Urban Impervious Surface Extraction Based on Multi-Features and Random Forest

Impervious surface data has become an indicator of the degree of urbanization and the environmental quality, which inspires the widely use of remote sensing technology in extracting impervious surface. In order to reduce the confusion between impervious surface and other landcover types, this paper proposed an effective urban impervious surface extraction method based on multi-features and Random Forest (RF). First, Sentinel-2 multispectral data and Luojia 1–01 images are employed to extracted multi-features, including spectral features, texture features and temporal features. Then, a feature selection method with null importance is proposed to remove irrelevant features. Finally, Probabilistic Label Relaxation (PLR) is introduced into RF to obtain the classification results and impervious surface. The main urban areas of Zhengzhou and Hangzhou are selected as study areas. The experiment results show that the integration of the multi-features can significantly improve the overall accuracy of classification and the extraction accuracy of impervious surface. And the classification accuracy can be further improved after the feature selection with null importance. Besides, the PLR method can effectively reduce the salt-and-pepper phenomenon of classification results using random forest, which presupposes the optimal number of iterations. The method proposed in this paper can effectively improve the estimation of impervious surface and provide an important reference for the extraction of impervious surface based on pixel level.


I. INTRODUCTION
Impervious surface is natural or artificial surface that prevents surface water from penetrating into the soil and thereby altering the flow of flood runoff, such as roads, car parks and rooftops made of cement concrete, glass, asphalt, plastics, tiles, metals, etc. High-precision impervious surface estimation is of great significance for environmental management, urban planning, urban construction, hydrological environment and economic development [1]- [7]. Remote sensing technology has become the main technique for impervious surface mapping due to its large monitoring range, fast acquisition speed and low cost. In recent years, a variety of methods for extracting impervious surface using remote sensing data have been successively proposed and successfully applied to urban planning, environmental monitoring, water resource The associate editor coordinating the review of this manuscript and approving it for publication was Hongjun Su. management and other fields, which are mainly divided into sub-pixel methods [8]- [10] and traditional per-pixel methods [11]- [13]. Pixel-based methods are easy to operate and understand. However, the spectral confusion is ineluctable due to the heterogeneity of urban landscape and the spectral resolution of images, particularly the confusion between impervious surface and bare land.
In order to reduce the confusion, the integration of multi-features has been proposed. Many indexes have been produced to enhance the information of impervious surface and can be served as auxiliary spectral features to extract impervious surface, including Normalized Difference Built -up Index (NDBI) [14], Index -based Built-up Index (IBI) [15], Enhanced Built-Up and Bareness Index (EBBI) [16], Vegetated Adjusted NTL Urban Index (VANUI) [17], Modified Normalized Difference Impervious Surface Index (MNDISI) [18], etc. Additionally, the impervious surface areas are often artificial geographical objects VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ with regular geometric shapes, which highlights the role of texture features in distinguishing impervious surface from bare land and improving the extraction accuracy of impervious surface [19]- [21]. Besides, compared with the areas covered with soil, the spectrum of the impervious surface has less change in time series, because the areas overlaid with soil can be vegetated for some time, which makes it effective to improve the estimation of impervious surface by embedding temporal information [22]. Spectral features, texture features and temporal features are all helpful to improve the extraction accuracy of impervious surface, but few researches integrate these three kinds of features. The method of integrating multi-features can greatly enrich the information content of remote sensing data, but it will cause information redundancy and decrease the rate of classification, which is not conducive to the effective utilization of remote sensing data. Therefore, it is necessary to prioritize all features and eliminate redundant ones. The most commonly used methods to quantify the importance of features are linear models and decision trees, but linear classifiers fail to find complex dependencies in training data and decision trees lack stability and smoothness [23]. The feature ranking based on Random Forest (RF) can effectively assist us to overcome these problems, but it has been found that RF-based importance measures are biased when the categorical variables have a large number of categories [24]. Moreover, when one of the correlated features is employed by the model, the other will have decaying importance. Altmann et al. [23] proposed a permutation importance method to correct the measurement of feature importance, which can improve the prediction accuracy of the random forest model and enhance the interpretability of the model. Based on the permutation importance method, we proposed the feature selection method with null importance which can be severed as a solution of mass features.
Machine learning algorithm is an advanced method with comparative advantages in impervious surface extraction which can overcome the subjective factors and improve the concrete algorithm automatically in the process of learning [25]. The Artificial Neural Network (ANN), Support Vector Machine (SVM), Classification and Regression Trees (CART) and Random Forest (RF) are most widely used in machine learning algorithm [26]- [28]. However, ANN has difficulty in determining initial weight and it is difficult to for SVM to explain how its kernel function selection affects classification accuracy. CART is sensitive to data noise and has poor learning ability. By contrast, RF is a nonparametric method and it has stronger generalization ability and learning ability, which makes it perform well in classification using remote sensing images [29]. Hayes and Murphy [30] used the ensemble Random Forest classifier to classify land cover from National Agricultural Imagery Program (NAIP) image of southeast Wyoming in 2009 with a spatial resolution of 10 m. Torbick and Corbiere [31] combined National Land Cover Dataset (NLCD) products and Landsat images to extract high-precision urban impervious surface by the way of Classification And Regression Tree (CART) using random forest, which combined with indices including Normalized Difference Vegetation Index (NDVI), Land Surface Water Index (LSWI), Soil-adjusted Total Vegetation Index (SATVI) and Modified Soil-adjusted Vegetation Index (MSAVI). Although the random forest algorithm achieves a high classification accuracy, the strategy based on the voting method fails to consider the spatial neighborhood information of pixels. Some researcher introduced Markov Random Field (MRF) into classification to describe the spatial neighbor information, and experiments proved that the use of spatial neighborhood features can effectively improve the image classification accuracy [32], [33]. Compared with the classical Markov Random Field, the Probabilistic Label Relaxation (PLR) can provide a better understanding of the spatial neighborhood relationship of pixels and its calculation is simpler [34], [35].
The purpose of this research is to improve the extraction accuracy of urban impervious surface. A combination of multi-features and random forest classifier method is adopted. Besides, feature selection with null importance and probabilistic label relaxation are employed to refine the strategy of extracting impervious surface areas. The main contributions of our work are listed as follows: 1. Multi-features including spectral features, texture features and temporal features are integrated. It can give full play to the advantages of various features in the identification of impervious surface. 2. A feature selection method with null importance is proposed to eliminate redundant which corrects the measurement of feature importance in common RF. 3. PLR is introduced into RF which can reduce the salt and pepper phenomenon by taking the spatial neighbor information into account.

A. STUDY AREAS AND DATASETS
To demonstrate the effectiveness of the proposed method, the main urban areas of Zhengzhou City and Hangzhou City are selected as the study areas. Zhengzhou City is located in the south of the North China Plain and the lower Yellow River, known as an important comprehensive transportation hub of China. In recent years, the urbanization of Zhengzhou has developed rapidly, and the city has expanded unprecedentedly, which includes five administrative units as districts, Zhongyuan, Jinshui, Erqi, Guancheng and Huiji, covering an area of 990 square kilometers ( Figure 1). Hangzhou, located in the south of the Yangtze River Delta in China, is an important central city of the Yangtze River Delta and a transportation hub in southeastern China. The urbanization process of Hangzhou is accelerating and the impervious surface areas are increasing rapidly. The study areas of Hangzhou City cover an area of 680 square kilometers including six administrative units as districts, Xihu, Gongshu, Jianggan, Shangcheng, Xiacheng and Binjiang ( Figure 1).  Major datasets used in this research include Sentinel-2 multispectral data, Luojia 1-01 images and high spatial resolution images from Google Earth, which are outlined in Table 1. Considering that some of bare land can be covered by vegetation in summer which is easy to distinguish from impervious surface, the Sentinel-2 images in summer were selected for the extraction of impervious surface in both study areas. For Zhengzhou, the Sentinel-2 image acquired on July 7, 2019 was used to extract the impervious surface. For Hangzhou, the Sentinel-2 image acquired on August 17, 2019 was employed. However, luxuriant trees in summer can shade the surrounding impervious surface, which lead to the underestimation of impervious surface. In order to reduce this effect, single image in spring, autumn and winter was selected for temporal feature extraction. For Zhengzhou, Sentinel-2 images were acquired on April 12 (spring), October 4 (autumn) and December 12 (winter). For Hangzhou, Sentinel-2 images were acquired on March 20 (spring), October 20 (autumn) and December 10 (winter). The Luojia 1-01 images of Zhengzhou and Hangzhou acquired on October 30, 2018 were used to enhance impervious surface information. In this study, all images need to be resampled to 10 m resolution and the coregistration was conducted by several manually selected control points for two kinds of images. Combined with the field data of the research areas and the high spatial resolution images of Google Earth, a total of 9,152 sample points were collected for Zhengzhou and 8,765 sample points were collected for Hangzhou, which were well distributed. They were divided into 4 types which were impervious surface, bare land, water body and vegetation. For Zhengzhou, 8,152 points were used as training samples and the remaining 1,000 points were used as validation samples. For Hangzhou, the training samples included 7,765 points and the validation samples were 1,000 points.

B. MULTI-FEATURES EXTRACTION AND SELECTION
As the multi-spectral image, Sentinel-2 has strong ability to identify landmarks, but every single band contains limited information. Based on the original multi-spectral bands, the Vegetation Adjusted NTL (nighttime light) Urban Index (VANUI), the first and second principal component of Principal Component Transform (PCT), and Normalized Difference Built-up Index (NDBI) were extracted as auxiliary spectral features to improve the classification results, VOLUME 8, 2020 especially the accuracy of impervious surface. PCT is a common method for image enhancement and it can reduce the dimensions of data and eliminate redundant information. NDBI can serve as a worthwhile alternative for quickly mapping built-up areas, which is expressed as follows: where ρ SWIR1 and ρ NIR are the band 11 and band 8 of Sentinel-2 images. Vegetated Adjusted NTL Urban Index (VANUI) is an urban index developed by combining NTL with NDVI to reduce light saturation and increase variation of data values in the city center [17] which has been proved has a strong correlation with impervious surface [36]. Since Luojia 1-01 data has fine spatial resolution and high radiometric quantization, it is combined with NDVI derived from sentinel-2 images to obtain the VANUI. The VANUI is calculated from the following expression: where NDVI is derived from Sentinel-2 and NTL is collected by Luojia 1-01. The NDVI is defined by where ρ NIR and ρ RED are the band 8 and band 4 of Sentinel-2 images. The extraction of impervious surface supplemented by texture information is helpful to draw the distinction between impervious surface and other ground objects. Statistical methods are most widely used to extract texture parameters, among which Gray-Level Co-occurrence Matrix (GLCM) is the top. The GLCM is used to extract the texture by calculating the frequency of the grayscale of pixel pairs with fixed relative positions. Haralick [37] extracted 14 texture features based on GLCM to describe the texture information of the image. Combined with previous researches and many experiments, the size of the moving window was determined to be 3 × 3 and the first and second principal components of PCT were implemented to calculate eight texture features based on GLCM, which contains Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment and Correlation.
Impervious surface of urban areas will be sheltered by lush trees within one period, and using images in this period to extract impervious surface can lead to the underestimation of its result. To counteract this effect, for Zhengzhou, the NDVI acquired form the sentinel-2 images on April 12, July 7, October 14 and December 12,2019 were employed as the temporal features. And for Hangzhou, they are acquired form the images on April 12, July 7, October 14 and December 12,2019. Because in different seasons, impervious surface may be covered by trees in varying degrees, which provides a potential possibility to identify impervious surface.
This paper proposed a method of integrating multi-features including spectral features, texture features and temporal features. Each feature obtained in this study corresponds to a layer, and the method of layer overlay is used for feature fusion. The integration of multi-features can greatly enrich the information content of remote sensing data. But it will also cause information redundancy and decrease the rate of classification. This study proposed a feature selection method with null importance, which corrected the RF-based importance measures. It permutes land cover types to test actual importance significance against the distribution of feature importance using shuffled class labels. The importance is quantified by the total gain across all splits the feature is used in.
Firstly, we need to create the distribution of null importance for all features by continuously shuffling the land cover types and training the model. The performance of each feature when fitted to shuffled class labels can help to figure out how the model makes sense of this feature irrespective of the land cover types. Secondly, it is necessary to fit the model on the actual land cover types and gain the feature importance. This provides a benchmark whose significance can be tested against the null importance distribution. Finally, we use the log actual feature importance divided by the 75 percentile of null distribution to score features. Higher score indicates that the feature is more important in this model. The score value less than 0 indicates that this feature has no relationship with the land cover type. We plan to integrate features with score greater than 0.

C. COMBINATION OF RF AND PLR
The Random Forest (RF) classifier is an integrated machine learning algorithm based on decision trees. Bagging method is adopted in RF to generate independent identically distributed training sample sets for each decision tree and the final classification result of the RF depends on the voting of all decision trees. RF performs well in remote sensing image classification, but it fails to consider the spatial neighborhood information of pixels. Probabilistic Label Relaxation (PLR) is a kind of iterative labeling algorithm considering the neighborhood characteristics around pixels [35]. This basic idea is based on the fact that the adjacent pixels in the image have a certain correlation and continuous areas of certain sizes are to be expected in reality. The category belonging information of each pixel is related to its local observation information and neighborhood information. Using PLR is helpful to remove isolated pixels caused by noise or misclassification in the classification results and can make the classification results consistent in spectrum and space.
Before using probabilistic relaxation, it is assumed that the probability that the pixel n falls into class w j is already known, which is p n (w j ). First, the prior compatibility coefficient p on w i | w j is defined to describe the probability that the pixel o belongs to w i when the neighborhood pixel n belongs to category w j , where ρ quantifies the probability that adjacent pixels are homogeneous. The probability of the current pixel o having class label w i when the neighborhood pixel n belongs to class w j can be written as a neighborhood function denoting the cumulative contribution of the neighborhood pixel labels to the current pixel class can be defined as Utilizing the spatial neighborhood, the class probability of pixel o can be revised as where t indicates the current iteration step. p t o (w i ) of pixel o should be updated during each iteration. After enough times of iterations, p o (w i ) will converge, and the termination condition of iteration is This paper combines the RF and PLR to improve the classification accuracy and the precision of impervious surface extraction. The algorithm flow is as follows: (1) The trained random forest model is used to count the votes of each pixel belonging to various classes, so as to calculate the category probability of each pixel.
(2) According to the PLR model, the category probability of pixels is modified by using the neighborhood characteristics between pixels.
(3) The final classification result is obtained by the corrected class probability based on the maximum a posteriori criterion.
According to the characteristics of the impervious surface, this paper proposed to extract spectral features, texture features and temporal features, and selected optimal features with null importance and finally adopt a neighborhood-supported classification combining RF and PLR. The methodologic flowchart of this research is shown in Figure 2.

A. INTEGRATION OF MULTI-FEATURES AND FEATURE SELECTION
Combined with Sentinel-2 multispectral data and Luojia 1-01 images, multi-features were extracted including spectral features, texture features and temporal features. Table 2 shows that there are a total of 34 features extracted in this study.
In the process of features selection, for Zhengzhou, the number of decision trees of random forest was set at 111 (Table 3), and the number of features considered in searching for optimal segmentation was √ N , where N represented the total number of input features, and the shuffle times of class label was 80. For Hangzhou, the number of decision trees was set at 121 (Table 3) and the number of features and shuffle times were the same as Zhengzhou. The scores and ranking results of the feature importance are shown in Figure 3. If the value of feature score is less than 0, this feature has almost no influence on classification results. For Zhengzhou, 20 bands with scores greater than 0 were finally reserved as feature selection results. For Hangzhou, there exist 19 bands whose scores were greater than 0, which were selected to extract impervious surface. VOLUME 8, 2020

B. CLASSIFICATION USING RF
To demonstrate the effectiveness of the integration of multifeatures and feature selection with null importance, this research designed three classification experiments base on RF by using following different input data: (1) original bands (2) original bands and multi-features (3) feature selection results.
It is well known that the number of decision trees can directly affect the performance of random forest classifier. A small number of decision trees will make it impossible for  the random forest model to make full use of all features, while an excessive number of decision trees will reduce the training efficiency of the random forest. Grid Search algorithm is a method to optimize the model performance by traversing the given parameter combination, but its final performance is closely related to the initial data partition results. In order to deal with this situation, cross validation is often used to reduce the contingency. In this paper, Grid Search with Cross Validation was used to select the optimum number of decision trees for the six groups of experiments respectively. Table 3 shows the optimum number of decision trees for six classification experiments and their cross validation scores.
The optimal number of decision trees in each group was used for its random forest classification respectively and the number of features considered in searching for optimal segmentation was the root of the total number of input features. Table 4 provides a better understanding of urban land cover classification results. For Zhengzhou, when using only original bands, the confusion between impervious surface and bare land was serious, as well as the confusion between impervious surface and vegetation. In the classification results, 58 pixels of impervious surface were misclassified as bare land and 26 pixels were misclassified as vegetation. However, after the integration of multi-features, those two numbers were reduced to 33 and 6. The integration of multi-features made the overall accuracy and Kappa coefficient increased by 8.6% (=90.60%-82.00%) and 0.1266 (=0.8604-0.7338). For Hangzhou, when using only original bands, 43 pixels of impervious surface were misclassified as bare land and 25 pixels were misclassi-VOLUME 8, 2020 fied as vegetation, which indicated the confusion between impervious surface and these two classes. Nevertheless, those two misclassification numbers were reduced to 21 and 11 after the integration of multi-features. And the overall accuracy and Kappa coefficient were increased by 5.8% (=93.00%-87.20%) and 0.0892 (=0.9091-0.8199). This result indicated that the integration of multi-features could effectively improve the classification results by reducing the confusion between impervious surface and bare land or vegetation.
The strategy of feature selection proved its strength in improving the classification results. For Zhengzhou, when using feature selection results, the overall accuracy increased from 90.60% to 91.10%, and the Kappa coefficient was improved from 0.8604 to 0.8678. For Hangzhou, the overall accuracy increased from 93.00% to 93.90%, and the Kappa coefficient was improved from 0.9091 to 0.9137. In the case of sufficient training samples, when there are too many features involved in classification, the features of redundancy and poor performance may lead to the reduction of classification accuracy. After feature selection with null importance, the unnecessary features can be eliminated, which not only improves the speed of modeling, but also improves the classification accuracy.

C. IMPERVIOUS SURFACE EXTRACTION WITH PLR-RF
The Probabilistic Label Relaxation (PLR) method can reduce the noise interference in classification results and eliminate the salt-and-pepper classification error. With the increase of PLR iteration times, isolated pixels in the classification results will decrease, but too many iterations will lead to the expansion of the neighborhood, and irrelevant spatial information will mislead the classification results. To find the optimal number of iterations, the experiments set different iteration times of PLR algorithm on the basis of random forest classifier using selected features and calculated the classification accuracy. For both two study areas, it indicated that with the increase of iteration times, the overall accuracy and Kappa coefficient all increased first and then decreased ( Figure 4). For Zhengzhou, when the number of iterations was 3, the overall accuracy and Kappa coefficient reached the maximum. As the number of iterations continued to increase, the classification accuracy showed a downward trend. When the number of iterations increased to 5, the classification result was even worse than the situation when the PLR method was not used. The experiments proved that the optimal number of iterations of PLR for Zhengzhou was 3. For Hangzhou, the optimal number of iterations of PLR was 2. For both two study areas, when the number of iterations passed the optimal value, the overall accuracy and Kappa coefficient all showed a sharp drop. Moreover, the rate of decline increased with the increase of iteration number. The results prove that PLR can make full use of the spatial information based on the optimal number of iterations, but exceeding the optimum number will lead to the expansion of the neighborhood, and the introduction of invalid spatial information will lead to misclassification. The success of PLR depends on its iteration number.    Table 5 shows the confusion matrices for the landcover classification with optimal iteration number of PLR for two cities. For Zhengzhou, we could know that the overall accuracy and Kappa coefficient were 94.10% and 0.9123, which increased by 3% (=94.10%-91.10%) and 0.0445 (=0.9123-0.8678) respectively compared with the classification results without using PLR in Table 4. For Hangzhou, the overall accuracy and Kappa coefficient were 96.00% and 0.9433, which increased by 2.1% (=96.00%-93.90%) and 0.0296 (=0.9433-0.9137). These results highlighted the advantage of PLR with optimal iteration number in reducing the salt and pepper phenomenon and improving the classification results.
When extracting the impervious surface, other three kinds of land covers including the bare land, water body and vegetation were combined as non-impervious surface. VOLUME 8, 2020  New confusion matrices were calculated in Table 6 in order to make the results more intuitive. The results in Table 6 show the consistency with that in Table 5, but the overall accuracy is generally higher for the removal of confusion among three non-impervious classes. To illustrate the strength of proposed method, Figure 5 and Figure 6 show the impervious surface extraction results with a local zoom for two cities. For Zhengzhou, when using only original bands, the road could be extracted very well (Figure 5(b)). However, impervious surface with small areas was not easily identifiable, which could be improved by the integration of multi-features and feature selection (Figure 5(c)). Water body with high  Red bars represent the actual importance of the feature, and blue bars represent the null importance with shuffled landcover class. The total number of shuffle is 80. The importance is quantified by the total gain across all splits the feature is used in. If the actual importance of the feature is more significant than null importance, this feature is useful, and greater difference indicates greater importance. sediment concentration was easily confused with impervious surface and it caused the salt and pepper phenomenon. The PLR method could significantly reduce this interference in the classification by taking spatial information into account ( Figure 5(d)). For Hangzhou, it could be seen that many pixels of bare land were mistakenly classified as impervious VOLUME 8, 2020 surface, especially those pixels near the impervious surface ( Figure 6(b)). This led to an overestimation of the impervious surface. The integration of multi-features and feature selection could improve the overestimation (Figure 6(c)) and PLR could reduce the salt and pepper phenomenon ( Figure 6(d)). Figure 7 shows the impervious surface extraction results using proposed method. For Zhengzhou city, the impervious surface areas are mainly concentrated in its central and southern part, and the impervious surface areas in Huiji District are significantly less than that in the other four districts, which is consistent with the slow development in the northwest of Zhengzhou. For Hangzhou, the impervious surface areas are mainly concentrated in its eastern and northern part, and the Xihu District has the smallest impervious surface areas. This result also accords with the actual situation of urban development.

IV. DISCUSSION
In this paper, the impervious surface extraction results of GlobalLand30 in 2020 (http://www.globallandcover.com/) are collected to compare with the impervious surface extracted by our method. GlobalLand30 mainly uses Landsat images with a resolution of 30 meters, and it can be found that the impervious surface areas in the core of the both two cities are significantly overestimated (Figure 8(a), (b)).
In GlobalLand30, a large number of non-impervious surface areas near impervious surface areas are misclassified, and impervious surface at the edge of the city cannot be effectively identified (Figure 9(a), (b)). Compared with GlobalLand30, the method proposed in this paper uses the sentinel-2 data with higher resolution and combines the night-time light images to extract multi-features. Therefore, it can identify the impervious surface with small areas and it is more sensitive to the impervious surface at the edge of the city. GlobalLand30 is suitable for the mapping of impervious surface at large scale, and our method is dedicated to extracting the impervious surface with high precision.
Multi-features employed in this research depend on their importance measures. For both two cities, the feature selection results show that all original bands of Sentinel-2 images are useful, especially the B11 and B12. When creating the null importance distribution, the actual feature importance distribution of B11 and B12 is obviously far from the feature importance distribution of shuffled category labels (Figure 10(a)∼(d)), which indicates the signification of these two bands. The B11 and B12 are short-wave infrared bands and they are sensitive to high temperature targets to some extent. The impervious surface has small specific heat capacity, which makes it easy to reach high temperature in summer, so short-wave infrared can effectively detect it. Both Red bars represent the actual importance of the feature, and blue bars represent the null importance with shuffled landcover class. The total number of shuffle is 80. The importance is quantified by the total gain across all splits the feature is used in. If the actual importance of the feature is more significant than null importance, this feature is useful, and greater difference indicates greater importance.
NDBI and VANUI have a strong correlation with impervious surface, so the comparison results are more obvious (Figure 10(e)∼(h)). Remarkably, for both two cities, we can clearly see the outline of the urban runoff through the VANUI images, which may be cause by the reflection of light on the water surface ( Figure 11). It indicates that the combination of Sentinel-2 and Luojia 1-01 images may have the potential to identify urban river networks. As for texture features, only Mean_1 and Mean_2 have outstanding advantages in impervious surface extraction for both two cities (Figure 12(a)∼(d)). This situation just proves that feature selection is indeed necessary. For both two cities, the importance of NDVI1, NDVI2, NDVI3 and NDVI4 are all significant, especially NDVI2 (Figure 12(e)∼(l)). It shows the importance of temporal features. The cause that NDVI2 is the most significant one may lie in the fact that vegetation and bare land can transform into each other in different seasons, which affects the correlation between classification results in summer and NDVI derived from other seasons.
The impervious surface with high albedo has similar spectral characteristics with bare land, which makes it impossible to distinguish between the two perfectly. The results indicate that the integration of multi-features can give full play to the advantages of various features in the identification of impervious surface, and significantly improve the classification results of impervious surface and bare land. However, the confusion between impervious surface and non-impervious surface can't be eliminated completely due to the presence of mixed pixels. The RF classifier performs well in classification, but it fails to reduce the salt and pepper VOLUME 8, 2020 phenomenon. Therefore, we take spatial information into account by using PLR. It is important to note that the success of PLR is based on its optimal number of iterations. The number is different for different study areas, which need to be determined by plenty of experiments.
In generally, the combination of multi-features and RF is a successful attempt to obtain a high-precision impervious surface estimation. Additionally, feature selection with null importance and PLR also play important roles in the process of extracting impervious surface areas. Our method can not only accurately extract impervious surface but also serve as a solution to difficulties in distinguishing easily-confused land classes in pixel scale. The limitation of the method proposed in this paper is that the result relies on the manual selection of samples and it fails to achieve automatic classification. Therefore, when selecting samples, we need to be very careful to avoid the wrong classification results caused by improper selection of samples.

V. CONCLUSION
High-precision impervious surface information is of great scientific significance to the sustainable development of cities, and therefore attracts unprecedented attention. The similarity of different land classes' spectrum has always been a technical difficulty when using remote sensing data to extract impervious surface in per-pixel classification. In order to overcome this difficulty and obtain a high-precision impervious surface estimation, this paper proposed the method of combining multi-features and neighborhood-supported classification. Spectral features can enhance image quality and increase the contrast between impervious surface and other land cover classes. The texture features are based on the regular geometry of impervious surface and the temporal features are used to reduce the influence of tree shelter on impervious surface. PLR algorithm is introduced into RF model, which can reduce the salt and pepper phenomenon. Additionally, feature selection with null importance plays important roles in the improvement of impervious surface extraction results. The method proposed in this paper can effectively improve the impervious surface estimation and provide a solution to difficulties in distinguishing easily-confused land classes in pixel scale. In addition, it provides an important reference for the application of random forest in the classification of remote sensing images. However, the spatial resolution of Sentinel-2 images used in this research is 10 m, which is not high enough to void the occurrence of mixed pixels. In the future, higher resolution remote sensing data should be used to obtain more accurate extraction results of impervious surface. At present, deep learning method is attracting more and more attention in the impervious surface extraction, which is considered to be combined with our proposed method for further research. MIAO YANG is currently an Assistant Engineer with Henan Provincial Communications Planning and Design Institute Company Ltd. His current research interests include remote sensing, image processing, and research on the design of water transport and highway. VOLUME 8, 2020