Get Our e-AlertsSubmit Manuscript
Journal of Remote Sensing / 2022 / Article

Research Article | Open Access

Volume 2022 |Article ID 9847835 | https://doi.org/10.34133/2022/9847835

Xiandie Jiang, Shuai Zhao, Yaoliang Chen, Dengsheng Lu, "Exploring Tree Species Classification in Subtropical Regions with a Modified Hierarchy-Based Classifier Using High Spatial Resolution Multisensor Data", Journal of Remote Sensing, vol. 2022, Article ID 9847835, 16 pages, 2022. https://doi.org/10.34133/2022/9847835

Exploring Tree Species Classification in Subtropical Regions with a Modified Hierarchy-Based Classifier Using High Spatial Resolution Multisensor Data

Received26 Feb 2022
Accepted06 Jun 2022
Published23 Jun 2022

Abstract

Tree species distribution is valuable for forest resource management. However, it is a challenge to classify tree species in subtropical regions due to complex landscapes and limitations of remote sensing data. The objective of this study was to propose a modified hierarchy-based classifier (MHBC) by optimizing the classification tree structures and variable selection method. Major steps to create an MHBC include automatic determination of classification tree structures based on the -score algorithm, selection and optimization of variables for each node, and classification using the optimized model. Experiments based on the fusion of Gaofen-1/Ziyuan-3 panchromatic (GF-1/ZY-3 PAN) and Sentinel-2 multispectral (MS) data indicated that (1) the MHBC provided overall classification accuracies of 85.19% for Gaofeng Forest Farm in China’s southern subtropical region and 94.4% for Huashi Township in China’s northern subtropical region, which had higher accuracies than random forest (RF) and classification and regression tree (CART); (2) critical variables for each class can be identified using the MHBC, and optimal variables of most nodes are spectral bands and vegetation indices; (3) compared to results from RF and CART, MHBC mainly improved the accuracies of the lower levels of classification tree structures (difficult classes to separate). The novelty in using MHBC is its simple and practical operation, easy-to-understand, and visualized variables that were selected in each node of the automatically constructed hierarchical trees. The robust performance of MHBC implies the potential to apply this approach to other sites for accurate classification of forest types.

1. Introduction

Subtropical ecosystems in China play important roles in regional and global carbon regulations and climate changes because of their rich tree species, dense forest coverage, and high carbon sequestration [1]. Tree species maps are regarded as a basic product for quantitative measurement and evaluation of forest ecosystems [2]. Accurate mapping of subtropical forest distributions in a timely manner is also needed for making proper decisions for the development of forest management strategies [3]. Remote sensing technology has become an effective tool for analyzing, monitoring, and evaluating the changes of forest resources [4]. With the transformation of forest resource management from extensive to intensive, forest classification has gradually shifted to accurate identification of detailed tree species. The subtropical region has complex terrain conditions and dense forest coverages but with fragmental and diverse patch sizes due to frequent disturbances from humans and nature [5]; thus, classification of tree species classes is a challenge using remote sensing technologies [6].

Sentinel-2, with its improved spectral and spatial resolutions compared to Landsat, is often used for land cover or forest classification [7, 8]. Previous studies show that near-infrared (NIR), red-edge, and shortwave infrared (SWIR) bands are valuable for tree species classification [7, 9]. However, considering the fragmented forests and diverse forest patch sizes in subtropical regions, the mixed-pixel problem in Sentinel-2 imagery with 10/20 m spatial resolution may be an important factor leading to poor classification [6]. Although high spatial resolution data, such as Worldview, Gaofeng-1/2/6, and Ziyuan-3, are rich in spatial information, the limited number of spectral bands (blue, green, red, and NIR) make it difficult to distinguish different tree species [6, 10]. Thus, integration of multisensor data may be an effective way to improve the classification accuracy [6, 11]. However, increased spatial resolution can cause spectral heterogeneity within the same land cover, resulting in a large number of “salt-and-pepper” pixels when the pixel-based classifier is implemented [10]. In this situation, the object-based classifier has been proven to provide a better accuracy than pixel-based approaches [6, 10]. As different ancillary data became available, the combination of remotely sensed and ancillary data is an alternative to improve the classification accuracy [10, 12, 13].

Variable selection is a critical step to improve classification accuracy [11, 14]. As summarized by Lu and Weng [15], some potential methods such as graphic analysis, statistical methods, and the fuzzy-logic expert system have been used to identify optimal combination of variables. An alternative is to use random forest (RF) method because of the ability to provide importance ranking of variables [3, 16, 17]. Too many variables used in a classification procedure cannot guarantee the best classification result, but selection of the optimal variable combination from a wide range of variables is critical for separation of specific classes [6]. Many previous studies have proven that, when multisource data are used, machine learning algorithms (e.g., artificial neural networks (ANN), support vector machine (SVM), classification and regression trees (CART), and RF) have advantages over traditional classification methods/techniques (e.g., maximum likelihood classifier (MLC) and minimum distance) in dealing with complex data, thus, providing better classification [10, 18, 19]. In recent years, deep learning due to its powerful data mining ability has been employed for tree species mapping based on high spatial resolution and/or hyperspectral images [2022]. Because different tree species have similar spectral signatures, distinguishing tree species types using spectral bands alone is often difficult. However, different tree species have their own characteristics in crown size and tree height, and incorporation of Lidar-derived variables that reflect canopy or tree features into spectral features is proven effective in improving tree species classification [22].

Of the machine learning algorithms, CART and RF, due to their efficiency in computation and good classification performance, have been extensively used for land cover or forest classification [11, 23]. The tree structure of CART is divided recursively to ending points or terminal nodes according to preset criteria. CART determines which binary division of a single variable best reduces the bias of response variables by analyzing all input variables [24, 25]. The result of CART analysis is a dichotomous decision or classification tree in which each internal node represents a test on a variable, each branch represents a test output, and each leaf node represents a class. Therefore, the tree can be viewed as a series of rules to predict likely class membership [25]. However, the processing time of CART generally increases exponentially with the height of the tree, while the larger width of the tree with shallow depth provides poor accuracy. CART with automatic selection variables has been widely used in land cover or forest classification [26, 27].

RF is an ensemble classifier using a large number of decision trees (DTs) to overcome the weaknesses of a single DT and using voting strategies to obtain the final classification result [28, 29]. The randomness is reflected in the random selection of training samples and feature sets [24]. RF requires much less processing time for parametric optimization than other machine learning algorithms such as SVM [10, 30], but the classification results can maintain a stable accuracy [11, 31]. The advantages of RF in processing high-dimensional data, having a certain fault tolerance, and being fast in the training stage make it popular in regional and global land cover or forest classification [30, 32, 33]. The disadvantage is that it is impossible to visualize the trees, which is equivalent to completing the classification in a black box and directly outputting the results [11, 34].

Compared with CART and RF, the hierarchy-based classifier can effectively determine the optimal variables for each tree, thus, may produce more accurate classification [11, 35]. However, how to select an optimal combination of variables from the high-dimensional features and how to ensure the classification accuracy while reducing the complexity of the model are particularly important. Chen et al. [11] proposed a hierarchical procedure with optimized node variables and thresholds to classify multiple tree species in North China, with an overall accuracy of over 86%. However, this method is strongly influenced by the user’s experience in designing the tree structure and the complexity of tree species classes. Another problem in this method is adopting the single classifier decision-making mode, similar to the shortcoming in CART. Zhao et al. [35] determined the optimal variables by iterating RF at each parent node based on testing accuracy value to conduct urban vegetation classification. Based on previous exploration [11, 35], our research further modifies the method to optimize the tree structure by introducing the -score test. Specifically, the contribution of this research is to propose a new method to optimize the classification tree structures and variable selection method and to explore the robustness of the modified hierarchy-based classifier (MHBC) in different landscapes.

2. Materials and Methods

2.1. Study Areas

Gaofeng Forest Farm, Nanning, Guangxi Zhuang Autonomous Region was selected as an experimental site (Figure 1). This farm covers approximately 220 km2 on Daming Mountain in the southern subtropical region with an elevation range from 70 to 900 m above sea level. The climate is subtropical monsoon with average annual temperature of about 21°C and average annual precipitation of about 1300 mm. The dominant forest types are plantations of Masson pine, Chinese fir, eucalyptus, Chinese anise, and other broadleaf evergreens [36].

Huashi Township in Jinzhai County, Anhui Province was selected as a validation site for exploring the transferability of our proposed approach (Figure 1). This area has a northern subtropical humid monsoon climate with elevations from 186 to 1302 m, an annual precipitation of 1381.5 mm, and average annual temperature of 15.5°C. The forest types mainly include Masson pine, Chinese fir, Quercus, and bamboo.

2.2. Methods

The framework of this research is illustrated in Figure 2, including (1) collection of multiple data sources and extraction of digital surface model (DSM) data from ZY-3 stereo images; (2) data fusion of Gaofen-1/Ziyuan-3 panchromatic (GF-1/ZY-3 PAN) and Sentinel-2 multispectral (MS) data and image segmentation using eCognition; (3) extraction of potential variables from the fused data; (4) implementation of image classification using MHBC, CART, and RF; and (5) comparison and evaluation of classification results.

2.2.1. Data Collection and Organization

Different optical sensor data (GaoFen-1, ZiYuan-3, and Sentinel-2), ancillary data (forest distribution maps), and field survey data were collected (Table 1). All data were registered to the UTM Zone 49 N/WGS-84 projection system. (1)Field survey data collection and organization


Study areaDatasetAcquisitionDescription

Gaofeng Forest FarmGF-1 PAN2019/12/10Panchromatic band with 2 m spatial resolution.
Sentinel-2 MS2019/12/07Four bands (three visible bands and one near infrared (NIR) band) at 10 m spatial resolution, six bands (three red-edge bands, one narrow NIR band, and two shortwave infrared (SWIR) bands) at 20 m spatial resolution.
ZY-3 stereo2018/03/10
2020/01/30
Nadir-view image with 2.1 m, backward and forward view images with 3.5 m spatial resolution.
Forest distribution map2016The forest distribution map with digital vector format was provided by Gaofeng Forest Farm. The subcompartments in this map contain detailed forest attributes such as tree species, age, average stand height, and stem volume.
Field survey2017/12
2019/09
A total of 2166 samples were collected [6]. The samples collected in 2017 were checked to make sure they are not changed in 2019 so that all samples can be used as one population

Huashi townshipZY-3 PAN2019/12/13
2020/04/09
Panchromatic band with 2.1 m spatial resolution.
Sentinel-2 MS2019/12/11
2020/04/09
Same as above for Gaofeng Forest Farm
ZY-3 stereo2019/12/13Same as above for Gaofeng Forest Farm
Forest distribution map2019Forestland “one map” database of Anhui Province, showing different forest types
Field survey2020/05Land-cover types and associated locations were recorded and then organized and refined. A total of 526 samples were obtained.

The field surveys were conducted at Gaofeng Forest Farm in December 2017 and September 2019 and at Huashi Township in May 2020. The samples were first randomly separated into two groups: 60% of the samples were used for modeling and 40% for validation. Then, the samples of modeling were separated into training samples and test samples according to the same ratios. The numbers of training, test, and validation samples are summarized in Table 2. Considering the tree species types and objectives of this research, the classification systems were designed with 10 land-cover classes in Gaofeng Forest Farm and 6 in Huashi Township (see Table 2). Because of different area amounts of forest types in these study areas, the numbers of collected samples for training, test, and validation varied considerably. (2)Remotely sensed data collection and preprocessing


Study areaLand-cover typeTraining samplesTest samplesValidation samples

Gaofeng Forest FarmMasson pine674676
Chinese fir523558
Eucalyptus12484140
Chinese anise312135
Schima231627
Other broadleaf trees422847
Bamboo forest282033
Shrub231627
New plantation271831
Other land covers644473
Total481328547

Huashi townshipMasson pine372542
Chinese fir221526
Quercus342338
On-year bamboo221525
Off-year bamboo271831
Other land covers422951
Total184125213

Note: other land covers include impervious surface area, bare soils, and water.

Considering that the spatial resolutions in openly accessed DEM data such as SRTM and ASTER GDEM are too coarse to meet the requirement of this research, we used ZY-3 stereo images to generate DSM data with 2 m spatial resolution using Geomatica PCI software. The major steps include conducting relative orientation with a surface 3D model, conducting absolute orientation through manually selecting ground control points, creating tie points connecting two images, and finally developing an epi-polar image and extracting corresponding DSM. Xie et al. [10] provided the detailed procedure of DSM extraction. Since DSM represents the land surface elevation, it is necessary to conduct postprocessing to create a substitute of the DEM. Our previous research in this region had explored a minimum filtering algorithm with a window size of 5 by 5 pixels, followed by a median-filtering algorithm with the same window size to generate the DEM data [6]. Two ZY-3 stereo images were mosaiced at the Gaofeng Forest Farm, and the panchromatic imagery was used as reference. Both Sentinel-2 and DEM images were registered to the UTM Zone 49 N coordinate system with root mean square error of less than 0.5 pixels.

The panchromatic bands from Gaofen-1 (GF-1) and from Ziyuan-3 (ZY-3) with 2 m spatial resolution were used for two sites, respectively. The panchromatic bands were orthorectified, then atmospherically calibrated using the dark-object subtraction method, and resampled to a pixel size of 2 m. The Sentinel-2 data (Level-1C product) with 10 spectral bands were used. Sen2cor was used to process the image into L2A (atmosphere-corrected reflectance) products, and Sen2Res was used to enhance the spatial resolution of the products with 10 m/pixel [37]. The approach was then used to implement topographic correction for both GF-1/ZY-3 PAN and Sentinel-2 data [38], as this algorithm has a good correction effect for complex terrain areas. This process was completed in the topographic correction module of ENVI software. DEM raster, sun azimuth, and sun elevation are used as input. In addition, two parameters are needed to be set: kernel size (the larger the kernel size value, the smoother the terrain factor calculated) and grid size (sampling interval of sample points involved in the regression operation for the empirical parameter C). Default values were used for both parameters in this study.

2.2.2. Image Segmentation Based on Fused Imagery

Our previous research had explored a comparative analysis of multisensor/multiresolution data fusion scenarios using a High Pass Filter (HPF) fusion algorithm and found that the ZY-3 PAN and Sentinel-2 MS fused data provide the highest accuracy [6]. Thus, fusion results based on Sentienl-2 MS and GF-1/ZY-3 PAN were used in this research. Considering the effectiveness of using the object-based classification approach with high spatial resolution images and knowing the importance of obtaining high-quality image segmentation [39], we proposed a new segmentation method to improve the accuracy of image segmentation. The major steps include (1)Chessboard Segmentation Based on Vector Data. The forest distribution map as a thematic layer was added to the eCognition software for image segmentation using the chessboard segmentation algorithm. The object size was often set to be larger than the number of image rows and columns, and the same segmentation object as the vector data was output(2)Multiresolution Segmentation Based on Upper-Level Segmentation. Multiresolution segmentation was set to perform at the image object level, not pixel level. The segmentation object of chessboard segmentation was chosen to be the image object level. “Creat below” used as level usage means that multiresolution segmentation completes within each object. Four key parameters—scale of segment, weight of input layers, weight of spectra and shape, and weight of compacts and smoothness—need to be defined in this segmentation algorithm. In our research, the fused image was used as the basis for segmentation. The weight of visible bands was set as 1, and the rest such as NIR, red-edge, and SWIRs were set as 2 to make use of the differences of tree species in these bands. Shape weight and compact weight were set as 0.3 and 0.5, respectively, after a substantial number of adjustments. Estimation of scale parameters was used to identify several candidate segmentation scales, and 200 were finally selected by checking the results of these segment scales

A comparative analysis was performed using conventional multiresolution segmentation to verify the usability of the proposed segmentation algorithm. The comparison of image segmentation results indicated that the forest distribution map can better outline the boundaries of different forest types and makes the overall segmentation avoid fragmentation (② in Figure 3). As shown in ① of Figure 3, the plot is obviously a newly felled area; however, the forest distribution map is wrong due to the delay of this data. The segmentation method in this study is to perform multiscale segmentation again directly on the basis of forest distribution map, to redraw the boundary in this changing object. In addition, ③ in Figure 3 are difficult to segment using spectral bands alone and basically show fragmental patches in multiscale segmentation (Figure 3(a)). In this study, this method provides a good basis for the following classification, especially for those plantations with obvious planting shape characteristics.

2.2.3. Extraction of Variables

For optical sensor data, common variables can be pixel-based (spectral bands and vegetation indices) and spatial-based (textures and segments). Spectral bands are the most common variables used for vegetation classification, while vegetation indices may highlight some special information that individual spectral bands do not have. Based on our previous research [10, 11, 40], six vegetation indices (Table 3) were selected.


VariablesEquations or descriptionsReferences

(1) Spectral bands (10)Ten spectral bands of the fused imagery:
(a) Blue, green, red, near infrared (NIR), and three red-edge (RE1, RE2, and RE3) bands
(b) One narrow NIR band (NNIR)
(c) Two shortwave infrared bands (SWIR1 and SWIR2)
(2) Vegetation indices (6)
Normalized difference vegetation index
[42]
Normalized difference water index[43]
Normalized difference infrared index[44]
MERIS terrestrial chlorophyll index[45]
Red-edge normalized difference vegetation index 1[46]
Red edge normalized difference vegetation index 2[47]
(3) Textural variables (80)Mean, standard deviation, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation[48]
(4) Segment polygon features (6)Shape, area, length, width, rectangle, and ellipse
(5) Topographic factors (3)Elevation, slope, and aspect

Textures and segments are often used spatial features in forest classification. A textural image is a complex combination of texture measure, spectral band, window size, and direction [41]. In order to reduce the selection of textural images based on different window sizes, this research employed grey-level cooccurrence matrices (GLCM) to calculate textural images based on segments which were developed using the above proposed segmentation algorithm. Considering different patch sizes of plantations, segmented polygon features (e.g., shape, area, and length) were also used as extra variables for forest classification. In addition to remote sensing-derived variables, elevation, slope, and aspect were included because of the relationship between tree species distribution and terrain conditions.

2.2.4. Design of the Hierarchy-Based Classifier

(1)Automatic determination of classification tree structures

In this study, value was introduced to quantitatively evaluate the performance of each variable in the separability of different forest classes. The variables of training samples were normalized to the dimensionless using Eq. (1), and the value of each variable in any two classes was calculated using Eq. (2). where and represent the original and normalized variables of the training samples, respectively; and are the maximum and minimum values of variables; , , and are the mean, standard deviation, and number of the samples; and represent different forest types. A larger value implies a higher discrimination ability of this feature. The difference between average values of two forest types in the numerator directly determines the value. The variance and number of the samples in the denominator are tuned to take the sample quality into account.

The automatic stratification strategy includes three parts: feature selection, cluster center search, and group determination. (2)Selection and optimization of variables for each node

Step 1. According to Eq. (2), all values of any two classes based on different variables are calculated to form the dataset . is the number of variables, and is the number of classes.

Step 2. The maximum value () in was used to determine the most representative variable and the two most different classes and , namely, the two cluster centers for the top layer (the easiest classes to separate).

Step 3. The classes were grouped by comparing the similarity of other classes with these two cluster centers, followed by lower layers (relatively easy classes to separate) until all classes are arranged in the classification trees.

The variable selection and node accuracy setting were conducted for each node. The RF algorithm was used to identify importance rankings of all variables. Pearson correlation analysis was conducted to examine the relationships between all importance-ranking-sorted variables. A variable having high correlation with other variables while having less importance in the ranking list was removed, and the potential key variables for each node were determined. In addition, users can set an acceptable accuracy for each forest type in each node.

In order to best distinguish between classes, identification of one best variable or more is needed. Here, an iterative approach was used to select the optimal variables in each loop by comprehensively taking the test and setting accuracies into account. If the test accuracy is higher than the setting accuracy, the variables are selected, and the classifier moves to the next parent node. Otherwise, iteration of adding new variables continues, and the classification of the next node is not started until the test accuracy meets the user demand or all potential variables are added to the model. If the latter situation occurs, the maximum of the test accuracy of each loop is recorded, and its corresponding variables are selected in the final RF model. (3)Classification using the optimized MHBC model

The optimized MHBC model is composed of three parts: automatic determination of classification tree structures, selection and optimization of variables for each node, and node binary classification. First, the classification tree structure was established through the training samples of all forest types. Then, variable optimization was carried out for each node. It is worth noting that each node uses the samples of all nodes below the node. After optimal variables for all nodes were selected, RF was integrated into the MHBC model as a classifier for each node. Three parameters of RF are required to optimize: ntree (the number of decision trees), mtry (the number of variables randomly sampled as predictors for each split (default value is one-third of the total variables)), and node size (default is one). Node size and mtry are often kept as default values, and ntree as 500 was finally selected [31]. Based on training samples and optimal variables as input, the segmented objects were classified and marked with the trained RF at each node. When entering the next node, the classification mark was updated based on the classification results of this layer until all nodes were completed. The MHBC relied on the integration of sklearn (open source library) and QGIS 2.18 (open source platform). The complete source code is shared on the GitHub platform (https://github.com/YGwork123/Auto-Hierarchy).

2.2.5. Comparison of MHBC with Other Classifiers

For the purpose of evaluating the classification performance using the MHBC, RF and CART were selected because they have similar classification strategies. CART was implemented using “rpart” in R package. Feature selection and pruning are the core items in the CART algorithm. “gini” was selected as the parameter of split method. Since a single CART was used here, the pruning step was not needed. “randomForest” in R package was used for RF classification. The training samples in Table 2 and all selected variables were used in this RF algorithm. Three parameters used by RF are the same as those set by MHBC method: ntree was set as 500, and mtry and node size were kept as default values. The classification results were evaluated using the confusion matrix approach [49].

A total of 547 verification samples were selected randomly (see Table 2). From the confusion matrix, overall classification accuracy and kappa coefficient were used to assess the overall classification results; producer’s and user’s accuracies were used to evaluate each land-cover classification result. Meanwhile, the nonparametric McNemar test [50] was used to evaluate whether a statistically significant difference in the overall accuracies among the selected models was present. The test was implemented in SPSS software. The classification results of all validation samples in different algorithms were grouped as right or wrong. The values for the McNemar test were approximated by the chi square distribution ().

2.2.6. Transferability of MHBC

The experiments described above were conducted in Gaofeng Forest Farm. To explore the robustness of the MHBC, Huashi Township was selected as a validation site. Based on the field survey data, a classification system including Masson pine, Chinese fir, Quercus, Moso bamboo (on-year and off-year), and other land covers was designed. Considering the separation of deciduous and evergreen forests, two scenes of ZY-3 images acquired 13 December 2019 (leaf-off season) and 9 April 2020 (leaf-on season) were used in Huashi Township. The ZY-3 PAN data from different seasons and Sentienl-2 MS from the same periods were fused. The Forestland “One map” database was used to assist in image segmentation, and three parameters (weights of (a) input layers, (b) spectra and shape, and (c) compacts and smoothness) were consistent with the Gaofeng Forest Farm, except that the segmentation scale of 250, instead of 200, was selected, based on visual interpretation. The same procedure for variable extraction, classification, and accuracy assessment was used at this validation site.

3. Results

3.1. Analysis of Classification Tree Structures and Optimal Variables at Each Node

The values of any two classes (Figure 4) indicated importance of variables as follows: . Spectral variables have obvious advantages in distinguishing forest from nonforest (e.g., water, impervious surfaces, and bare soils) and between coniferous and broadleaf forests. However, it is difficult to distinguish detailed coniferous tree species or broadleaf tree species using spectral variables alone. For example, the value between Masson pine and Chinese fir is only 5.9, and the value between different broadleaf tree species is approximately 10. Although the overall values of textural variables and segmented polygon features are not high for broadleaf tree classes, they are higher than spectral variables. Figure 4 also shows that topographical-based variables alone cannot distinguish tree species, but the values for nonforest, bamboo forest, and Schima (generally located on ridge lines) are higher than some other forest types. The values in Figure 4 indicate that different kinds of variables have various roles in distinguishing tree species classes. Thus, identification of an optimal combination of variables has the potential to successfully separate different tree species.

The determined classification tree structure and corresponding value (Table 4 and Figure 5) show that the optimal variables in each node are different. For instance, spectral bands and vegetation indices are often selected in the nodes and have higher values than other variables, indicating the importance of using spectral variables in differentiating tree species. The values of the nodes at the lower levels (e.g., P8–P11) are below 15, indicating the difficulty of tree species classification. Table 4 also shows that some nodes, for instance, nodes P6 and P9, include topographic and shape features, indicating some tree species classes cannot be separated using remotely sensed features alone, and incorporation of ancillary data is needed.


NodeKey feature valueSelected variables

P1NIR89.1NIR, SWIR1, and HOMRed
P2NDII77.2NDRE2, NDII, CORRed, and DISRed
P3NIR38.7NIR, SWIR1, MTCI, RE1, HOMRE1, and shape
P4NDII37.6SWIR1, NDII, RE1, HOMNNIR, HOMSWIR2, NIR, and elevation
P5NDII26.9Blue and DISSWIR1
P6Shape22.8Slope, shape, NDII, NNIR, and elevation
P7RE215.4MTCI, RE2, RE1, and NDII
P8DISRE312.0HOMBlue, HOMGreen, DISRE3, DISRE2, HOMRE1, DIS SWIR2, NDRE1, STDSWIR2, HOMNIR, HOMRE2, and MTCI
P9Shape10.6Slope, shape, elevation, CORSWIR2, NDII, NNIR, CORBlue, and aspect
P10NDII8.5NDII, SECRed, and ENTNIR
P11NIR5.9NDWI, blue, elevation, NNIR, MTCI, HOMRE1, SWIR1, HOMSWIR1, CORSWIR2, and MENIR

Notes: HOM: homogeneity; COR: correlation; ENT: entropy; DIS: dissimilarity; ME: mean; STD: standard deviation; SEC: second moment; RE1, RE2, and RE3 represent three red-edge bands. See Table 3 for additional terminologies.
3.2. Comparative Analysis of Classification Results

Comparison of classification results among three classifiers indicates that MHBC has the highest overall accuracy of 85.2%, and CART has the lowest accuracy of 75.7% (Table 5). MHBC improved overall classification by 9.5% and 4.8% compared to CART and RF, respectively, and showed statistically significant improvement in McNemar’s test. The comparison of confusion matrices among three classification results (Table 6) indicates that MHBC yields greatly improved results for some tree species classes. For example, with MHBC, eucalyptus has significantly higher user’s accuracy (90.0%) and producer’s accuracy (96.4%) than with RF (87.3% and 88.6%) and CART (89.2% and 88.6%). Similar situations occur for Masson pine and Chinese fir. However, other tree species, such as Chinese anise, had a relatively lower producer’s accuracy using MHBC than using RF and CART. Schima, shrub, and new plantation had a relatively higher user’s accuracy (100%, 95.2%, and 96.9%) using RF than using MHBC and CART. We found that the number of validation samples of these four forest types is relatively small, less than 35. The misclassification of several samples leads to a large difference in the final results.


ClassifierAccuracy assessmentPairsMcNemar’s test
Overall accuracyKappa value

MHBC85.190.83MHBC-RF19.530.000
RF80.440.77MHBC-CART44.850.000
CART75.690.72RF-CART22.320.000

Notes: MHBC: modified hierarchy-based classifier; RF: random forest; CART: classification and regression tree; value < 0.05 indicates a significant difference between two classifiers.

TypesMPCFEUCASCOBTBBFSHNPOLCUAPA

MHBCMP60170612000069.878.9
CF14390100000072.267.2
EU011351211000090.096.4
CA2102300000088.565.7
SC0000211000095.577.8
OBT0054333010071.770.2
BBF0000003130091.293.9
SH0000002210091.377.8
NP0000000231191.2100.0
OLC00000000072100.098.6

RFMP52220612000062.768.4
CF22360100000061.062.1
EU001241314000087.388.6
CA1002511000089.371.4
SC00001800000100.066.7
OBT10162430010055.663.8
BBF0000003250086.597.0
SH0000001200095.274.1
NP0000000031196.9100.0
OLC0000000107298.698.6

CARTMP49270523000057.064.5
CF23300010000055.651.7
EU001241211001089.288.6
CA0102614000081.374.3
SC2011183020066.766.7
OBT20152326000054.255.3
BBF0000002430185.772.7
SH0000007200074.174.1
NP0000002128382.490.3
OLC0000000126995.894.5

Notes: MP: Masson pine; CF: Chinese fir; EU: eucalyptus; CA: Chinese anise; SC: Schima; OBT: other broadleaf trees; BBF: bamboo forest; SH: shrub; NP: new plantation; OLC: other land covers; UA: user’s accuracy; PA: producer’s accuracy.

It is interesting to note that the lower levels of the classification tree structures, including P8 (eucalyptus and other broadleaf trees), P10 (bamboo forest and shrub), and P11 (Masson pine and Chinese fir) (see Figure 5), show a significant accuracy improvement with MHBC. Nevertheless, the accuracy of Masson pine and Chinese fir was still lower than 80%. Overall, large proportions of eucalyptus and coniferous forests occupy the study area, Chinese anise is mainly distributed in the eastern part, and bamboo forests and shrub occupy a small proportion widely dispersed around the study area (Figure 6). New plantations and other land covers (mainly bare soils) occupy a large area in the study area, indicating active forest management activities that include deforestation in Gaofeng Forest Farm.

3.3. Transferability of the MHBC

The classification tree structure of Huashi Township in Figure 7 shows that the first layer of the tree consists of vegetation and nonvegetation, but Quercus is classified as nonvegetation. The key feature of this layer is the NDII extracted from the winter imagery, because Quercus is a deciduous tree and has a spectral signature similar to bare soil during winter. Except for nodes 1 and 7, the key variables of other nodes were mainly from the variables in April images, such as NDIIA at node 5 and SWIRA at node 6, indicating that the tree species were more easily distinguished in spring. The selected variables for the nodes at the bottom of the classification tree structure were mainly SWIR or RE bands in the leaf-on and leaf-off seasons (Table 7), where the vegetation index was sufficient to distinguish on-year and off-year bamboo forests at node 5, while two homogeneous texture variables (Table 4) were also used for Masson pine and Chinese fir at node 11 in Figure 5.


NodeKey feature valueSelection of variables

P1NDIID34.3NDIID, SWIR2D, NDWIA, RE3A, and MTCIA
P2SWIR1A21.9Green A, RE3A, and SWIR1D
P3RE3A20.1NIRA, MEA-RE1, NDWID, NIRD, SWIR1D, and MEA-Green
P4SWIR2A19.8RE1A and NDVIA
P5NDIIA11.9NDVIA, SWIR1A, NDIID, MTCID, and NDRE2D
P6SWIR1A8.3SWIR1A, NDIIA, RE3D, HOMA-RE3, HOMA-SWIR1, and NNIRD
P7BlueD6.2BlueD and NDIID

Notes: A: the fused image in April; D: the fused image in December; HOM: homogeneity; ME: mean; RE1, RE2, and RE3 represent three red-edge bands. See Table 3 for additional terminologies.

Application of MHBC to the Huashi Township test site yielded the overall accuracy of 94.37% and Kappa of 0.93 (Table 8), indicating the robustness of this MHBC classification procedure. The McNemar’s test results also prove that the MHBC has significantly better performance than RF and CART. However, the value at this test site is significantly lower than that at Gaofeng Forest Farm because the tree species classes are much simpler and fewer in number. Comparing the classification accuracy assessment results in Table 6 with contents in Table 8 shows the considerably different overall classification accuracy. The main reason is that use of phenological features in the northern subtropical region (Huashi Township) played an important role in distinguishing tree species classes, especially the deciduous tree species. As shown in Figure 8, different forest types have their own spatial patterns, for example, Quercus occupied a large proportion of this test site, in particular, in west and northwest part; Masson pine, and Chinese fir distributed in the central part and bamboo dispersed in different locations of this test site.


ClassifierAccuracy assessmentPairsMcNemar’s
OAKAMPCFQUONBOFBOLC

MHBC94.40.9389.089.710088.095.1100MHBC-RF4.0000.039
RF91.10.8985.788.898.788.789.6100MHBC-CART4.2670.035
CART90.10.8887.293.997.385.390.796.2RF-CART0.0830.774

Note: MHBC: modified hierarchy-based classifier; RF: random forest; CART: classification and regression tree; OA: overall accuracy; KA: Kappa; MP: Masson pine; CF: Chinese fir; QU: Quercus; ONB: on-year bamboo; OFB: off-year bamboo; OLC: other land covers; value < 0.05 indicates a significant difference between two models.

4. Discussion

4.1. Design of a Proper Hierarchical Structure for Classification

The classification strategy based on hierarchical structures is not new and has been applied for land-cover classification [5153]. In previous research, the design of hierarchical structures was subjective and relied on the researcher’s experiences and characteristics of the study area under investigation [11, 53]; thus, the classification approach was not transferable. Improper hierarchical structure may reduce the accuracy of vegetation classification. In this study, construction of the hierarchical structure was automatically developed based on -score statistics. The tree structure is similar to the artificially established hierarchical system, but has its own characteristic. For example, the easily confused forest types are at the bottom of the tree, such as Masson pine and Chinese fir, impervious surfaces and bare soil, eucalyptus, and other broadleaf trees. However, the first layer of Gaofeng Forest Farm separates the water body first, rather than the traditional vegetation and nonvegetation. Although the first layer with parent node 1 of Huashi Township is split into vegetation and nonvegetation groups, Quercus, which has an obvious deciduous phenomenon in winter, is also temporarily classified as nonvegetation. This is different from the design of the hierarchical structure based on visual interpretation and expert knowledge by Chen et al. [11]. However, this kind of tree structure design may not be optimal, because the hierarchical structure depends on the existing experience of the experimenter. It has two main disadvantages: it is easy to make subjective judgment errors, resulting in grouping errors. The hierarchical structure is not from simple to complex. This research was conducted on the basis of the perspective of easy differentiation using -score based on training sample data and variables. This method has been proved to automatically adjust the hierarchical structure in different study areas and land-cover types to avoid the influence of human subjectivity and landscape heterogeneity.

Compared to other similar classifiers, such as CART and RF, the MHBC in this research has the following advantages: (1) classification tree structure was determined automatically based on existing samples and features, especially in subtropical areas with complex and diverse tree species, and is not influenced by subjective factors. (2) Compared to the RF algorithm that makes the overall optimal variable selection, MHBC selects the optimal variables for each node. (3) Compared to a black box in RF, the variables used in each node of the MHBC procedure are visible and easy to operate, and they are also explainable.

4.2. Selection of Optimal Variables from Multisource Data

Many previous studies have conducted comprehensive and comparative analyses of the effects of different data sources on classification accuracy [6, 19]. It is generally believed that making full use of the advantages of different data sources can improve classification accuracy [2, 10]. Spectral signatures and textural variables are fundamental for land cover and forest classification, and combinations of these variables are needed, especially for high spatial resolution imagery [19]. However, these kinds of data bring problems such as increased data volume and high correlation between spectral bands, which pose a great challenge to remote sensing classification. Considering the special characteristics of forest stand structures (e.g., canopy height and density) among different forest types, incorporation of Lidar-derived variables such as percentiles into spectral-based variables will be an effective way to improve tree species classification, as previous research confirmed [22, 54, 55]. In recent years, deep learning algorithms become increasingly popular for tree species classification based on high spatial resolution multisource data [21, 22]. Traditional statistical test techniques such as the Jefferies-Matusita distance are often used to test distance/separability between two tree species; however, these methods often do not automatically filter variables. As summarized by Pu [19], three popular feature selection methods—decision tree (i.e., CART or RF), correlation-based feature selection (CFS), and stepwise variable selection—can be used to identify optimal variables for specific tree species classes. In fact, these three methods are often used in combination. First, RF is used to rank variables importance for all tree species; then, the CFS method is used to remove the variables that are highly correlated with other variables but relatively less important; finally, automated forward and backward stepwise feature selection based on testing accuracy is used to determine key variables. Although RF is composed of numerous decision trees, it is difficult to find the important variables for each species [11]. Hierarchical selection of variables can solve this problem. Zhao et al. [35] used a hierarchy-based classification approach to select variables in each tree node based on integration of spectral, spatial, and canopy features and improved vegetation classification accuracy by 3.08%. This is consistent with the research results of our study.

5. Conclusions

The experiments in tree species classification in subtropical regions using high spatial resolution imagery indicate that the MHBC, through automatic determination of classification tree structure and selection of optimal variables at each node, confirmed its effectiveness and robustness. Major conclusions include the following: (1) the MHBC provided the highest overall classification accuracy of 85.19% for 10 land-cover classes and improved overall accuracies by 4.75% and 9.5% compared to RF and CART, respectively. (2) Optimal variables for each class can be identified using this proposed procedure. The important rankings of the different types of variables are . Spectral variables are often selected in the nodes and have a higher value than other variables. (3) The top layers (the easier classes to identify) of hierarchical structures need few variables to produce highly accurate results, but the lower layers require more variables to separate the relevant classes. (4) The proposed MHBC is robust and can be successfully transferred to other study areas to obtain better classification accuracy than RF and CART.

Data Availability

The complete source code is shared on the GitHub platform (https://github.com/YGwork123/Auto-Hierarchy).

Conflicts of Interest

The authors declare that there is no conflict of interest regarding the publication of this article.

Authors’ Contributions

X.J. and D.L conceived the ideas and designed methodology; X.J., S.Z., and D.L collected the data; X.J., S.Z., and D.L analyzed the data; X.J., Y.C., and D.L led the writing of the manuscript. All authors contributed critically to the drafts and gave final approval for publication.

Acknowledgments

This research is financially supported by the National Key R&D Program of China (2021YFD2200401). The authors would like to thank Ms. Yunhe Li for her help in image preprocessing of the validation site.

References

  1. G. Yu, Z. Chen, S. Piao et al., “High carbon dioxide uptake by subtropical forest ecosystems in the East Asian monsoon region,” Proceedings of the National Academy of Science, vol. 111, no. 13, pp. 4910–4915, 2014. View at: Publisher Site | Google Scholar
  2. F. E. Fassnacht, H. Latifi, K. Stereńczak et al., “Review of studies on tree species classification from remotely sensed data,” Remote Sensing of Environment, vol. 186, pp. 64–87, 2016. View at: Publisher Site | Google Scholar
  3. G. Ewa, F. David, and O. Katarzyna, “Evaluation of machine learning algorithms for forest stand species mapping using Sentinel-2 imagery and environmental data in the Polish Carpathians,” Remote Sensing of Environment, vol. 251, article 112103, 2020. View at: Publisher Site | Google Scholar
  4. A. M. Lechner, G. M. Foody, and D. S. Boyd, “Applications in remote sensing to forest ecology and management,” One Earth, vol. 2, no. 5, pp. 405–412, 2020. View at: Publisher Site | Google Scholar
  5. S. Liu, X. Wei, D. Li, and D. Lu, “Examining forest disturbance and recovery in the subtropical forest region of Zhejiang province using Landsat time-series data,” Remote Sensing, vol. 9, no. 5, p. 479, 2017. View at: Publisher Site | Google Scholar
  6. X. Yu, D. Lu, X. Jiang et al., “Examining the roles of spectral, spatial, and topographic features in improving land-cover and forest classifications in a subtropical region,” Remote Sensing, vol. 12, no. 18, p. 2907, 2020. View at: Publisher Site | Google Scholar
  7. A. Axelsson, E. Lindberg, H. Reese, and H. Olsson, “Tree species classification using Sentinel-2 imagery and Bayesian inference,” International Journal of Applied Earth Observation and Geoinformation, vol. 100, article 102318, 2021. View at: Publisher Site | Google Scholar
  8. L. Ghayour, A. Neshat, S. Paryani, H. Shahabi, and A. Ahmad, “Performance evaluation of Sentinel-2 and Landsat 8 OLI data for land cover/use classification using a comparison between machine learning algorithms,” Remote Sensing, vol. 13, no. 7, p. 1349, 2021. View at: Publisher Site | Google Scholar
  9. L. Li, N. Li, D. Lu, and Y. Chen, “Mapping Moso bamboo forest and its on-year and off-year distribution in a subtropical region using time-series Sentinel-2 and Landsat 8 data,” Remote Sensing of Environment, vol. 231, article 111265, 2019. View at: Publisher Site | Google Scholar
  10. Z. Xie, Y. Chen, D. Lu, G. Li, and E. Chen, “Classification of land cover, forest, and tree species classes with ZiYuan-3 multispectral and stereo data,” Remote Sensing, vol. 11, no. 2, p. 164, 2019. View at: Publisher Site | Google Scholar
  11. Y. Chen, S. Zhao, Z. Xie, D. Lu, and E. Chen, “Mapping multiple tree species classes using a hierarchical procedure with optimized node variables and thresholds based on high spatial resolution satellite data,” GIScience & Remote Sensing, vol. 57, no. 4, pp. 526–542, 2020. View at: Publisher Site | Google Scholar
  12. D. Lu, G. Li, E. Moran, L. Dutra, and M. Batistella, “A comparison of multisensor integration methods for land cover classification in the Brazilian Amazon,” GIScience & Remote Sensing, vol. 48, no. 3, pp. 345–370, 2011. View at: Publisher Site | Google Scholar
  13. J. K. Gillan, J. W. Karl, M. Duniway, and A. Elaksher, “Modeling vegetation heights from high resolution stereo aerial photography: an application for broad-scale rangeland monitoring,” Journal of Environmental Management, vol. 144, pp. 226–235, 2014. View at: Publisher Site | Google Scholar
  14. O. Stromann, A. Nascetti, O. Yousif, and Y. Ban, “Dimensionality reduction and feature selection for object-based land cover classification based on Sentinel-1 and Sentinel-2 time series using Google Earth Engine,” Remote Sensing, vol. 12, p. 76, 2020. View at: Publisher Site | Google Scholar
  15. D. Lu and Q. Weng, “A survey of image classification methods and techniques for improving classification performance,” International Journal of Remote Sensing, vol. 28, no. 5, pp. 823–870, 2007. View at: Publisher Site | Google Scholar
  16. B. I. Anis, “Variable selection using support vector regression and random forests: a comparative study,” Intelligent Data Analysis, vol. 20, no. 1, pp. 83–104, 2016. View at: Publisher Site | Google Scholar
  17. S. Rajbhandari, J. Aryal, J. Osborn, A. Lucieer, and R. Musk, “Leveraging machine learning to extend ontology-driven geographic object-based image analysis (O-GEOBIA): a case study in forest-type mapping,” Remote Sensing, vol. 11, no. 5, p. 503, 2019. View at: Publisher Site | Google Scholar
  18. A. E. Maxwell, T. A. Warner, and F. Fang, “Implementation of machine-learning classification in remote sensing: an applied review,” International Journal of Remote Sensing, vol. 39, no. 9, pp. 2784–2817, 2018. View at: Publisher Site | Google Scholar
  19. R. Pu, “Mapping tree species using advanced remote sensing technologies: a state-of-the-art review and perspective,” Journal of remote sensing, vol. 2021, pp. 1–26, 2021. View at: Publisher Site | Google Scholar
  20. K. Cao and X. Zhang, “An improved Res-Unet model for tree species classification using airborne high-resolution images,” Remote Sensing, vol. 12, no. 7, p. 1128, 2020. View at: Publisher Site | Google Scholar
  21. B. Zhang, L. Zhao, and X. Zhang, “Three-dimensional convolutional neural network model for tree species classification using airborne hyperspectral images,” Remote Sensing of Environment, vol. 247, article 111938, 2020. View at: Publisher Site | Google Scholar
  22. J. Mayra, S. Keski-Saari, S. Kivinen, T. Tanhuanp, and P. Vihervaara, “Tree species classification from airborne hyperspectral and lidar data using 3D convolutional neural networks,” Remote Sensing of Environment, vol. 256, article 112322, 2021. View at: Publisher Site | Google Scholar
  23. X. Pan, Z. Wang, Y. Gao, X. Dang, and Y. Han, “Detailed and automated classification of land use/land cover using machine learning algorithms in Google Earth Engine,” Geocarto International, vol. 1–18, pp. 1–18, 2021. View at: Publisher Site | Google Scholar
  24. L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp. 123–140, 1996. View at: Publisher Site | Google Scholar
  25. S. Elmahdy, T. Ali, and M. Mohamed, “Regional mapping of groundwater potential in Ar Rub Al Khali, Arabian Peninsula using the classification and regression trees model,” Remote Sensing, vol. 13, no. 12, p. 2300, 2021. View at: Publisher Site | Google Scholar
  26. Y. Shao and R. Lunetta, “Comparison of support vector machine, neural network, and CART algorithms for the land-cover classification using limited training data points,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 70, pp. 78–87, 2012. View at: Publisher Site | Google Scholar
  27. Y. Chen, X. Song, S. Wang, J. Huang, and L. Mansaray, “Impacts of spatial heterogeneity on crop area mapping in Canada using MODIS data,” ISPRS Journal of Photogrammetry & Remote Sensing, vol. 119, pp. 451–461, 2016. View at: Publisher Site | Google Scholar
  28. K. Fawagreh, M. M. Gaber, and E. Elyan, “Random forests: from early developments to recent advancements,” Systems Science & Control Engineering, vol. 2, no. 1, pp. 602–609, 2014. View at: Publisher Site | Google Scholar
  29. F. Zhang and X. Yang, “Improving land cover classification in an urbanized coastal area by random forests: the role of variable selection,” Remote Sensing of Environment, vol. 251, no. 10, article 112105, 2020. View at: Publisher Site | Google Scholar
  30. P. Gong, J. Wang, L. Yu, Y. Zhao, and J. Chen, “Finer resolution observation and monitoring of global land cover: first mapping results with Landsat TM and ETM+ data,” International Journal of Remote Sensing, vol. 34, no. 7, pp. 2607–2654, 2013. View at: Publisher Site | Google Scholar
  31. Y. Gao, D. Lu, G. Li et al., “Comparative analysis of modeling algorithms for forest aboveground biomass estimation in a subtropical region,” Remote Sensing, vol. 10, no. 4, p. 627, 2018. View at: Publisher Site | Google Scholar
  32. C. Pelletier, S. Valero, J. Inglada, N. Champion, and G. Dedieu, “Assessing the robustness of random forests to map land cover with high resolution satellite image time series over large areas,” Remote Sensing of Environment, vol. 187, pp. 156–168, 2016. View at: Publisher Site | Google Scholar
  33. X. Zhang, L. Liu, X. Chen, S. Xie, and Y. Gao, “Fine land-cover mapping in China using Landsat datacube and an operational speclib-based approach,” Remote Sensing, vol. 11, no. 9, p. 1056, 2019. View at: Publisher Site | Google Scholar
  34. H. Zhang and M. Wang, “Search for the smallest random forest,” Statistics and Its Interface, vol. 2, no. 3, pp. 381–388, 2009. View at: Publisher Site | Google Scholar
  35. S. Zhao, X. Jiang, G. Li, Y. Chen, and D. Lu, “Integration of ZiYuan-3 multispectral and stereo imagery for mapping urban vegetation using the hierarchy-based classifier,” International Journal of Applied Earth Observations and Geoinformation, vol. 105, article 102594, 2021. View at: Publisher Site | Google Scholar
  36. Y. Yu, M. Zhang, Z. Zhong, and J. Lin, “Dynamic change analysis of forest resources in Guangxi Gaofeng Forest Farm during 1989–2018,” Journal of Fujian Forestry Science and Technology, vol. 192, no. 3, pp. 109–114, 2020. View at: Publisher Site | Google Scholar
  37. N. Brodu, “Super-resolving multiresolution images with band-independent geometry of multispectral pixels,” IEEE Transactions on Geoscience & Remote Sensing, vol. 55, no. 8, pp. 4610–4617, 2017. View at: Publisher Site | Google Scholar
  38. S. A. Soenen, D. R. Peddle, and C. A. Coburn, “SCS+C: a modified sun-canopy-sensor topographic correction in forested terrain,” IEEE Transactions on Geoscience & Remote Sensing, vol. 43, no. 9, pp. 2148–2159, 2005. View at: Publisher Site | Google Scholar
  39. D. Liu and F. Xia, “Assessing object-based classification: advantages and limitations,” Remote Sensing Letters, vol. 1, no. 4, pp. 187–194, 2010. View at: Publisher Site | Google Scholar
  40. Y. Chen, Z. Peng, Y. Ye, X. Jiang, D. Lu, and E. Chen, “Exploring a uniform procedure to map _eucalyptus_ plantations based on fused medium -high spatial resolution satellite images,” International Journal of Applied Earth Observations and Geoinformation, vol. 103, article 102462, 2021. View at: Publisher Site | Google Scholar
  41. D. Lu, G. Li, E. Moran, L. Dutra, and M. Batistella, “The roles of textural images in improving land-cover classification in the Brazilian Amazon,” International Journal of Remote Sensing, vol. 35, no. 24, pp. 8188–8207, 2014. View at: Publisher Site | Google Scholar
  42. J. W. Rouse, R. W. Haas, J. A. Schell, D. W. Deering, and J. C. Harlan, “Monitoring the Vernal Advancement and Retrogradation (Greenwave Effect) of Natural Vegetation,” in NASA/GSFCT Type III Final Report, Greenbelt, MD, USA, 1974. View at: Publisher Site | Google Scholar
  43. S. Mcfeeters, “The use of the normalized difference water index (NDWI) in the delineation of open water features,” International Journal of Remote Sensing, vol. 17, no. 7, pp. 1425–1432, 1996. View at: Publisher Site | Google Scholar
  44. E. R. Hunt and B. N. Rock, “Detection of changes in leaf water content using near- and middle-infrared reflectances,” Remote Sensing of Environment, vol. 30, no. 1, pp. 43–54, 1989. View at: Publisher Site | Google Scholar
  45. J. Dash and P. J. Curran, “The MERIS terrestrial chlorophyll index,” International Journal of Remote Sensing, vol. 25, no. 549, pp. 151–161, 2004. View at: Publisher Site | Google Scholar
  46. A. A. Gitelson and M. N. Merzlyak, “Spectral reflectance changes associated with autumn senescence of _Aesculus hippocastanum_ L. and _Acer platanoides_ L. leaves. Spectral features and relation to chlorophyll estimation,” Journal of Plant Physiology, vol. 143, no. 3, pp. 286–292, 1994. View at: Publisher Site | Google Scholar
  47. J. G. P. W. Clevers and A. A. Gitelson, “Remote estimation of crop and grass chlorophyll and nitrogen content using red-edge bands on Sentinel-2 and -3,” International Journal of Applied Earth Observation and Geoinformation, vol. 23, pp. 344–351, 2013. View at: Publisher Site | Google Scholar
  48. R. Haralick, K. Shanmugam, and I.'. H. Dinstein, “Textural features for image classification,” Studies in Media and Communication, vol. SMC-3, no. 6, pp. 610–621, 1973. View at: Publisher Site | Google Scholar
  49. R. G. Congalton and K. Green, Assessing the Accuracy of Remotely Sensed Data: Principles and Practices (Third Edition), CRC Press, Boca Raton, FL, 2019. View at: Publisher Site
  50. S. Raschka, “MLxtend: providing machine learning and data science utilities and extensions to Python’s scientific computing stack,” The Journal of Open Source Software, vol. 3, no. 24, p. 638, 2018. View at: Publisher Site | Google Scholar
  51. D. Sulla-Menashe, M. A. Friedl, O. N. Krankina et al., “Hierarchical mapping of Northern Eurasian land cover using MODIS data,” Remote Sensing of Environment, vol. 115, no. 2, pp. 392–403, 2011. View at: Publisher Site | Google Scholar
  52. D. Lu, S. Hetrick, E. Moran, and G. Li, “Application of time series Landsat images to examining land-use/land-cover dynamic change,” Photogrammetric Engineering & Remote Sensing, vol. 78, no. 7, pp. 747–755, 2012. View at: Publisher Site | Google Scholar
  53. L. Jiao, W. Sun, G. Yang, G. Ren, and Y. Liu, “A hierarchical classification framework of satellite multispectral/hyperspectral images for mapping coastal wetlands,” Remote Sensing, vol. 11, no. 19, p. 2238, 2019. View at: Publisher Site | Google Scholar
  54. A. Hovi, L. Korhonen, J. Vauhkonen, and I. Korpela, “LiDAR waveform features for tree species classification and their sensitivity to tree- and acquisition related parameters,” Remote Sensing of Environment, vol. 173, pp. 224–237, 2016. View at: Publisher Site | Google Scholar
  55. H. Zhao, Y. Zhong, X. Wang et al., “Mapping the distribution of invasive tree species using deep one-class classification in the tropical montane landscape of Kenya,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 187, pp. 328–344, 2022. View at: Publisher Site | Google Scholar

Copyright © 2022 Xiandie Jiang et al. Exclusive Licensee Aerospace Information Research Institute, Chinese Academy of Sciences. Distributed under a Creative Commons Attribution License (CC BY 4.0).

 PDF Download Citation Citation
Views39
Downloads25
Altmetric Score
Citations