commercial software Pix4D Mapper (Pix4D SA, Switzerland). This process involves Downwelling Light Sensor (DLS) and reflectance panel corrections. All surveyed fields are treated individually and the corresponding rasters were clipped and extracted from each orthomosaic with a surrounding area buffer using Geospatial Data Abstraction Library (GDAL) [25].
The blue and red reflectance orthomosaics were then used to compute a simple ratio vegetation index (VI) between the two reflectance bands as follows: blue-red simple ratio (BRSR) = blue/red [8], which enhances the spectral difference between soil and vegetation and has already been found to be useful in our previous study on bare soil field areas delineation [8].
Remote Sens. 2020, 12, x 4 of 25 The camera performs radiometric calibration of the acquired imagery with data from the downwelling light sensor positioned on the top of the UAV, while the sensors are looking at the nadir position with respect to the UAV flight path. The Micasense RedEdge system also incorporates a GNSS receiver for the accurate geolocation of the captured imagery. To allow the creation of reflectance orthomosaics, a MicaSense RedEdge reflectance calibration panel is employed before every survey to measure the incoming radiation. The UAV flights were conducted at an altitude of 100 m, at noontime for minimizing shadowing effects and under clear sky conditions to avoid cloud shadowing. The flight paths were planned so that there is a minimum of 75% overlapping and side-lapping for neighboring images.

Data Processing
For this study, the visualization, maps generation, and general Geographic Information Systems (GIS) procedures were handled through the QGIS software [23]. Two image processing methods were tested and assessed for their value in the delineation of field areas and boundaries that was performed with two different levels of information: firstly, on the pixel-level information (PBIA methodology) and secondly on the object-level information (GEOBIA methodology). The steps and procedures used for delineating field areas and boundaries are described in Figure 2 for the PBIA methodology and in Figure 3 for the GEOBIA methodology [15,24].   In both PBIA and GEOBIA methodologies, the first step was to generate georeferenced reflectance orthomosaics from the blue, green, red, red-edge, and NIR images from each flight campaign through the photogrammetric commercial software Pix4D Mapper (Pix4D SA, Switzerland). This process involves Downwelling Light Sensor (DLS) and reflectance panel corrections. All surveyed fields are treated individually and the corresponding rasters were clipped and extracted from each orthomosaic with a surrounding area buffer using Geospatial Data Abstraction Library (GDAL) [25].
The blue and red reflectance orthomosaics were then used to compute a simple ratio vegetation index (VI) between the two reflectance bands as follows: blue-red simple ratio (BRSR) = blue/red [8], which enhances the spectral difference between soil and vegetation and has already been found to be useful in our previous study on bare soil field areas delineation [8].

Random Forests
Both methodologies utilize RF for classification. The classification scheme produces three classes: soil, crop, and other vegetation. With respect to the input features for classification, while being initially tested, Haralick et al. [26]'s Gray Level Co-occurrence Matrix (GLCM) textural features were not used because the classification accuracies were already high enough without using textural features.
Since the RF classifier was introduced by Breiman [27], it has been widely applied in remote sensing [28] due to its robustness and advantages, being easily parametrized and fast. RF is a non-parametric decision-tree classification algorithm that does not assume a normal distribution of the data [27,29]. RF is an ensemble classification model, aggregating a user-defined number of uncorrelated classification trees, every one of which is grown with a set of randomly chosen features from the feature space at each tree node. The final class decision is made through the majority voting of the full ensemble of the trained trees. In this study, we deployed the off-the-self RF implementation from the R programming language v3.5.1 [30]. The number of trees grown ntree was set to 500 and the number of random features mtry for the growth of each decision tree in the ensemble was kept as its default value which is the square root of the total number of features, rounded down.
To determine the classification accuracy, RF performs internally a procedure resulting in an out-of-bag (OOB) error rate. RF works by bootstrapping a sample dataset from the original training data for every tree grown. From the original data, an approximate 37% become "out-of-bag" data

Random Forests
Both methodologies utilize RF for classification. The classification scheme produces three classes: soil, crop, and other vegetation. With respect to the input features for classification, while being initially tested, Haralick et al. [26]'s Gray Level Co-occurrence Matrix (GLCM) textural features were not used because the classification accuracies were already high enough without using textural features.
Since the RF classifier was introduced by Breiman [27], it has been widely applied in remote sensing [28] due to its robustness and advantages, being easily parametrized and fast. RF is a non-parametric decision-tree classification algorithm that does not assume a normal distribution of the data [27,29]. RF is an ensemble classification model, aggregating a user-defined number of uncorrelated classification trees, every one of which is grown with a set of randomly chosen features from the feature space at each tree node. The final class decision is made through the majority voting of the full ensemble of the trained trees. In this study, we deployed the off-the-self RF implementation from the R programming language v3.5.1 [30]. The number of trees grown ntree was set to 500 and the number of random features mtry for the growth of each decision tree in the ensemble was kept as its default value which is the square root of the total number of features, rounded down.
To determine the classification accuracy, RF performs internally a procedure resulting in an out-of-bag (OOB) error rate. RF works by bootstrapping a sample dataset from the original training data for every tree grown. From the original data, an approximate 37% become "out-of-bag" data due to the replacement strategy in the sampling. These OOB data are then parsed down and classified by the trees that are not trained with them to estimate the classification error by adding up all the discrete OOB errors. The OOB error rate is the complementary percentage of the overall classification accuracy of the RF classification and is a highly robust accuracy indicator. RF also provides a confusion matrix indicating the misclassifications for each class. This matrix allows us to compute the class user's and producer's accuracies as well as the overall classification accuracy [31].
Additionally, RF provides importance values to indicate the input features significance for the classification with two metrics in the R implementation: (1) the mean decrease in the Gini index of node impurity when a feature is split when a node is made and (2) the mean decrease of prediction when a feature is permutated. In this study, we display the MeanDecreaseAccuracy plot using R's ggplot2 [32]. It graphically represents the value of the MeanDecreaseAccuracy metric for each feature which is the difference between the prediction accuracy of the original OOB data and the prediction accuracy when the values of that feature are randomly permuted with the OOB data and predicted down the trained forest. The final mean difference in prediction errors for every feature is a measure of feature importance as it exhibits a decrease in accuracy when the feature is assigned random but realistic values. For unimportant features, the permutation should have minimal to no effect on the accuracy, whereas, for important variables, the accuracy should be significantly reduced.
For training of the RF classifiers for both the PBIA and GEOBIA methodologies, spatially representative and uniformly spread training sites for the three classes were delineated on the original mosaics. These training sites are common for both methodologies.

Jeffries-Matusita Distance
The spectral separability of the three classes was assessed by the distance between the random probability distributions within the feature space in pairs of classes. The metric used is the Jeffries-Matusita (J-M) distance [33,34] (Equation (1)) which considers the Bhattacharya (B) distance [35] (Equation (2)). J-M values range from 0 to 2, with 0 implying the two distributions are entirely correlated and thus the classes are spectrally inseparable, while the J-M upper asymptotic limit of 2 means a full non-correlation between classes and is considered as an indication of excellent class separability, making the J-M transformation more convenient compared to B which falls in the [0, +∞) range.
For each pair of classes C 1 and C 2 , which are two multivariate distributions, assuming data normality, with means µ 1 , µ 2 and covariance matrices σ 1 , σ 2 M C 1 ,C 2 is the root Mahalanobis distance [36] between the class means with respect to σ computed by Equation (4):

Pixel-Based Image Analysis (PBIA)
For the PBIA methodology, the RF classification uses the five reflectance bands (blue, green, red, red-edge, NIR) and the BRSR VI as input features. As a result, for 6 features, mtry = 2.
All the training areas were randomly sampled for a number of pixels that is proportional to the area of each training site. For each class, approximately 10,000 random pixels in total were employed for training, resulting in a robust training set with minimized spatially induced bias. The number of training areas and pixels for each field is shown in Table 3.

Pre-Segmentation Processing
To generate objects in the GEOBIA methodology optimally, we performed the following pre-segmentation processing steps for each reflectance band and the BRSR VI used in the multi-resolution algorithm, in order to have a more consistent segmentation:

1.
Set the pixels of the lowest and highest 2% values to the 2% and 98% limit pixel values based on the image histogram to minimize the number of outliers following [37].

2.
Normalize all pixel values in order to avoid features of greater value ranges dominating the ones in smaller ranges. A linear transformation similar to the linear minimum-maximum normalization as described in [38,39], is used to rescale pixel values to a new range between 0 and 100 using the following (Equation (5)) with a = 2% limit pixel value and b = 98% limit pixel value.

Segmentation
The fundamental step of a GEOBIA pipeline is the segmentation of the image. For very high-resolution UAS imagery, which usually has high spectral variability, GEOBIA can be challenging when trying to construct and parametrize the segmentation algorithm for generating meaningful objects. We performed the image segmentation with eCognition 9.4 (Trimble Inc., USA) using the multiresolution segmentation algorithm [40]. This algorithm is one of the most utilized in eCognition. It is an iterative clustering process that minimizes heterogeneity through a local optimization procedure [41]. It begins at the individual pixel level by merging bottom-up regions until convergence to a threshold that represents the object variance limit that is parametrized by the "Scale" parameter, which in turn is weighted by the "Shape" and "Compactness" parameters ranging from 0.1 to 0.9. The larger the Scale value, the higher the allowed variability within each segment, resulting in larger objects generation tolerating more deviation within the homogeneity rules. The Shape parameter determines the spectral information weight that one would like to give to the objects. The higher the Shape value, the less the influence of the spectral information on the segmentation. The Compactness parameter determines how well-defined objects will be with respect to the shape criterion in order to create clear edges. We applied a Scale value of 30, which was selected by a trial and error procedure as it is usually done in GEOBIA, to maintain small objects at the field borders in order to discriminate well between different vegetation types and to consider the natural vegetation transition at the borders. The Shape parameter was set to a value of 0.1 to use the full weight on the spectral information (color = 1-shape) for good discrimination between the different vegetation signatures. Finally, for the Compactness parameter, we used a value of 0.4 which is the median of its value range. The Compactness parameter does not have any significant impact on the objects because of the very low value of the Shape parameter. All the parameter values were kept constant for every field. Finally, all features were weighted by an equal value of 1, as we consider all the input features equally important. Table 4 shows the number of objects generated for each field and their mean area (m 2 ). Object Feature Generation The selected object features are related to spectral properties of the objects-but not their textural or spatial (size-shape) properties-because spectral differences are the most meaningful for object classification in land cover classes in the case of crops. The resulting vector file that has all the objects which were generated by the segmentation was exported as a georeferenced tiff file to be used in the following step that consists of generating object features from the original datasets. The following features were generated for each object from the BRSR VI and the blue, green, red, red-edge and NIR reflectance bands (Table 5). These features are related to the objects' spectral properties and are commonly used for spectral discrimination [42][43][44]: 1.
The mean reflectance or VI; 2.
The standard deviation (SD) of the reflectance or VI; 3.
The median reflectance or VI; 4.
The mean reflectance or VI calculated from the range of the 10th to the 90th percentiles of the pixel values distribution, removing outliers for more robust statistics; 5.
The SD of the reflectance or VI calculated from the range of the 10th to the 90th percentiles of the pixel values distribution.

Classification
For the GEOBIA RF classification, we used a total of 30 features, resulting in mtry = 5, and the forest was trained with all the objects that fall within the training polygons. The number of training objects per class and their mean area (m 2 ) are shown in Table 6. On average, the training objects represent 8.9% of the study site areas.

Vectorization
The final map with the field borders and areas is produced by inserting every classified image into a vectorization pipeline. For the PBIA classified image, the System for Automated Geoscientific Analyses (SAGA) Majority filter with a radius of 5 pixels [45] was first applied to the classified image in order to clear misclassifications and salt and pepper noise. For the GEOBIA classified image, the filtering step is omitted because the objects have the desired homogeneity and consistency. Afterward, the polygonize function of the GDAL library was employed to vectorize the PBIA majority-filtered image and the GEOBIA classified image. Finally, the field area is defined by extracting the polygon refined from holes that has the largest area, after smoothening the borders with vector buffering and debuffering.

Actual Field Boundaries and Areas
The accuracies provided by RF give only a comparison between the classified image and the training areas, but a better accuracy assessment should be made by comparing the field boundaries and areas to the actual ones. Actual field boundaries and areas can be measured in the field using GNSS equipment that records in situ border waypoints at regular intervals. Such equipment has a measurement accuracy that depends considerably on how many GNSS networks the equipment can receive and how many satellites are present for every measurement. The accuracy is also related to factors such as the surrounding environment (trees) and the weather conditions. Also, the method is quite expensive and time-consuming. An alternative is to manually delineate the field boundaries and areas on the RGB raster composite made with the UAS imagery, such as was done in our previous work [8]. Indeed, the RGB composite provides enough visual details because it has a pixel size of~8cm as it was made with the photogrammetrically stitched UAS images that have an excellent geolocation accuracy.

Area Goodness of Fit (AGoF)
The first accuracy metric of the method compares the manually-and machine-delineated areas. It is an area similarity measure that is called the area goodness of fit (AGoF) [46]. AGoF calculates the overlapping percentage between the manually-and machine-delineated areas for a given field by Equation (6): where: A is the manually delineated field area (ha); B is the machine-delineated field area (ha); C is the area of the intersection between the manually-and the machine-delineated crop polygons (ha); AC is the absolute value of the total area difference between A and C (ha): AC = |A − C|; BC is the absolute value of the total area difference between B and C (ha): BC = |B − C|.

Boundary Mean Positional Error (BMPE)
The second accuracy metric of the method compares for a given field the manual-and the machine-delineated boundaries. It is a positional similarity measure that is called the boundary mean precision error (BMPE) [8]. To compute BMPE, sequential geographical points are first sampled at a 0.5 m interval along the machine-delineated boundaries. Secondly, the minimum distance between each of these points and the manually delineated boundary is calculated. BMPE is finally the mean distance between the N machine-delineated boundary points and the manually delineated boundary polygon (Equation (7)): MinDist i (7) where: N = number of sample points from the machine delineated boundary; MinDist i = minimum distance between the i-th point of the machine-delineated boundary and the manually delineated boundary (m).

Jeffries-Matusita (J-M) Distance
For the PBIA classification, the J-M distances are computed for each pair of the bare soil, crop, and other vegetation classes, using the original feature space of the BRSR VI, blue, green, red, red-edge, NIR reflectance and the PBIA training areas (Table 7). On average, the J-M distances are very high, indicating very good class spectral separabilities. The lowest average (1.75) is between crop and other vegetation classes, with the lowest one being for the Barley3 field (1.32). The spectral separability for the Crop-Soil and Soil-Vegetation pairs are excellent, with average J-M distances higher than 1.96. For the GEOBIA classification, the J-M distances are computed for each pair of bare soil, crop, and other vegetation classes from the training objects with the blue, green, red, red-edge, NIR reflectance, and BRSR images ( Table 8). The feature space of the J-M distances is the mean, SD, and median from all the pixel values within each object and the mean and SD from the pixel values after removing their 10% lower and higher values. All three pairs of classes have excellent class separabilities, with a mean J-M distance higher than 1.99 and a J-M distance for each field higher than 1.96. The J-M distances for the GEOBIA classification are on average and for every individual field higher than those with the PBIA training areas, showing that a better classification can be potentially achieved with the GEOBIA classification and that the feature selection is very important for discriminating between the probability distributions of each pair of classes. Table 8. Jeffries-Matusita distance between the three classes in the case of the geographic object-based image analysis classification using the full GEOBIA feature list of Table 5.

Classification Accuracy
The OOB error rate from the RF is presented in Table 9 for each field in the case of the PBIA and GEOBIA classification. For both classifications, the lowest OOB error rate occurs for the Corn3 field, while the highest OOB error rate occurs for the Barley2 field. Table 9. Random forest pixel-based image analysis and geographic object-based image analysis out-of-bag error rates (%) as a function of the field. The related confusion matrices including the user's, producer's, and overall classification accuracies (UA, PA, OA) and errors of omission and commission (EO, EC) are shown for each field in Table 10 for the PBIA RF classifier applied to the blue, green, red, red-edge, NIR reflectance and BRSR VI features and in Table 11 for the GEOBIA RF classifier applied to objects' generated features for the blue, green, red, red-edge, NIR reflectance and the BRSR VI. In both cases, the highest overall classification accuracies and the lowest OOB error rates were observed for the Corn3 field. This is the opposite in the case of the Barley2 field, mainly because of confusion between the crop and the bordering vegetation due to a vegetation transition and mixture at the field borders. On average, the overall classification accuracy with the PBIA classification was slightly lower than with the GEOBIA classification (97.06% versus 97.49%). This is also the case for every individual field. Table 10. Confusion matrices and associated class user's accuracies, producer's accuracies, overall accuracies, errors of omission and errors of commission obtained by applying the PBIA random forest classifier to the blue, green, red, red-edge, near-infrared reflectance and blue-red simple ratio images. The bold diagonal number elements are the correctly classified pixels for each class.

Random Forest (RF) Variable Importance
The variable importance plots show by decreasing order each feature's MeanDecreaseAccuracy value for the RF classification. These values have no physical meaning apart from a feature importance comparison metric within the feature space. For the PBIA-RF classification, the red-edge and NIR reflectance images appear to be the most important input features (Figure 4). For the GEOBIA-RF classification, the red and NIR reflectance images appear to be the most important input features ( Figure 5).

Field Area and Border Maps
Both PBIA ( Figure 6) and GEOBIA (Figure 7) methodologies show very satisfying results in regard to the final border vector products. The field border maps after the vectorization pipelines show a good fit comparing the machine-and the manually delineated borders. The GEOBIA classification field borders are cleaner and more robust. The PBIA misclassifications between all classes have to be dealt with using greedy expensive algorithms for noise removal. For PBIA, the border comparison displays the best result for Corn1 and worst for Oat ( Figure 6). For GEOBIA, the best result is achieved for Corn1 and the worst for Barley1 (Figure 7). A visual comparison between the manual and machine delineated field borders allows the detection of several minor factors that lead to misclassifications and border skewness explaining most of the divergence between the machine-and the manually delineated borders. For all the crop fields at their borders, such factors are surrounding tree canopies and canopy shadows on top of the crops and transitional mixture of wild weeds and crops.
Both PBIA ( Figure 6) and GEOBIA (Figure 7) methodologies show very satisfying results in regard to the final border vector products. The field border maps after the vectorization pipelines show a good fit comparing the machine-and the manually delineated borders. The GEOBIA classification field borders are cleaner and more robust. The PBIA misclassifications between all classes have to be dealt with using greedy expensive algorithms for noise removal. For PBIA, the border comparison displays the best result for Corn1 and worst for Oat ( Figure 6). For GEOBIA, the best result is achieved for Corn1 and the worst for Barley1 (Figure 7). A visual comparison between the manual and machine delineated field borders allows the detection of several minor factors that lead to misclassifications and border skewness explaining most of the divergence between the machine-and the manually delineated borders. For all the crop fields at their borders, such factors are surrounding tree canopies and canopy shadows on top of the crops and transitional mixture of wild weeds and crops.       Table 12 compares the AGoF (in %) between the PBIA and GEOBIA classifications. There is no important difference between both classifications for the mean AGoFs (98.91% and 98.78%, respectively). Concerning BMPE (in m), it is on average slightly higher with the GEOBIA classification (0.76 m) than with the PBIA classifications (0.68 m), indicating that PBIA is a slightly more accurate method on average (Table 13). For most of the fields, the difference is less than 0.1 m and thus is not important.   Table 12 compares the AGoF (in %) between the PBIA and GEOBIA classifications. There is no important difference between both classifications for the mean AGoFs (98.91% and 98.78%, respectively). Concerning BMPE (in m), it is on average slightly higher with the GEOBIA classification (0.76 m) than with the PBIA classifications (0.68 m), indicating that PBIA is a slightly more accurate method on average (Table 13). For most of the fields, the difference is less than 0.1 m and thus is not important.

Discussion
We performed PBIA and GEOBIA with RF classification to delineate crop field boundaries and extract field areas using UAS multispectral imagery acquired over corn, oat, and barley vegetated fields. In both cases, we achieved an average classification accuracy higher than 97%, a mean AGoF greater than 98%, and a mean BMPE lower than 0.8 m. All these values are on the same order of magnitude as Vlachopoulos et al. [8] who applied the mean-shift clustering and the RF classifier to perform field areas and boundaries delineation from UAS imagery acquired over bare soil fields. The average overall classification accuracies in this study were higher than De Luca et al. [9] (89-97.6%) who applied RF and SVM to RGB and NIR UAS imagery and on the same order of magnitude as Laliberte and Rango [10] (95-100%), who used GEOBIA with decision tree analysis on UAS RGB imagery soil, shrubs, and grass demarcation. The lower accuracies we obtained with the Barley1 and Oat fields are mainly due to the surrounding tree shadows and tree canopies on top of the crops, the transitional mixture of wild weeds and crops at the field boundaries, and the sparsely crop-planted areas at the borders.
Since the AGoF and BMPE values do not differ too much between the two methodologies, this indicates that the methods can be considered equally strong. However, each methodology has its advantages and disadvantages. The PBIA methodology has fewer steps but is more timeand computing resource-consuming mainly because of the majority filtering step. For the GEOBIA methodology, such filtering is not needed because the classified objects are homogenous and do not have individual pixels, as shown in Figure 7. Because the GEOBIA methodology classifies objects but not pixels, the number of training samples are much lower than those of PBIA, thus resulting in a quicker RF training and classification, even if the feature space for the GEOBIA method has a higher number of input features than that of the PBIA method. The feature space in each methodology is related to the fact that we used either pixels or objects. In PBIA, the features are five reflectance bands and one VI. For GEOBIA, the feature space is the mean, median, standard deviation of the pixel values and the mean and standard deviation of the pixel values in the range of the 10th and 90th percentile associated with the objects for each reflectance band and the BRSR VI. As a result, a direct comparison cannot be done for the two feature spaces. However, it is worth mentioning that we use the same reflectance bands and VI both for PBIA and GEOBIA, thus making the original information a common ground in both methodologies. The high number of features used in GEOBIA is needed firstly to add discriminatory information to each object and secondly because the training dataset in GEOBIA is limited as it is related to a limited number of objects. In order to construct a robust training dataset and an informative feature space, one needs to have multiple spectral characteristics of the objects that allow having the appropriate number of input features for the classification, by excluding outliers.
The GEOBIA method utilizes the eCognition software for image segmentation, which is a commercial product that may not always be available. It would be interesting to explore other segmentation methods such as the open-source Orfeo Toolbox (OTB) [47] and compare their performance.
Both methods rely on random forests as the classifier. RF is an open-source ML algorithm that is easily implemented, cost-effective, and conveniently replicated and generalized to other crops and fields. The training data for the RF algorithm have a very high level of randomness for both PBIA and GEOBIA due to the RF methodology used for feature selection and bootstrapping. Specifically, training samples for PBIA are randomly selected from all the training areas. This randomness reduces the bias from both the user and algorithm sides making both methodologies scalable. The proposed methodologies are better performing in comparison to the convolutional neural network method proposed by Crommelinck et al. [6] as the latter requires distinct features such as roads and water bodies to detect cadastral plot boundaries. Our methodologies do not need multitemporal series of images such as by DeLuca et al. [9] and Persello et al. [1]. The PBIA methodology used in this study does not need commercial software, whereas Fetai et al. [7] employ ENVI which may be inaccessible due to commercial costs.
The time to perform both methodologies depends on the computational power of the computer used and the size of the UAS imagery. For example, to apply the proposed methodologies, computational time might be in the order of seconds or a few minutes with a standard server-driven solution and the order of minutes with a personal computer. For the fields we tested, the proposed methodologies were more efficient than the manual delineation. Additionally, manual delineation is quite demanding as fields can be very irregular. Also, it is highly laborious to outline the borders of a crop field with hundreds or thousands of geographical points. Another advantage of the proposed methodology is that it can be used on any number and size of fields. It is therefore service-driven, scalable, and refraining from user bias.

Conclusions
This study presents a method to delineate field areas and boundaries from UAS multispectral images that were acquired over 7 vegetated fields having various crops (barley, corn, and oat). The imagery was classified using the non-parametric supervised classifier random forests which was applied either to the raw images (pixel-based image analysis or PBIA) or the images that were previously segmented with the multiresolution segmentation algorithm implemented in eCognition (geographic object-based image analysis, or GEOBIA). Both methodologies classify the images in three classes (soil, crop, other vegetation) using the blue, green, red, red-edge, NIR reflectance bands and the BRSR VI. For PBIA, the RF classification used the class statistics derived from the training areas. The GEOBIA classification used the following statistical parameters for each object from the same training areas: mean reflectance or VI; standard deviation (SD) of the reflectances or VIs; median reflectance or VI; mean reflectance or VI after removing the lowest and highest 10% reflectance or VI pixel values; SD of the reflectances or VIs after removing the lowest and highest 10% reflectance or VI pixel values. The accuracy of both methods was assessed using the classification accuracies and two metrics following Vlachopoulos et al. [8]: area goodness of fit (AGoF) for the field area and boundary mean positional error (BMPE) for the crop borders. Both classifications performed exceptionally well with an average accuracy higher than 97%, leading to a mean AGoF greater than 98% and a mean BMPE lower than 0.8 m. Also, both classification methods rely on random forests, an open-source ML algorithm that is easily implemented, highly efficient, and can be replicated and generalized to other crops and fields.
Our excellent results indicate only some minor divergences between the photo-interpreted and machine-delineated field areas and boundaries which depend on some field-specific characteristics, such as tree shadows, tree canopies, the transitional mixture of vegetation and crop plants at the field boundaries, and sparsely cropped areas at the borders. Further research is needed to investigate the effect of these factors on the results. While there were no significant differences between both classification methods in terms of AGoF and BMPE values, the GEOBIA classification shows more promising results than the PBIA classification, because the first method is faster to perform, gives higher classification accuracies and class separability, does not need a post-classification filtering process and it is less intensive in terms of data processing. However, compared to the PBIA classification, it depends on the use of the eCognition software which is not free. Future research is needed to perform GEOBIA segmentation with open-source segmentation algorithms. Also, the GEOBIA classification we tested uses a single segmentation step and further work is needed to assess whether an eCognition hierarchy of processes can refine and simplify the segmentation and thus the classification. Our results were obtained using multispectral UAS imagery acquired under specific flight altitudes and clear sky conditions as well as over particular crops (barley, oat, corn). Further work is needed to test UAS imagery acquired at a different flight altitude that leads to a different GSD. It will also be interesting to test UAS imagery acquired under diverse weather conditions and on other crops. Finally, the scalability of the proposed methodologies can be tested and generalized with sufficient training data from other crops and land-cover classes. Our study is an important step towards the development of an operational and efficient methodology for the automatic delineation of field boundaries and areas from UAS multispectral data with a combined ML pipeline and vectorization steps, to quickly acquire field areas and boundaries for various applications in precision agriculture, for example, determining crop insurance premiums.