A Comparative Analysis of Machine Learning with WorldView-2 Pan-Sharpened Imagery for Tea Crop Mapping

Tea is an important but vulnerable economic crop in East Asia, highly impacted by climate change. This study attempts to interpret tea land use/land cover (LULC) using very high resolution WorldView-2 imagery of central Taiwan with both pixel and object-based approaches. A total of 80 variables derived from each WorldView-2 band with pan-sharpening, standardization, principal components and gray level co-occurrence matrix (GLCM) texture indices transformation, were set as the input variables. For pixel-based image analysis (PBIA), 34 variables were selected, including seven principal components, 21 GLCM texture indices and six original WorldView-2 bands. Results showed that support vector machine (SVM) had the highest tea crop classification accuracy (OA = 84.70% and KIA = 0.690), followed by random forest (RF), maximum likelihood algorithm (ML), and logistic regression analysis (LR). However, the ML classifier achieved the highest classification accuracy (OA = 96.04% and KIA = 0.887) in object-based image analysis (OBIA) using only six variables. The contribution of this study is to create a new framework for accurately identifying tea crops in a subtropical region with real-time high-resolution WorldView-2 imagery without field survey, which could further aid agriculture land management and a sustainable agricultural product supply.


Introduction
In recent years, due to global warming, many regions of the Earth have suffered from drastic climate change, which has increased the frequency of extreme climate events. Not only does this cause certain inconveniences for the lives of human beings, but it also threatens our lives and property. The growth of crops has an intimate relationship with the climate, so any climatic change could impact the amount and the quality of crop production [1][2][3][4][5][6][7][8][9][10][11]. In the past, in order to control the production surface area and the amount of production, governments would generally monitor the sensitive crops with prices that fluctuate according to the weather, disasters, or public preferences so as to prevent price collapses, which could cause financial damage to farmers. Therefore, it is important to have a comprehensive understanding of the acreage of crop production and the market situation of various crops in order to propose important contingency measures in response to the economic losses and potential food crises caused by extreme climate phenomena. In the sub-tropical region in Taiwan, for example, tea is a sensitive crop of high economic value and a featured export product. About 3306 tons of teas of various special varieties are exported annually, and the number is still growing. According to the United Nations Food and Agriculture Organization (FAO), the global production of tea grew from 3.059 million tons in 2001 to 4.162 million tons in 2010, while the global area planted with tea grew According to the abovementioned reasons and background, the main objectives of this study include: (1) to investigate the classification accuracy of a combined workflow applied on WorldView-2 imagery to identify tea crop LULC; (2) to compare the capabilities of several ordinary LULC classification algorithms (e.g., maximum likelihood (ML), logistic regression (LR) random forest (RF) and support vector machine (SVM) based on pixel and object-based domains) in identifying tea LULC; and (3) to explore the potential of using high spatial/spectral/time resolution WorldView-2 imagery to identify tea crop LULC classification, and the details about image segmentation and feature extraction will also be discussed. The main content of this paper is divided into four major sections. Sections 1 and 2 include descriptions and introductions of the study, introductions to on site investigation and satellite imagery data sets, frameworks of feature extraction and classification developed for this study. Section 3 contains the main findings and breakthroughs of this study, followed by their discussion. The final part consists of the conclusions.

Study Area and Workflow
The study area is located in part of Nantou County in central Taiwan (between 120.620°E to 120.683°E and 23.8492°N to 23.8894°N). The total area is about 263 hectares. According to the 2014 agricultural statistic data of Taiwan, the annual tea production of this area is about 2.3% of the total output. The climate of this region is sub-tropical, with average temperatures of 22~25 °C and an annual rainfall of 1500~2100 mm. The terrain is mainly a flat plain near low slope hills. This region contains typical agricultural villages with abundant and minimally changed     (2) other crops (e.g., pineapple and betel palm (Areca catechu)); (3) water and ponds; (4) buildings and roads; and (5) forest (evergreen broadleaf forest). Other crops in the study area were mainly pineapple and betel nut. We re-categorized these five LULC types into "tea" and "non-tea" LULC by integrating all land features except the tea areas. The results of the field survey showed that there are around 33 hectares of tea and 230 hectares of non-tea regions in our study area.   (2) other crops (e.g., pineapple and betel palm (Areca catechu)); (3) water and ponds; (4) buildings and roads; and (5) forest (evergreen broadleaf forest). Other crops in the study area were mainly pineapple and betel nut. We re-categorized these five LULC types into "tea" and "non-tea" LULC by integrating all land features except the tea areas. The results of the field survey showed that there are around 33 hectares of tea and 230 hectares of non-tea regions in our study area. (2) other crops (e.g., pineapple and betel palm (Areca catechu)); (3) water and ponds; (4) buildings and roads; and (5) forest (evergreen broadleaf forest). Other crops in the study area were mainly pineapple and betel nut. We re-categorized these five LULC types into "tea" and "non-tea" LULC by integrating all land features except the tea areas. The results of the field survey showed that there are around 33 hectares of tea and 230 hectares of non-tea regions in our study area.

WorldView-2 Imagery
The main LULC of the study area are tea crops and pineapples, but it was not easy to do artificial interpretation even with VHR orthogonal aerial photographs due to the similar arrangements of the plants. For this reason, WorldView-2 imagery with both high spatial and spectral resolution was applied as a main material. The WorldView-2 imagery satellite (DigitalGlobe, Inc., Westminster, CO, USA) is a commercial satellite with eight multispectral bands and very high resolution. The average amount of time needed to revisit and acquire data for the exact same location is 1.1 day, and the spatial resolution is 0.46~0.52 m (panchromatic band, 450~800 nm) and 1.84~2.08 m (multispectral bands). The eight multispectral bands contained coastal blue (400~450 nm), blue (450~510 nm), green (510~580 nm), yellow (585~625 nm), red (630~690 nm), red edge (705~745 nm), NIR1 (770~895 nm) and NIR2 (860~1040 nm). The added spectral diversity of WorldView-2 satellite such as coastal blue, yellow and red edges provides the ability to perform detailed environmental and landscape change detection better than with other satellite images (e.g., Formosat-2, SPOT6, Kompsat) and DMC aerial photographs with only four bands or less. We applied cloud-free imagery (<5%) of the study area as the original material. The original digital number (DN) value of the imagery was converted to top-of-atmosphere spectral reflectance [66]. The imagery was also ortho-rectified and co-registered to the two degree zone Transverse Mercator projection coordinate system with a root mean square error of less than 0.5 pixels per image. The time of image acquisition for this study was acquired on 22 February 2015 corresponding to the date of ground reference data (4-5 March 2015). It worth mentioned that tea crops within study area are stable in LULC, farm shape arrangement, and not leafing out all years ( Figure 4). Therefore, if the tea LULC still remained, the image acquisition and interpretation would not be limited or affected by season or date within a period due to stable vegetation spectral dynamics. This is beneficial to ground verification and long term monitoring. The main LULC of the study area are tea crops and pineapples, but it was not easy to do artificial interpretation even with VHR orthogonal aerial photographs due to the similar arrangements of the plants. For this reason, WorldView-2 imagery with both high spatial and spectral resolution was applied as a main material. The WorldView-2 imagery satellite (DigitalGlobe, Inc., Westminster, CO, USA) is a commercial satellite with eight multispectral bands and very high resolution. The average amount of time needed to revisit and acquire data for the exact same location is 1.1 day, and the spatial resolution is 0.46~0.52 m (panchromatic band, 450~800 nm) and 1.84~2.08 m (multispectral bands). The eight multispectral bands contained coastal blue (400~450 nm), blue (450~510 nm), green (510~580 nm), yellow (585~625 nm), red (630~690 nm), red edge (705~745 nm), NIR1 (770~895 nm) and NIR2 (860~1040 nm). The added spectral diversity of WorldView-2 satellite such as coastal blue, yellow and red edges provides the ability to perform detailed environmental and landscape change detection better than with other satellite images (e.g., Formosat-2, SPOT6, Kompsat) and DMC aerial photographs with only four bands or less. We applied cloud-free imagery (<5%) of the study area as the original material. The original digital number (DN) value of the imagery was converted to topof-atmosphere spectral reflectance [66]. The imagery was also ortho-rectified and co-registered to the two degree zone Transverse Mercator projection coordinate system with a root mean square error of less than 0.5 pixels per image. The time of image acquisition for this study was acquired on 22 February 2015 corresponding to the date of ground reference data (4-5 March 2015). It worth mentioned that tea crops within study area are stable in LULC, farm shape arrangement, and not leafing out all years ( Figure 4). Therefore, if the tea LULC still remained, the image acquisition and interpretation would not be limited or affected by season or date within a period due to stable vegetation spectral dynamics. This is beneficial to ground verification and long term monitoring. As for the resolution of WorldView-2 imagery, the intervals between every tea plantation were not clear in the 2 m resolution MS bands in this study. Therefore, a pan-sharpening process was added to improve the spatial resolution of the MS bands from 2 m to 0.5 m. The pan-sharpened archives had enhanced interpretability in inferring forms and shapes of tea crop LULC. Each band of the pan-sharpened imagery was then standardized by using mean and standard deviation. The reason is that the reflectance value range of the WorldView-2 image varies between each band and makes the input variables contribute unequally to the PBIA and OBIA.

Spectral Variability Analysis
The common scientific challenge of mapping vegetation with very high spatial resolution imagery is to distinguish between different vegetation classes due to the superimposition of spectral signatures. In order to test the LULC classification capability of eight MS band WorldView-2 imagery, and the difference or similarity of tea crops in different season, spectral separability analysis was undertaken as the initial step. Figure 4 presents the box-whisker-plots of spectral variability of the training pixels of five basic LULC classes. Reflectance of buildings and roads in NIR1, NIR2, coastal As for the resolution of WorldView-2 imagery, the intervals between every tea plantation were not clear in the 2 m resolution MS bands in this study. Therefore, a pan-sharpening process was added to improve the spatial resolution of the MS bands from 2 m to 0.5 m. The pan-sharpened archives had enhanced interpretability in inferring forms and shapes of tea crop LULC. Each band of the pan-sharpened imagery was then standardized by using mean and standard deviation. The reason is that the reflectance value range of the WorldView-2 image varies between each band and makes the input variables contribute unequally to the PBIA and OBIA.

Spectral Variability Analysis
The common scientific challenge of mapping vegetation with very high spatial resolution imagery is to distinguish between different vegetation classes due to the superimposition of spectral signatures. In order to test the LULC classification capability of eight MS band WorldView-2 imagery, and the difference or similarity of tea crops in different season, spectral separability analysis was undertaken as the initial step. Figure 4 presents the box-whisker-plots of spectral variability of the training pixels of five basic LULC classes. Reflectance of buildings and roads in NIR1, NIR2, coastal blue wavelengths and reflectance of water and ponds in all red edge wavelengths were significantly different from other classes. Besides, in most bands, tea LULC, forest canopies, and other crops (pineapple crop and betel palm) showed similar spectral variability in the blue, green, yellow, red bands and normalized difference vegetation index (NDVI, ranges from minus one (´1) to plus one (+1)), and little difference in NIR1 and NIR2 bands. In other words, tea farm patches cannot be definitely distinguished from non-tea crops if we only use eight MS bands and NDVI without other information.
As for the evaluation of tea spectral dynamics, usually, when plants are harvested, they have a physiological response characterized by intensive vegetation growth and high primary production to restore the area that was eliminated. This frequently translates into a displacement of the red edge position towards longer wavelengths. However, Figure 5 shows that the mean standardized reflectances of tea were almost the same in each band between 18 October 2012 and 22 February 2015. The same appearance was noted for pineapple as well. For this reason, we believed that the tea LULC and tea spectral dynamics are relatively stable. blue wavelengths and reflectance of water and ponds in all red edge wavelengths were significantly different from other classes. Besides, in most bands, tea LULC, forest canopies, and other crops (pineapple crop and betel palm) showed similar spectral variability in the blue, green, yellow, red bands and normalized difference vegetation index (NDVI, ranges from minus one (−1) to plus one (+1)), and little difference in NIR1 and NIR2 bands. In other words, tea farm patches cannot be definitely distinguished from non-tea crops if we only use eight MS bands and NDVI without other information.
As for the evaluation of tea spectral dynamics, usually, when plants are harvested, they have a physiological response characterized by intensive vegetation growth and high primary production to restore the area that was eliminated. This frequently translates into a displacement of the red edge position towards longer wavelengths. However, Figure 5 shows that the mean standardized reflectances of tea were almost the same in each band between 18 October 2012 and 22 February 2015. The same appearance was noted for pineapple as well. For this reason, we believed that the tea LULC and tea spectral dynamics are relatively stable.

Generation of Other Input Variables
This study assumed that the classification accuracy of tea and non-tea LULC separation would be increased if more sub-class bands or indices were involved in the procedure. Thus, eight WorldView-2 bands treated with pan-sharpening and standardization, 64 GLCM texture indices derived from these eight bands, and eight principal components, for a total of 80 variables, were used as the input variables. In regard to the spectral variables, first, the multispectral images of the eight bands of WorldView-2 and the panchromatic images of the single band were treated with fusion using hyperspherical color spaces (HCS) [67]; this is followed by standardization which would be more convenient for the subsequent feature selection by using logistic regression (LR) analysis, which is ORIGIN 1-8 in Table 1.

Generation of Other Input Variables
This study assumed that the classification accuracy of tea and non-tea LULC separation would be increased if more sub-class bands or indices were involved in the procedure. Thus, eight WorldView-2 bands treated with pan-sharpening and standardization, 64 GLCM texture indices derived from these eight bands, and eight principal components, for a total of 80 variables, were used as the input variables. In regard to the spectral variables, first, the multispectral images of the eight bands of WorldView-2 and the panchromatic images of the single band were treated with fusion using hyperspherical color spaces (HCS) [67]; this is followed by standardization which would be more convenient for the subsequent feature selection by using logistic regression (LR) analysis, which is ORIGIN 1-8 in Table 1.
The local mean grey level value in a given area. This measure helps distinguish the spectral difference between tea and other crop fields.
A measure of heterogeneity. Variance increases when the grey level values differ from their mean. Tea crop fields usually show lower variance compared with other crops.
Measures the linear dependency of grey levels of neighboring pixels. High contrast represents heavy textures. Tea crop fields usually show lower contrast compared with other crops.
GLCM Dissimilarity 1-8 Defines the variation of grey level pairs in an image. It is the closest to Contrast with a difference in the weight. Contrast will always give slightly higher values than Dissimilarity.
GLCM Homogeneity 1-8 Measures the level of variation in a given area. High homogeneity refers to textures that contain ideal repetitive structures. Tea crop fields usually show higher homogeneity compared with other crops.
measure of grey level linear dependence between the pixels at the specified positions relative to each other. Tea crop fields usually show higher correlation compared with other crops.
easures the level of chaos in a given area. A completely random distribution would have very high entropy because it represents chaos. Solid tone image would have an entropy value of 0. Tea crop fields usually show lower entropy compared with other crops.
Measures the textural uniformity that is pixel pair repetitions. High ASM values occur when the grey level distribution is constant or periodic. Tea crop fields usually show higher ASM values compared with other crops.
In addition, PC 1-8 was generated from the analysis of ORIGIN 1-2 in the principal component analysis. In regard to the spatial characteristics, the texture indices adopted in this study consisted of mean, variance, contrast, dissimilarity, homogeneity, correlation, entropy, angular second moment in GLCM. These indices are supposed to be helpful for distinguishing the structure differences between tea and other crop fields, such as the fields of pineapple and betel palm. With regards to plant structure, the canopies of tea trees have a more compact structure compared with other crops. In addition, although the tea trees and other crops in the study area are mostly planted line by line, the canopies of tea trees usually cover most of the soil surface. Therefore, tea crop fields show smoother texture than other crop fields. The relationship between the above ecological meaning and the GLCM indices along with the formulas of each of the index are shown in Table 1. Among them, i is the row number, j is the column number; p i,j is the normalized frequencies at which two neighboring pixels separated by a constant shift occur in the WorldView-2 imagery; N is the number of gray levels presented in the imagery.
With the purpose of proving that the increase of "sub-categories" or "factors" could enhance the LULC classification, this study adopted four spectral distance measurements: Euclidean, divergence, transformed divergence (TD), Jefferies-Matusita (J-M), to calculate the degree of separation of the training samples of tea trees and non-tea trees (forest, building and road, water bodies and ponds and other non-tea LULC) in the spectrum. In the TD assessment, 2000 was the best categorical resolution. Basically, above 1900 represents a good categorical resolution. The closer the selected sample from the training area is to 0, the worse the categorical separation efficiency is, so training data shall be re-selected. J-M has a similar concept towards TD; the closer to 1414, the better the separability is ( Table 2). The results showed that the larger the increase of sub-categorical bands or indicators is, the more effective it is to increase the separability between tea and non-tea farmland. This study subsequently investigated which sub-categorical band was the most helpful.

Feature Selection-LR Analysis
We applied LR analysis, a log-linear model suitable for quantitative analysis, in two different stages of this research. In feature selection stage, LR analysis was a filter to reduce the dimension of the data from 80 to less than 80. The next part we applied LR analysis was to evaluate the best LULC classifier, and LR model was one of the classifiers.
LR analysis is useful when the dependent variable is categorical (e.g., yes or no, presence or absence) and the explanatory (independent) variables are categorical, numerical, or both [68][69][70]. Generally speaking, a LR model can be expressed in the following forms [71]: In Equation (1), P`y i " 1˘means the conditional probability of landslide changes; x i is the vector of independent variable; if p " P`y i " 1˘, the above equation can be revised to: In Equation (2), p is the occurrence probability; α is the intercept; and β 1 , β 2 , ... β n are the regression coefficients of the input variables x 1 , x 2 , ... x n . In order to interpret the occurrence probability, one has to use the coefficients as a power to the natural log (e). The result represents the corresponding coefficient of each responding independent variable is the effect on odds (tea and non-tea occurrence), and when β n = 0 it means that the odds will not change. In this way, we can sum up the main factors related to the land cover. Last, Cox & Snell test, Omnibus test, Hosmer and Lemeshow test and odds ratio are used for the LR model as to achieve the assessments of the explanatory power of the factors and the model fit. We applied SPSS for calculating LR here, and all input row data (.txt) were generated from original layer files (.grd) by using the conversion tools of ArcGIS software.

Classification Procedure
The classification in this study basically used PBIA and OBIA mainly because the high spatial resolution of pan-sharpened WorldView-2 0.5 m images may present high spectral heterogeneity of the surface features, and the LULC classification may be affected if one only includes PBIA [72]. The difference of OBIA and PBIA is that the image shall be segmented before the classification and the objects are created; the establishment of objects from image segmentation mainly uses the heterogeneity, shape and boundary of surface features to define a object. This study segmented the images with multi-scale segmentation by using the eCognition image processing software concept. During the image segmentation, the scale, shape and compactness were the three parameters which produced the most reasonable results. The reasonability defined in this study refers to: (1) the drawing of the tea tree plantation in the study area using the boundaries; and (2) the object is still complete and does not include more than one kind of LULC classification after the image segmentation. This study used ORIGIN variable layers as the image segmentation basis. The best parameters obtained were 160 in scale, 0.1 in shape and 0.5 in compactness within the study area. After obtaining the reasonable segmented results, a classification of the LULC for the segmented images follows. PBIA classifies LULC based on the value of the single pixel, yet the OBIA would assign the LULC classes with the characters of homogeneous segmented features.
In both PBIA and OBIA, the aforementioned 80 variables were a huge amount which would cause a great burden to the system operation. For this reason, LR model was applied to reduce data dimensions before putting variables into ML, LR, RF, and SVM classifiers. The area of training samples for tea and non-tea was 6.2 and 16.1 hectares separately for PBIA classifiers, and 36 and 132 segments for OBIA classifiers. The following section provides a detailed description of tea tree classification with RF and SVM.

Classification Procedure-RF Classifier
We employed a newly developed RF algorithm as one of the LULC classifiers in this study. The main concepts of the RF algorithm are based on a non-parametric decision-tree, and it allows various kinds of inputs with different scales and types, including data sources recorded on different measurement scales [73]. This advantage allows us to use three different sets of variables involving original spectral bands and their derivative GLCM and PCA. The process of RF starts by selecting many bootstrap (with replacement) observations from the original data. The data source of a single classification tree comes from each bootstrap sample, but the predictor variables decrease with the binary partition classification tree at each node. The observations eliminated from bootstrapping at each decision-tree are called the "out-of-bag" (OOB) observations. The error rate for the entire RF classifier was generated from the comparison between the observed classification and the predicted classification for all observations derived from the out-of-bag vote tallies [74]. We thus built a separate RF model for three sets of predictors:

1.
Two control parameters were set: one for the amount of the decision trees (n tree ), while the other was for the number of characteristic parameters (m try ); 2.
We took the original training samples and then put them back with a bootstrap method; the best number and sample of decision tree Xi (i = number of iterations) were randomly collected from the original samples (dataset) X. Each sample stage contained about one third of X. The element that was not included in the sample was called OOB; 3.
Every decision tree would have a complete growth without pruning. This step differs from a traditional decision tree; 4.
In every iteration of the bootstrap, the prediction did not include the dependent variables within OOB, and the results generated from all the decision trees were to be averaged; 5.
When all OOB data was included in the generation of decision tree, the importance of every factor could be ensured through the calculation of RMSE with increased percentage; thus, the correlations of these factors with the dependent variables could be established; 6.
Last vote was to elect the classification results into tea tree or non-tea tree.

Classification Procedure-SVM Classifier
The theoretical concept of SVM was basically developed from the optimal conditions of the hyperplane from the linear or non-linear separability. The optimal hyperplane not only correctly separates two types of samples but also has the largest class intervals. For a linearly separable binary classification, if the training data with k number of samples is represented as {X i , y i }, i = 1, . . . , k where X P R N is an N-dimensional space and y P t`1,´1u is a class label with two classes, the SVM classifier can be created by a hyperplane given by: where W is a vector perpendicular to the linear hyperplane and b is a scalar showing the offset of the discriminating hyperplane from the origin. For the two classes y, two hyperplanes can be used to discriminate the data points in the respective classes. The two hyperplanes can be expressed as: pW¨X i q`b ě`1 for all y "`1, i.e., class 1, and pW¨X i q`b ď´1 for all y "´1, i.e., class 2 For a nonlinear classification, the input data is mapped onto a higher dimensional feature space using a non-linear mapping function. The classification is then carried out in the mapped space rather than the input feature space. In order to increase the computational efficiency, a kernel-trick is usually used to definite kernel function. In this study, we chose the kernel function contributing to the best tea and non-tea LULC classification from the four basic options: linear, polynomial, radial and sigmoid. Besides, two crucial parameters, cost of constraints violation (C) and sigma (σ), were also considered to create the best SVM classifier. C accounts for the over-fitting of the model while σ decides the shape of the hyperplane [65].

Accuracy Assessment
For accuracy assessment of PBIA results, we converted the ground reference data from vector to grid file format with a cell size of 0.5 m (i.e., the same cell size as the spatial resolution of pan-sharpened WorldView-2 images) and calculated the areas classified correctly and incorrectly, while for OBIA we calculated the number of segmented objects rather than the areas with correct and incorrect classification results. Indices including producer's accuracy (PA), user's accuracy (UA), overall accuracy (OA) and kappa index of agreement (KIA) were then applied to evaluate the PBIA and OBIA results. The KIA measures the difference between the actual agreement and the chance agreement in the error matrix [75,76]. The KIA is computed as: where r is the number of rows in the error matrix; x ii is the number of observations in row i and column i (on the major diagonal); x i+ is the total of observations in row i; x +i is the total of observations in column i; N is the total number of observations included in the matrix.
To assess if the statistical difference between the results of PBIA and OBIA exists, McNemar's test based on the standardized normal test statistic was used in this study. The difference in accuracy between two approaches is statistically significant if |z| > 1.96, and the sign of z indicates which the more accurate approach is [77,78]. The z-score is computed as: f GP´fPG a f GP`fPG (7) where f GP indicates the number of samples correctly classified by OBIA and incorrectly classified by PBIA and f PG indicates the number of samples correctly classified by PBIA and incorrectly classified by OBIA.

Results
In this section, we describe the feature selection results, followed by the classification ones. When discussing the results, we mostly focus on the tea cultivation area.

Feature Selection
Although both PBIA and OBIA had 80 input variables, yet their "basic-unit values" were different; thus, LR model filter shall be carried out separately as to select the best variables for pixel-based and object-based approaches.
In terms of feature selection in a pixel-based approach, the eight multispectral bands, the eight principal components and the 64 GLCM textures were used to choose the significant impact variables in the LR analysis. The analytical result obtained with Cox & Snell R square test was 0.481. The p value of the Omnibus test was 0.000, while p < 0.01 meant the explained variance of independent variables was significantly greater than the variance of the unexplained ones. This also represented that there was at least one independent variable which can effectively explain and predict the dependent variable. The p-value of Hosmer and Lemeshow test was 0.511. The insignificant p-value here meant LR model was indicated of good fit. With the 34 significant variables listed in Table 3, the best classification accuracy (87.5%) gained from the training data can be achieved.  Among them, the 1st to the 21th represents the texture variables. MEAN2 means the MEAN texture of the second band in WorldView-2 fusion image. The others may also be induced in the same way. From the results, the ENTROPY and SECOND MOMENT textures in each band were both considered as not significant, so they were both removed from the classifiers. MEAN7 has the highest B coefficient among all texture variables and this implies that it is the most beneficial index to differentiate tea from non-tea LULC. PC1 means the first main component. Except for PC6, the rest of the main components were significant; ORIGIN, which was the original band of WorldView-2 fusion image, was not significant in the first and fifth band of the model; thus, only the other six bands were kept. The variables from the aforementioned Table 3 were the input variables of the four pixel-based classifiers. In terms of feature selection step in OBIA, the Cox & Snell R square value of the analytical results was 0.547, meaning the intensity of correlation and the predictive power of the input variables and dependent variables are both good. The result of the Omnibus test was the value of p = 0.000, meaning that the explained variance was significantly greater than the unexplained variance, and among the input variables, at least one variable can effectively explain and predict the dependent variable. The p-value of Hosmer and Lemeshow test was 0.977. It meant LR model was indicated of good fit. The six significant variables listed in Table 4 can be used to achieve the best classification accuracy (94.7%) from the training data. The difference between OBIA and PBIA consisted in the variables which were related to spectral characteristics, such as MEAN, ORIGIN and PC, and they were not considered as significant; however, the similarity among them is that CONTRAST, DISSIMI and CORREL were considered as significant.
This means when the classification was carried out with OBIA, the spatial characteristics would have more referential meaning than the spectral characteristics. This feature could avoid the high spectral heterogeneity phenomenon in high spatial resolution images [72]. The variables in the aforementioned Table 4 were taken as the input variables of the four object-based classifiers.

Classification with the Selected Input Variables
The following two sub-sections demonstrated the results generated from ML, LR, RF and SVM classifiers with the use of 34 and six selected variables for the pixel-based and object-based approaches respectively.

Pixel-Based Results
The study evaluated the classification accuracies of ML, LR, RF and SVM classifiers seperately in PBIA. First, in the LR analysis, we used Equation (2) and the B coefficient in Table 3 to derive the occurrence probability p; pixels with the value of p < 0.5 were assigned as the class of tea while non-tea if p ě 0.5. RF classifier takes the best n tree = 500 and m try = 11 as the final parameter setting. SVM classifier obtained the best classification results with radial kernel function and gamma value of 0.1. The best results obtained with the use of the aforementioned classifiers were shown in Table 5 and Figure 6.  In terms of classification accuracy, SVM was the highest (OA = 84.70% and KIA = 0.690) followed by RF analysis, ML and LR. In terms of McNemar's test, the accuracy of SVM which was the best in classification was significantly higher than the second best RF (|z| = 35.961), while LR was significantly the lowest (with |z| = 72.109 compared with ML).
For more detailed information about the PA and UA values of each classifier, we found that the PA value of ML classifier was much higher than others (Figure 6a) since over 50% of study area was In terms of classification accuracy, SVM was the highest (OA = 84.70% and KIA = 0.690) followed by RF analysis, ML and LR. In terms of McNemar's test, the accuracy of SVM which was the best in classification was significantly higher than the second best RF (|z| = 35.961), while LR was significantly the lowest (with |z| = 72.109 compared with ML).
For more detailed information about the PA and UA values of each classifier, we found that the PA value of ML classifier was much higher than others (Figure 6a) since over 50% of study area was interpreted as tea tree LULC. It obviously reduced the omission error in tea tree classification (i.e., 1-PA); however, it also caused several non-tea areas to be commissioned as the classof tea and decreased the UA of the class of tea and the PA of class of non-tea.

Object-Based Results
In the LR analysis, the same as the pixel-based approach, objects with the value of p < 0.5 were assigned as the class of tea while they were non-tea. RF classifier obtained the best classification results when n tree = 500 and m try = 3. SVM classifier obtained the best classification results with radial kernel function and gamma values of 0.1. The best results gained with the aforementioned classifiers are shown in Table 6 and Figure 7. In terms of the accuracy of classification, ML was the one with the highest member in OBIA. OA and KIA may both reach to 96.04% and 0.887. LR, SVM analysis and RF followed. The one with the lowest accuracy was RF in OA and KIA with 91.20% and 0.733 respectively. The results of McNemar's test showed that ML was with the best classification accuracy which was significantly higher than the one of the second best LR (|z| = 138.082), while the accuracy of SVM analysis was significantly higher than RF (|z| = 38.051). Evaluating the details of PA and UA of the classifiers, the PA of tea tree with ML was higher than the other classifiers, which was the main reason for it to give the best overall classification results. The PA of tea tree and the UA of non-tea tree with RF were not high, meaning that many tea tree areas were classified as non-tea tree areas, which was the main reason for it to be the worse results among the classifiers.

Overall Comparison
Using the same classifier to compare the results obtained with the use of PBIA and OBIA, the classification accuracy of OBIA among ML and LR analysis classifiers was higher, while the one among the results gained with the use of PBIA in SVM and RF was higher. Although the result of McNemar's test indicated that both results had significant differences, the significant values |z| of ML and LR analysis classifiers were larger than the ones of RF and SVM. Overall, it turned out that

Overall Comparison
Using the same classifier to compare the results obtained with the use of PBIA and OBIA, the classification accuracy of OBIA among ML and LR analysis classifiers was higher, while the one among the results gained with the use of PBIA in SVM and RF was higher. Although the result of McNemar's test indicated that both results had significant differences, the significant values |z| of ML and LR analysis classifiers were larger than the ones of RF and SVM. Overall, it turned out that the results obtained with the use of OBIA were better than the results obtained with PBIA in both OA and KIA (Table 7). What is worth attention is that the accuracy of OBIA + ML was not only higher than that of PBIA + ML but also the best among all the results obtained with PBIA and OBIA. In addition, if we take a look at the ML results of the two approaches, the PA of tea tree LULC is the highest among all the classifiers mainly because the ML, whether among PBIA or OBIA results, tended to classify more areas as tea tree LULC as to avoid the error caused by omission. The performance of all classifiers of PBIA and OBIA approaches were unstable while the OA varied a lot between 82.97% and 96.04%. However, the results of RF in this study showed a difference from those of the previous studies for other LULC classification which showed the highest OA result than other classifiers [65,79]. It was speculated that perhaps in the experiment design of this study, only two categories-tea and non-tea LULC were provided-tea tree and non-tea tree LULC so that RF cannot show its advantage in classification.

Discussion
In the feature selection stage, adding variables such as GLCM layer would increase the interpretation capability of land-use/land cover, but it requires a lot of time and efforts. This study tried to find the most useful variables to be added here so we can promote the efficiency of interpretation with lowest cost. As mentioned in Figure 2, the feature selection stage based on LR model delineated 34 useful variables in PBIA and 6 GLCM variables in OBIA. All selected variables were significant related factor for LR function (p < 0.05, and B coefficient can be considered as the weights of importance). It means that we only need to input 34 variables in PBIA classifier, and six variables in the OBIA classifier. We think it is important for saving money if the government (e.g., Council of Agriculture) needs to lower the cost of crop cultivation investigation.
Previous studies have already discussed the use of four newly available spectral bands (coastal, green, yellow, red edge) for the classification of the tree species [39,80]. In this research, the results of feature selection using LR analysis showed that the influence of band 6 (red edge) and band 4 of its derivative texture indices were significant to the tea classification. Figure 8 shows the signature comparison between tea, betel nut, pineapple and forest fields using input variables with different band combination. Obviously, the visual differences became clear when standardized CONTRAST6, DISSIMI6, HOMO2, CORREL6 and PC2 were included in classification process. This means red edge spectrum and some texture indices were useful in distinguishing between tea and non-tea LULC. It also supports the conclusions of the previous studies concerning vegetation classification [39,81]. The superiority of the red edge spectrum indicates that it can reflect the chlorophyll concentration and the physiological status of crops [82]; and thus it can also be helpful for land cover classification [81]. Besides, principal component transformation is commonly considered a good method to reduce data dimension and to extract useful information. After the processing of 8-band WorldView-2 imagery, the most abundant information was transformed into the first PC (PC1) and the information decreases gradually to the last PC (PC8). According to the LR analysis results, however, PC2 to PC8 except PC6 were still helpful for tea and non-tea LULC discrimination. This implies that every component may have to be thought about when conducting binary classification. LR analysis also indicates the importance of each variable. Taking tea, betel nut, pineapple fields and forest as the examples, Figure 8 shows that the input variables with higher absolute B coefficient values differentiate the LULC above more than those with lower B coefficients.  Figure 8 shows that the input variables with higher absolute B coefficient values differentiate the LULC above more than those with lower B coefficients.  As for textural information, LR analysis in feature selection showed that GLCM MEAN and DISSIMI were the two most dominant variables for the tea LULC classification. GLCM MEAN represented the local mean value of the processing window and had high relation with the original spectral information.
The processing window also called moving window, is the basic unit for calculating grid layer features. The value of each pixel is calculated by shifting the window one pixel at a time until all pixels are calculated. The size of moving window was 3ˆ3 pixels in this study. B coefficients of the MEAN which were higher than those of other variables suggested that the spectral information was more useful than the structural characteristics of WorldView-2 imagery. Nevertheless, the dissimilarity measures, which were based on empirical estimates of the feature distribution, were also critical for enhancing the structural difference between tea and non-tea LULC.
Though previous studies seldom used MEAN and DISSIMI for tea classification, these two measures had been proven effective in improving classification accuracy for agricultural and forest areas [65,83] and the idea also indirectly support the findings of this study. The tea trees in the study area were planted in rows with totally different characteristics from those of the crops observed on the ground in structure and arrangement. However, in Table 3, the overall spectral variables of B coefficient are still larger than the structural variables when using pixel-approach. This means that in pixel-based classification, the spectral variables were more effective in distinguishing the differences of the ground coverage of tea trees and non-tea trees comparing to the structural variables. This could be explained with the two examples in Figure 9. Even in the pan-sharpened WorldView-2 imagery with 0.5-m high spatial resolution, there are not all characteristics of the line by line formation could be clearly presented due to the narrow interval between tea trees. For this reason, it seems to be much harder for the structural variables to play to their strengths.
As for textural information, LR analysis in feature selection showed that GLCM MEAN and DISSIMI were the two most dominant variables for the tea LULC classification. GLCM MEAN represented the local mean value of the processing window and had high relation with the original spectral information. The processing window also called moving window, is the basic unit for calculating grid layer features. The value of each pixel is calculated by shifting the window one pixel at a time until all pixels are calculated. The size of moving window was 3 × 3 pixels in this study. B coefficients of the MEAN which were higher than those of other variables suggested that the spectral information was more useful than the structural characteristics of WorldView-2 imagery. Nevertheless, the dissimilarity measures, which were based on empirical estimates of the feature distribution, were also critical for enhancing the structural difference between tea and non-tea LULC.
Though previous studies seldom used MEAN and DISSIMI for tea classification, these two measures had been proven effective in improving classification accuracy for agricultural and forest areas [65,83] and the idea also indirectly support the findings of this study. The tea trees in the study area were planted in rows with totally different characteristics from those of the crops observed on the ground in structure and arrangement. However, in Table 3, the overall spectral variables of B coefficient are still larger than the structural variables when using pixel-approach. This means that in pixel-based classification, the spectral variables were more effective in distinguishing the differences of the ground coverage of tea trees and non-tea trees comparing to the structural variables. This could be explained with the two examples in Figure 9. Even in the pan-sharpened WorldView-2 imagery with 0.5-m high spatial resolution, there are not all characteristics of the line by line formation could be clearly presented due to the narrow interval between tea trees. For this reason, it seems to be much harder for the structural variables to play to their strengths. As for the issues for reducing data dimensions and the number of input variables, the comparative analysis between PBIA and OBIA in this study demonstrated a significant difference that using OBIA was more effective and sensitive in conducting variable filtering process than using As for the issues for reducing data dimensions and the number of input variables, the comparative analysis between PBIA and OBIA in this study demonstrated a significant difference that using OBIA was more effective and sensitive in conducting variable filtering process than using PBIA. The strength of OBIA was attributed to the less amount of samples and the reduction of spectral heterogeneity within each variable layer when producing objects from pixels. Taking this study for example, it means that people could spend less time generating input variables and achieve satisfactory results compared to using PBIA if OBIA is applied. However, it is worth noting that OBIA demands more computational resources, requires specialized software (i.e., eCognition Suite) and is hard to applied in a large study area even the number of objects is much less than the total number of pixels. The above mentioned factors often set limitations on the use of OBIA.
One parametric classifier (ML) and three non-parametric classifiers (LR, SVM and RF) were applied in this study. According to the accuracy assessment of PBIA and OBIA approaches, almost all OA values were found to exceed 82.97%. The use of OBIA + ML obviously got the best OA value among OBIA approaches, but the use of PBIA+ML did not. The main reason that the overall accuracy of ML classifier was not as good as non-parametric approach (SVM or RF) in PBIA comes from the bimodal data distribution of tea-farm training samples due to the alternate permutation of tea plants and corridors. The special arrangement might lead tea farms divided into two LULC types in PBIA when ML classier was applied. However, in OBIA, the basic unit was polygon that contained many WorldView-2 image pixels, the average treatment effect within tea farm training samples would decrease the difference of every basic unit, eliminate the spectral heterogeneity within every tea LULC patches, and lead sample distribution more like normal distribution. The ML classifier performs well for normally distributed data, this is why the classification result of the use of OBIA + ML was much better than PBIA + ML.
Minimum distance classifier has been commonly used to classify remote sensing imagery [84,85]. This classifier works well when the distance between means is large compared to the spread (or randomness) of each class with respect to its mean; moveover, it is very good approach dealing with non-normal distribution data. However, this study did not consider this classifier as one of the methods in PBIA and OBIA approaches. The reason is that the non-tea class spreads much wider in the feature space and thus the distance between the means of non-tea and tea classes is not large compared to the spread of non-tea class. In this case, if there is an unclassed sample situating near the boundary of non-tea and tea classes in the feature space, the classification result may not be ideal. As for the ML classifier, the occurrence frequencies of non-tea and tea classes are used to aid the classification process. Without this frequency information, the minimum distance classifier can yield biased classifications. Although salt and pepper effect may influence the distribution normality of the data, OBIA approach can decrease the possibility of salt and pepper effect. Therefore, we think ML classifier still has the superiority especially in OBIA approach.
As for SVM classifier, its superior classifying effects toward high-dimensional feature space, capability on processing non-linear data, characteristics to avoid over fitting and under fitting, and the stability to decrease the salt and pepper effect made it the best classifier of PBIA. However, it was not the best algorithm of OBIA. We inferred the most possible reasons: (1) the OBIA image segmentation stage in this study made the sample data look like a case of normal distribution, decreased the possibility of salt and pepper effect, and promoted the effects of ML algorithm; (2) OBIA image segmentation was possible to combine tea farms and pineapple farms together due to the line by line arrangements of tea trees and pineapples with unfixed interval. Although the advantages of SVM with soft margin were not presented in the application of OBIA, the OA and Kappa values were still good.
As for the LULC monitoring frequency, compared with many herbaceous crops, the monitoring task on tea trees with high spatial resolution commercial satellite imagery is much easier because of the long-lived perennial characteristics of tea tree. The relatively stable life cycle of tea production in Taiwan often starts on the 3rd year after the planting of saplings, and then there is going to be one to two harvests every season for at least 30 years. Therefore, people do not need to collect a series of period images as frequently as in the monitoring on herbaceous crops. Nevertheless, it is worth mentioning that this study still has some limitations. The first one is the study area of this research was limited to a flat and gentle sloping area, yet some tea farms are in the high mountains where the terrain effect would certainly have an influence due to the shadow-effect of satellite images. For this reason, we suggest the users to apply our workflow in flat or gentle slope areas. Besides, at some places, physiological changes among non-tea plants might influence the spectral reflection of satellite images, so we considered that our workflow is suitable for the use in sub-tropical places where most plants are evergreen.
After all, in order to achieve tea production prediction or evaluation of high efficiency and low costs, we suggest the governments apply the methodology used in this study as a standardized framework, while field surveys and re-investigation with manpower resources should be the backup option for calibration or supplementary once people want to totally eliminate the commission and omission errors. The classification method in this study can be further applied in two different ways. The first one is of benefit to countries with the largest tea planting areas (e.g., China, India, Kenya, Sri Lanka, and Vietnam). Besides, it can also be applied in other crop classification studies such as banana, pineapple, sugar cane, and watermelon with particular type of plantation and farming arrangement.

Conclusions
Traditional field surveys or orthogonal photograph interpretation for the monitoring of sensitive crops are time-consuming and expensive. The assistance of high spatial and spectral resolution imagery has been proven useful for mapping several crops and helpful in agriculture land management and sustainable agricultural production. However, there is a gap in the research on mapping tea LULC in intensive and heterogeneous agriculture areas due to the spectral and spatial resolution restrictions. This study investigated the WorldView-2 imagery characteristics and the most applied significant variables and classification approaches used to develop a stepwise classification framework for tea tree mapping in Taiwan. The results show it is not enough for tea and non-tea crop classification if we only apply the original approach of eight MS bands of WorldView-2. Advanced variables such as GLCM texture indices are necessary. It is evident that logistic regression analysis in this study was helpful in feature selection, and the OBIA approach filtered more insignificant variables than the PBIA approach. In other words, the OBIA approach costs less but generated additional variables.
As for higher accuracy of tea tree classification, one parametric classifier (ML) and three non-parametric classifiers (LR, SVM and RF) were applied. The obtained data showed that all classifiers achieved overall accuracies ranging from 82.97% to 96.04%. Parametric ML showed significantly higher accuracy in OBIA while the lower accuracy of PBIA was due to the different spectral heterogeneity in the tea tree samples. We suggest that the OBIA+ML approach would offer the best combination of efficiency and stabilization in tea tree classification. Overall, the applicability of remote sensing imagery for the evaluation, and monitoring of tea trees was well confirmed.