Comparison Of Algorithms For Agave Detection On Unmanned Aerial Vehicle Images

In this study, six supervised classification algorithms were compared. The algorithms were based on cluster 40 analysis, distance, deep learning and object-based image analysis. Our objective was to determine which of 41 these algorithms has the highest overall accuracy in both detection and automated estimation of agave cover in 42 a given area to help growers manage their plantations. An orthomosaic with a spatial resolution of 2.5 cm was 43 derived from 300 images obtained with a DJI Inspire 1 unmanned aerial system. Two training classes were 44 defined: 1) sites where the presence of agaves was identified, 2) “absence”; where there were no agaves but 45 other plants were present. The object-oriented algorithm was found to have the highest overall accuracy (0.963), 46 followed by the support-vector machine with 0.928 accuracy and the neural network with 0.914. The algorithms 47 with statistical criteria for classification were the least accurate; Mahalanobis distance = 0.752 accuracy and 48 minimum distance = 0.421. We recommend that agave plantation managers use drones for their efficiency and 49 speed. We further recommend that the object-oriented algorithm be used, because in addition to having the 50 highest overall accuracy for the image segmentation process, it yields parameters that are useful for estimating 51 the coverage area, size, and shapes, which can aid in better selection of agave individuals for harvest. 52

Twenty-five species of agave have been recorded as being used to produce mezcal. The most used species are 61 wild; e.g., A. durangensis, A. salmiana, while some are cultivated; e.g., A. angustifolia (Carrillo-Trueba, 2007).

62
In the past two decades, the demand and overexploitation of agave for mezcal production has caused wild 63 populations to decline rapidly. One of the main reasons for the decline is that the plants are harvested before

80
As mentioned above, detection and counting has only been carried out for blue agaves using deep learning 81 algorithms, which require robust knowledge of image processing by users. This study therefore set out to 82 compare supervised classification algorithms for first-time detection of cenizo agaves (Agave duranguensis), 83 which are in danger of extinction. We compared algorithms based on statistical rules that are commonly used 84 in remote sensing, and deep learning algorithms that are available and contain flowcharts in commercial and 85 free software.

89
The Sampling was carried out in the Ejido Nombre de Dios, Durango, that is located at coordinates 23 ° 36 'and

152
A total of 860 random points were sampled, with at least 430 points for each class (Goodchild, 1994). The 153 following classes were considered: class a is presences of agaves mescaleros with 430 sites, class b is absences 154 of agaves with sites. These data were contrasted with the category to which each training pixel belongs, 155 corresponding to Georeferenced sites (Datum WGS-84, 13N) obtained in the field in September 2020.

156
Classification methods

157
Statistical rule-based algorithms. We used three types. The first was minimum distance, which uses the mean 158 vectors of each region of interest (ROI) and calculates the euclidean distance from each unknown pixel to the 159 mean vector for each class. Pixels are put into the closest ROI class unless the user specifies standard deviation 160 or distance thresholds, in which case some pixels may be unclassified if they do not meet the specified criteria.

161
The second algorithm was Mahalanobis distance. This is a direction-sensitive distance classifier that uses 162 statistics for each class. It assumes all class co-variances are equal and therefore is a faster method. The third 163 algorithm uses maximum likelihood. This method assumes that the statistics for each class in each band are 164 normally distributed and calculates the probability that a given pixel belongs to a specific class. Unless a 165 probability threshold is selected, all pixels are classified. Each pixel is assigned to the class with the highest 166 probability. The ROIs correspond to the sites with presence or absence of mezcal agaves (Richards, 1999).

167
Deep learning algorithms. We used two types. The first was a back-propagation artificial neural network 168 (BPNN) (Tan and Smeins, 1996). BPNN is widely used because of its structural simplicity and robustness in 169 modeling non-linear relationships. The first step in BPNN supervised classification is to enter the input layer,

175
The second type was a support vector machine algorithm. This method is based on statistical learning theory, 176 and often yields good classification results from complex and noisy data. It separates the classes with a decision surface that maximizes the margin between the classes. The surface can be called the optimal hyperplane, and 178 the data points closest to the hyperplane are called support vectors. The support vectors are the critical elements 179 of the training set. We used the ENVI implementation of SVM, which uses the pairwise classification strategy presents an approximate 95% confidence interval.

230
The BPNN and SVM deep learning algorithms improved the classification, yielding overall precision levels 231 greater than 0.90 (Table 1) In the estimation of errors in the counts of presence and absence of agaves using the Mahalanobis distance and 245 maximum likelihood algorithms, the weight (Wi) assigned to absences was greater than 0.60, which affected 246 estimation of the coverage both of area with agaves and area without the plant ( Table 2). The Wi were low in 247 the agave-present class in the deep learning and OBIA algorithms (Table 2). This makes sense, because agaves 248 cover less than 20% of the area studied and therefore the estimate of agave coverage improves notably ( Figure   249 2