Assessment of Machine Learning Methods for Urban Types Classification Using Integrated SAR and Optical Images in Nonthaburi, Thailand

Abstract

Urbanization and expansion in each city of emerging countries have become an essential function of Earth’s surface, with the majority of people migrating from rural to urban regions. The various urban category characteristics have emphasized the great importance of understanding and creating suitable land evaluations in the future. The overall objective of this study is to classify the urban zone utilizing building height which is estimated using Sentinel-1 synthetic aperture radar (SAR) and various satellite-based indexes of Sentinel-2A. The first objective of this research is to estimate the building height of the Sentinel-1 SAR in Nonthaburi, Thailand. A new indicator, vertical-vertical-horizontal polarization (VVH), which can provide a better performance, is produced from the dual-polarization information, vertical-vertical (VV), and vertical-horizontal (VH). Then, the building height model was developed using indicator VVH and the reference building height data. The root means square error (RMSE) between the estimated and reference height is 1.413 m. Then, the second objective is to classify three classes of urban types, which are composed of residential buildings, commercial buildings, and other buildings, including vegetation, waterbodies, car parks, and so on. Spectral indices such as normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and normalized difference built up the index (NDBI) are extracted from the Sentinel-2A data. To classify the urban types, a three-machine learning classifier, support vector machine (SVM), random forest (RF), and k-nearest neighbor (KNN) were developed. The classification uses randomly trained data from each 500 m focus study which are divided into a 100 × 100 m grid. Different models are examined using different variables, for example, classification using only building height and only spectral indices. The indices and estimated building height were used to classify the urban types. Not only the average of various satellite-based indices and building height of 100 × 100 m grid was used, but also the minimum, maximum, mean, and standard deviation were calculated from NDVI, NDWI, NDBI, and building height. There are a total of 16 variables used in the model. Eventually, the principal components analysis (PCA) was used to reduce the variables and get better performance of the models. SVM showed better accuracy than the other two, RF and KNN. The accuracies of SVM, RF, and KNN are 0.86, 0.75, and 0.76, respectively.

1. Introduction

Cities in developing countries face bigger urbanization problems than cities in industrialized countries. Additionally, urban growth and urbanization in each city of developing countries have become an important role on the Earth’s surface, with most people migrating from rural to urban areas. According to the United Nations Dataset, 55 percent of the world’s population lives in urban areas, a percentage that is expected to climb up to 68 percent by 2050, and almost 7 out of 10 people in the world will live in cities. Whereas Asia’s urbanization rate is approximately 50 percent, only 43 percent of the population in Africa lived in cities in 2018. Southeast Asia and other developing countries are currently among the world’s fastest-urbanizing regions; Thailand is one of the highest developing countries in terms of built-up areas [1], and Nonthaburi, Thailand is one of five vicinity provinces around Bangkok, Thailand that promote metropolitan expansion [2]. Rapid population development affects the infrastructures and transportation networks that have increased in urban areas. The strong impact of dense buildings and less vegetation in the urban zone can be influenced by the exponential growth in population, which can create a negative effect on municipal resources, urban sprawl, and disasters that are the global phenomena at present in several countries. Therefore, detailed information about the urban areas should be documented, and the number and density of buildings in urban areas should be identified and considered for sustainable development.

Urban dwellers constitute a limited proportion of all land cover groups, and the identification, and measurement of urban area extent, as well as population prediction, are critical factors in analyzing urban sprawl [3]. The urban area is being spread for infrastructures, and concerns about the degradation of its urban environment are growing as the urban region continues to grow rapidly. As a result, urban types identification is carried out for urbanization in an attempt to comprehend how urban forms evolve and to manage environmental sustainability, congestion, pollution, and natural disasters such as floods and earthquakes [4]. Site surveys, aerial imagery, and other traditional land survey techniques are overpriced and time-consuming. Therefore, the use of such remotely sensed data is becoming more common in distinguishing urban types for their high resolution and availability of information.

Many studies have used SAR to categorize land cover categories as urban or non-urban [3,5,6,7,8,9]. Researchers suggested urban land cover classification using SAR and optical data used in this research [10,11]. Landsat is used to classify ground cover types by separating them into sub-pixels [12]. The detection of the density of buildings and vegetated areas to classify urban types is unprecedented. [13] Demonstrated a connection between the interferometric coherence and the normalized difference vegetation index (NDVI). Convolutional neural networks to approximate the NDVI value with VV and VH polarization of SAR data was proposed in [14]. The accuracy of some optical remote sensing data in tropical regions could be low due to the presence of clouds. In contrast, synthetic aperture radar (SAR) is used to explore calm/shallow water with low backscattering because of the smooth surface, all-weather availability, free from cloud-cover conditions [5], and building inspection using backscattering with “double bounce” impact [3]. The use of SAR data in Thailand, which is situated in a tropical area with the coverage of clouds, is more relevant because of its advantages [15]. When optical remote sensing and SAR data are combined, the overall precision of urban classification is higher than when optical remote sensing or Sentinel-1 product is used alone. Machine learning for land use land cover (LULC) classification using Sentinel-1 and Sentinel-2A was investigated in [9,16]. Numerous methods for building height extraction from optical remote sensing images have been investigated [17,18]. The building height is potentially estimated using SAR because of the side-looking appearance and ability to provide images regardless of daytime and weather conditions. Detecting the heights of individual buildings by estimating the duration of layover from the single high-resolution TerraSAR-X image was proposed in [19,20]. The author [21,22] extracted the building height from the Sentinel-1 SAR images using VV and VH polarization. To fill this knowledge gap on urban height estimation and classifying different types of buildings in the study area, Nonthaburi Province, Thailand, we developed a building height model using the Sentinel-1 and Sentinel-2A data. The percentage of different buildings type using building height and greenspace (vegetation) is considered. Nonthaburi province was chosen as the study location because it is located in one of four metropolitan areas with a greater spread of many facilities, and it is close to the capital, the airport, government offices, department stores, and transportation services. Moreover, this province has slightly immigrated for habitation; however, the whole province is not covered, and eight focus areas are chosen to operate the processes of the research. SAR image was used to estimate and build a model for building height, the indices of NDVI, NDWI, and NDBI are extracted from Sentinel-2A optical data. The urban types were classified as (i) residential buildings, (ii) commercial buildings, and (iii) other buildings by using building height and satellite-based indices mentioned above.

2. Methods

2.1. Study Area

Nonthaburi province, Thailand, which is situated directly northwest of Bangkok on the Chao Phraya River, 13°51′32.7888″ N and 100°31′17.9472″ E (Figure 1). The province is divided into six districts, 52 subdistricts, and 433 villages. The population is 1.246 million in 2018, and the area of Nonthaburi province is 622 sq km which was registered as the second most populous city municipality in Thailand in the year 2019 (Department of Local Administration). As Nonthaburi is one of the metropolitan areas, there are very thick buildings such as residences, shopping malls, offices, and vegetation areas as well. Nonthaburi Province was registered as the 4th highest Human Achievement Index (HAI) in Thailand by the United Nations Development Program (UNDP) and the National Economic and Social Development Board (NESDB). The Nonthaburi research area is very wide, and the process is carried out in the eight study areas in the province. The eight target zones are 500 m and were created based on the possible combinations of different forms of the city, which are compact residential, industrial and high-rise buildings, as shown in Figure 2.

2.2. Data

There are three pieces of information, administration boundary, building blocks, and satellite images (see

Table 1. Details about dataset used.

Name	Acquisition Date	Resources
Sentinel-1 data	23 March 2020	European Space Agency
Sentinel-2A data	21 February 2020	European Space Agency
Building Blocks	2012	NOSTRA

), that are acquired from various sources. The first satellite, Sentinel-1, launched on 3 April 2014, and Sentinel-2A occurred on 23 June 2015, respectively. Both images are downloaded from the Copernicus open-access hub platform, which is freely available.

2.2.1. Building Blocks

Building block details containing the number of floors and the boundaries of buildings for each building in the Nonthaburi were collected from NOSTRA. Building block data define the character types such as offices, schools, shopping centers, and resident blocks which are grouped into commercial, residential, and others. These data are used to estimate and reference data for the building height. The data containing all polygons (about 460,000) of buildings in the Nonthaburi Province were surveyed in 2012, and the updated data of building blocks were acquired using google maps. Many new high-rise buildings were built after 2012 and fewer green areas were found. Moreover, the parking areas, roads, vegetation and water bodies did not include in NOSTRA dataset. Therefore, the buildings and polygon data after 2012 such as roads, water body and vegetation fields were checked with google maps and added

2.2.2. Sentinel-1 Images

The Sentinel-1 mission provides data from a dual-polarization C-band Synthetic Aperture Radar (SAR) instrument, which provides enhanced revisit frequency, coverage, timeliness, and reliability for operational services and applications requiring long time series. There are four bands in the Sentinel-1 product with 5 m spatial resolution, as shown in

Table 2. Band Information of Sentinel 1 IW GRD.

Band	Spatial Resolution (m)	Wavelength (GHz)
HH	5	5.405
HV	5	5.405
VV	5	5.405
VH	5	5.405

. This study operates on the Level-1 IW GRD and VH, VV polarization products that consist of focused SAR data that were detected, multi-looked, and projected to the ground range using an Earth ellipsoid model. In general, the building height is directly proportional to the backscatter coefficients of the Sentinel-1 SAR data; however, the relationship with VV and VH values can be varied. The backscatter coefficient of VH was used to estimate biomass and vegetation height in the earlier studies [22,23] and proposed that VH is a suitable proxy for categorizing vegetation’s vertical structure. The Sentinel-1 image that is used for this study was acquired on 23 March 2020.

2.2.3. Sentinel-2A Images

This research employed the Sentinel-2A multispectral optical sensor, which has a medium spatial resolution of 10 m of optical red, green, blue, and near-infrared bands. These multispectral data were downloaded through the use of the USGS website’s services. Sentinel-2A launched on 23 June 2015, and Sentinel-2B launched on 7 March 2017. The Sentinel-2A image that is used for this study was acquired on 21 February 2020. Many investigators prefer to employ images in their study because of its high spatial resolution, which is usually higher than that of other optical sensors. However, because these satellites are currently operational, time series analysis remains a restriction. Sentinel-2A products were in Sentinel-Standard Archive Format for Europe (SAFE) format, including image data in JPEG 2000 format, quality indicators, auxiliary data, and metadata.

2.3. Overall Methodology

The first objective of the research is to estimate the height of the building using a Sentinel-1 SAR image with the polarization of VV and VH. The second objective is the classification of urban types into three classes residential, commercial, and other areas, using 16 parameters; mean, minimum, maximum, and standard deviation of building height, NDVI, NDWI, and NDBI. Figure 3 describes that there are two images as inputs of the flow consisting of the Sentinel-1 and sentinel-2A.

In the data preprocessing section, SAR data were adjusted with noise reduction, calibration, speckle filtering, and terrain correction using desktop software SeNtinel Application Platform (SNAP) of the European space agency. Additionally, the preprocessing of Sentinel-2A, multispectral data include atmospheric correction and surface reflectance computation using SCP plug-in in quantum geographic information system [24], and the indices such as NDVI, NDWI, and NDBI were calculated in ArcMap. For the reference data, the raw building blocks from 2012 were checked and reorganized the data by adding or deleting the building blocks and calculating building heights from the number of floors. After preprocessing, VV and VH polarization were extracted from Sentinel-1, and the mean, minimum, maximum, and standard deviation of VV and VH were used to estimate building height for each building. For the classification of urban types, both the estimated building height from Sentinel-1 and NDVI, NDWI and NDBI from Sentinel-2A were applied. There are 16 parameters of the three satellite-based indices: NDVI, NDWI, and NDBI, and building height including mean, maximum, minimum, and standard deviation of four main parameters. The urban categorization was carried out as a grid-based classification using 194 samples in Nonthaburi’s eight zones. The three groups of urban forms (Figure 4) were discovered using machine learning algorithms random forest, support vector machine, and k-nearest neighbor. Several compositions of parameters of machine learning models were investigated during classification to determine which component produced the greatest accuracy; the model was run using building height, NDBI, and NDVI. The overall accuracy was tested using a confusion matrix, and an urban type categorized map was created.

2.4. Data Preprocessing

2.4.1. Preprocessing of Sentinel-1 SAR

The use and interpretation of SAR imagery requires a series of complex pre-processing procedures, which run on ESA’s SNAP software. A satellite pre-processing of the SAR images consists of the orbit application, data calibration, speckle filtering, and terrain correction.

In step 1, the metadata of the orbit file, we calculate the precise orbit data to improve the geocoding and other SAR processing results. After some days, the exact orbits of satellites are calculated and available days to weeks after product generation. By using a precise orbit available in SNAP, the orbit state vectors can be downloaded and updated automatically in the product metadata for each SAR scene to provide information on precise satellite positions and velocities. Step 2, the noise removal is performed before calibration to minimize noise effects in the inter-sub-swath texture, in particular, normalizing the backscatter signal within the entire Sentinel-1 scene and reducing discontinuities between sub-swaths for scenes in multi-swath acquisition modes. The thermal noise removal operator for Sentinel-1 data in SNAP can also reproduce the noise signal to be removed during level-1 product generation and upgrade product annotations so that the correction can be re-application. For step 3, calibration is the process by which digital pixel values are converted to radiometric SAR backscatter. The information needed to carry out the calibration equation is included within the Sentinel-1 GRD product; in particular, it enables the simple conversion of image intensity values into the naught sigma values by a calibration vector annotated in the product. The scaling factor applied to the production of level 1 is reversed, and a constant offset and range-dependent increase are implemented, including the absolute calibration constant. In step 4, for speckle reduction, the filter is applied to the data to eliminate the number of speckles at the cost of blurred features or decreased resolution characteristics. Both sigma 0 VV and VH were used as source bands. The last step 5 terrain correction, which can geocode the image by correcting SAR geometric distortions using a digital elevation model (DEM) and producing a map-projected product [25]. DEM SRTM with the 3-s resolution was used for this step. Resampled of the images was done using bilinear interpolation.

2.4.2. Sentinel-2A Optical Image Preprocessing

The image needed for this study is multispectral data containing energy from the sun to the earth’s surface. Sentinel-2A is a Level-1C product that is already geometrically adjusted and has top-of-atmosphere (TOA) reflectance. As a result, the next step is to utilize the SCP plugin in QGIS to execute Dark object subtraction (DOS) for atmospheric correction and surface reflectance since it is a useful tool for preprocessing various satellite images [26].

The first preprocessing step is atmospheric correction using dark object subtraction, which modifies atmospheric particles from satellite imagery of Top of Atmosphere reflectance to surface reflectance by removing 1% of the reflectance of the darkest object in a scene obtained from a shadow or absorptive object. In addition, reflectance values are determined using established Sentinel-2 sensor specifications. The output of atmospheric correction is an image that can determine surface reflectance, with each pixel containing a portion of solar energy and object reflection [26].

A semi-auto classification plugin is a tool that QGIS users could install with QGIS software [24]. Users can utilize SCP to perform satellite imagery activities such as downloading, preprocessing, classification, and postprocessing, as well as accuracy evaluation. As a result, SCP offers an option for assisting with atmospheric correction and reflectance computation for a Sentinel-2A image. As an outcome, employing the image band set and information from the image product, the SCP tool is used to perform atmospheric correction and reflectance calculation. For each band, the preprocessing produces atmospherically adjusted, and surface reflectance computed raster files. These images will be used to calculate spectral indices in further study.

2.5. Machine Learning Classification of Urban Types

Machine learning algorithms such as random forest (RF), support vector machine (SVM), and artificial neural network (ANN) has been used to classify and map the land use in the study area of Hangzhou City, China [27]. The data are derived from remote-sensing data, social media data, and other open internet data. The accuracies are varied regarding the urban land use level. In this study, RF and SVM models performed better than ANN in the urban land-use classification. [28] Proposed to separate the built-up regions and bare land using NDTI, the soil-adjusted vegetation index (SAVI), NDBI, the red-edge-based normalized vegetation index (NDVIre) and modified normalized difference water index (MNDWI) using Sentinel-2A level-1C in complex urban areas.

In this study, three classifiers, including SVM, RF, and K-nearest neighbors (KNN), were used. However, these three classifiers have their strong points and shortcomings; these models were chosen not only because of their efficacy in prior research but also because they are ideally suited for this type of categorization. The input dataset is loaded as CSV with the 194 samples, which include the mean, minimum, maximum, and standard deviation height of buildings which are estimated from the Sentinel-1 SAR, and the mean, minimum, maximum, and standard deviation value of NDVI, NDWI, NDBI from the eight study zones in the Nonthaburi Province. Among 194 samples, 135 samples are for training data, and 59 samples are used for testing datasets. The training data were sorted up randomly from each 500 m focus area which was divided into a 100 × 100 m grid for classification. Different spectral indices NDVI, NDWI, and NDBI are calculated from the Sentinel-2 optical data to input for the classification of the urban types by using ArcGIS, a raster calculator tool. Satellite-based indices are the spectral characteristics of reflection values that represent the chemical features of a substance on the earth’s surface. There are many indices that many scholars have chosen to determine differences in the similarity of building structures. Some indices offer a unique character of urban-type reflectance. This research identifies indices related to vegetation and building features that can be appropriately estimated from several indices. Many studies have used airborne hyperspectral and very high resolution satellite images to produce more spectral variables. This study, on the other hand, selects indices specified in the Sentinel Hub indices database that are related to vegetation, water body, and buildings. A random sample of 70 percent of data for the training classifier and 30 percent of the data for testing the accuracy of the results were used. To classify the urban zone, the focus zones were identified to train the model, including the three classes as residential, commercial, and other areas, which include vegetation, road, and water body.

Data standardization is applied for transforming numeric columns to a standard scale with Equation (1). PCA and standardization are methods for scaling and reducing the number of features. Initially, there were four main parameters (building height, three satellite-based indices) and the mean, minimum, maximum, and standard deviation of each parameter. The number of components utilized for classification is determined by the explained variance, which is estimated to be between 95 percent. This is done to minimize the number of parameters and the variance of parameters in the classification of learning. In this study, 16 parameters are reduced using PCA into 7, 6, and 5. There are ten experiments to distinguish urban types using all 16 parameters, only building heights, only three satellite-based indices, height with NDVI, height with NDWI, height with NDBI, and height with NDVI, NDBI.

$$Xsts=\left(X-mean\left(X\right)\right)/\left(sd\left(X\right)\right)$$

(1)

where X = Input column to transform, Xsts = Transformed column.

The model was developed in Python using the scikit-learn module. Model parameters were adopted to tune the appropriate dataset in this study. SVM is a classification learning model which draws lines of separation and assigns classifications to samples. Second, random forest is the decision tree-structured classification approach. It develops decision trees at random and learns via the dataset for all trees; the categorization result is voted on to assign a result class. Third, the KNN algorithm identifies unknown data points by determining the most frequent class among the k-closest samples. To summarize, urban types are classified into three classes; the dataset is divided into training and testing, and three learning algorithms will be used to train them. Following training, all models will be tested for accuracy using a confusion matrix.

3. Result and Discussion

3.1. Relationship of Building Heights and Backscatter Values

Building heights are usually associated with backscatter coefficients of sigma 0 (σ0) from Sentinel-1 data. Figure 5 shows that when the height is low, the sigma value of VV and VH is low as well. Although the VV and VH σ0 values show negative, the value of VV is higher than VH in this region. Cross-polarization values (HV or VH) are comparatively low, probably due to the relatively low incidence angle. In principle, VV polarization is more related to vegetation. The means of polarization values were extracted from the grid-based zone of the study area and made into a group with percentile with a step of 5% to 100%.

3.2. Developing of VVH Indicator

After preprocessing Sentinel-1 GRD data, a new indicator VVH was developed to build a height model using Equation (2) [21]. The VVH was calculated from VV and VH values which are percentile values; the output value changed according to the γ value. Among them γ = 1 produced the best coefficient of determination (R²) value (0.8561). When γ values were varied from 1 to 10, the R² of VVH drops from 0.856 to 0.1658, obviously because γ is a parameter to characterize the relative impact of VH, which has negative values and is lower than VV values to the derived VVH. The VVH performs better and has a reduced uncertainty range of predicted building heights.

$$\mathrm{VVH}={\mathrm{VV}\ast \mathsf{\gamma }}^{\mathrm{VH}}$$

(2)

3.3. Building Height Model

The building height model was developed by utilizing reference height data and VVH in the eight research zones of Nonthaburi at a scale of 500 m. It should be emphasized that the term “building height” in this study refers to the average height inside the 100 m grid, including buildings and non-buildings such as roadways and parking lots. To achieve a balance of model performance and spatial details of derived building heights, 100 m is the aggregation resolution with Equation (3) at γ = 1 of the VVH value in this research. The final parameters of a, b, and c are 0.2799, 1, and 5.727, respectively. The building height model using Sentinel-1 image and reference building height data performed well with a root mean square error of 1.413 m and 0.8557 of R² value (Figure 6).

$$lnH=a\ast {\mathrm{VVH}}^b+c$$

(3)

where H is the log-transformed building height; a, b, and c are three parameters in the building height model.

In the study zones, a few buildings are very high, about 186 m, and these kinds of buildings were removed because the unbalanced data created low accuracy. There are uncertainties in the calculated building height data from the criteria of building height, backscatter coefficients from Sentinel-1 data, and the suggested building height model. First, the term “building height” in the result refers to the mean height calculated from Sentinel-1 data inside the 500 m grid, which includes both buildings and non-buildings. Second, although the Sentinel-1 data were used in the year 2020, the reference height data were derived in the different years of 2012. So that this period gap could be the source of randomness. Overall, results from the model showed a good agreement with the reference building height.

3.4. Urban Types Classification

The classification of urban types were made in a variety of methods, including factors such as only satellite-based indices, building height, and both satellite-based indices and height. Three models were employed for classification and comparison to choose the best accurate training for urban types mapping. For splitting of training and testing data, 59 of 194 grid samples were chosen as validation data, and there were 135 grid samples as training data (30% test size).

Then, to minimize the number of features and processing time, PCA and standardization were used. The number of components in the PCA module was 0.95, which described 95% of the variation explained by all characteristics. Before employing PCA, feature scaling was applied to standardize. As a consequence, for all urban types, the number of components that represented 95 percent of the explained variation of features was 4, 5, 6, and 7 for different scenarios, respectively. There are ten trials, as shown in Table 6, to classify the urban types using all 16 parameters, only building heights, only three spectral indices, Height with NDVI, Height with NDWI, Height with NDBI, and Height with NDVI, NDBI.

Classification models were evaluated using accuracy, precision, and recall values for each model to choose the best model based on precision and recall for creating urban types maps and concluding the model to map the final urban map. The precision value is the proportion of classes that corrected the prediction. Furthermore, the recall value displays the number of corrected categories in one class in relation to the overall number of projected urban types. For the SVM classifier, the value for gamma is selected as 0.1, 0.01, and 0.001 for ‘rbf, linear’ kernel. The ‘C’ value is the regularization parameter tested as 100, 200, and 300, and the algorithm will choose the best suitable scores. For the RF classifier, four parameters were defined to optimize the model: the number of features to take into account when looking for the optimum split (max_features), the number of estimators (the required number of spanning trees) (n_estimators), the number of depths, and the number of jobs. For the KNN classifier, the number of neighbors was defined as 3, 5, 11, 19, and for the weight: uniform and distance were applied.

Urban types classification with 16 parameters consisting of the mean, minimum, maximum, and standard deviation of both building height and the three satellite-based indices were operated. The models used were support vector machine, random forest, and k-nearest neighbor. The classifications were conducted to generate the results of classifications using different models to compare the accuracy between each model shown in

Table 3. Classification result using building height with satellite-based Indices.

Models	Residential		Commercial		Other buildings		Weighted Avg		Overall Accuracy
	Precision	Recall	Precision	Recall	Precision	Recall	Precision	Recall
RF	0.74	1.00	0.8	0.33	1.00	0.12	0.78	0.75	0.75
SVM	0.90	0.95	0.8	0.67	0.75	0.75	0.86	0.86	0.86
KNN	0.75	0.97	1.00	0.5	0.5	0.12	0.76	0.76	0.76

. The three classes of residential, commercial, and other buildings were identified by all models.

Table 3 demonstrated that, when utilizing three indices, including mean, minimum, maximum, and standard deviation, the SVM model generated the greatest accuracy, precision, recall in predicting urban types when compared to others. While there is no statistically significant difference in accuracy between KNN and RF, the numbers demonstrated the precisions and recall for each class. Despite having the highest recall in residential buildings, other buildings have the lowest recall in both KNN and RF models. The other buildings have the least precision value in the KNN model.

Classification from the mean and maximum of the main four parameters provided better accuracy than classifying with mean values only. The residential building was the significant class for the study area, and the overall accuracy of all types was not different ranging from 0.68 to 0.71 for all models (

Table 4. Classification result using mean and maximum building height and satellite-based Indices.

Models	Residential		Commercial		Other buildings		Weighted Avg		Overall Accuracy
	Precision	Recall	Precision	Recall	Precision	Recall	Precision	Recall
RF	0.73	0.90	0.50	0.33	0.67	0.25	0.67	0.69	0.69
SVM	0.80	0.85	0.62	0.42	0.44	0.50	0.71	0.71	0.71
KNN	0.76	0.87	0.56	0.42	0.20	0.12	0.64	0.68	0.68

).

Table 4 demonstrated that the SVM model generated the higher overall accuracy and precision in predicting urban types when compared to others. While there is no statistically significant difference in accuracy between KNN and RF. RF had the highest recall in residential buildings but the lowest for commercial and other buildings. The other buildings have the least precision value in the KNN model.

Table 5. Classification using mean, maximum, and standard deviation of building height and satellite-based Indices.

Models	Residential		Commercial		Other Buildings		Weighted Avg		Overall Accuracy
	Precision	Recall	Precision	Recall	Precision	Recall	Precision	Recall
RF	0.70	0.97	1.00	0.17	0.67	0.25	0.76	0.71	0.71
SVM	0.67	0.79	0.22	0.17	0.25	0.12	0.52	0.58	0.58
KNN	0.67	0.97	1.00	0.08	0.00	0.00	0.64	0.66	0.66

, the mean, maximum, and standard deviation of building height and satellite-based indices yielded medium accuracy; however, the SVM classifier dropped its performance obviously as comparing the two cases above (Table 3 and Table 4). Moreover, despite the commercial buildings being classified correctly with precision value 1, the recall values were very low in RF and SVM models.

3.5. Urban Types Mapping

Various approaches and characteristics influenced the accuracy of urban-type categorization, according to the classification results. However, some models created their goodness and each model’s restrictions. Because certain classes reflected an unequal number of urban types for each zone, sample selection for training and validation may affect the accuracy of all models.

Additionally, this study tried to map dominant urban types by relying on classification results. The principle to selecting a suitable model is that the model should have the capability to classify three types of urban. To map urban types in Nonthaburi, the overall accuracy of classification using all parameters, including building heights, the three satellite-based indices, with the SVM classifier produced a better performance than others, and it was selected as the result of this study to make a map of urban types. Figure 7 shows the result of the urban types classification map in Nonthaburi, which was produced by the prediction of the support vector machine model with the combination of building heights, NDVI, NDWI, and NDBI. From this map, the percentage of each class was calculated to understand which type and how much the study area was occupied. The percentage of residential, commercial, and other buildings is 47.9%, 7.38%, and 44.7%, respectively. While residential and vegetation take place in almost the whole of Nonthaburi, very few buildings for commercial use can be seen on the map. As mentioned above, several grid zones are used as accommodation, grocery shops, and small restaurants. So those buildings could not be classified well because of the unbalanced sample data and building heights.

4. Discussion

4.1. Building Height Estimation

For the first objective, the estimated building heights of the study zones within 500 square meters from Sentinel-1 data have RMSE 1.413 m, about half of one floor. In the research of [21], the building height was estimated using LiDAR with ICESat heights for reference height in seven cities in the US. However, LiDAR data are more powerful and accurate in estimating building height; data costs are considerable and cannot be accessed everywhere. There are some challenges that the building heights which are low and high-rise buildings, could be similar because of the effect of backscatter coefficients on the ground layouts and materials of buildings. [29] Described that using only separated polarization from Radarsat 2 would not produce good accuracy in estimating the urban building height. The experimental results showed that mean of four polarized combinations produces much fewer errors than the single-polarization method for building height extraction. Therefore, building height extraction with SAR polarization, especially VV and VH, provided better accuracy than using other datasets. The suggested building height model could well be utilized in some areas of Pathum Thani where there was no ground truth data, and it is believed that the model can be used in any location on a worldwide scale.

4.2. Urban Types Classification

The urban types classification was carried out by using machine learning algorithms with random forest, support vector machine, and k-nearest neighbor classifier. In this research, the classification results were produced and compared by combining ten different scenarios (

Table 6. Parameters used and accuracy for classification on different cases. (Cases highlighted in grey are shown in results section).

Cases	Building Height	NDVI	NDWI	NDBI	RF	SVM	KNN
Case 1	mean, max, min, std	mean, max, min, std	mean, max, min, std	mean, max, min, std	0.75	0.86	0.76
Case 2	mean	mean	mean	mean	0.63	0.63	0.71
Case 3	mean, max	mean, max	mean, max	mean, max	0.69	0.71	0.68
Case 4	mean, max, std	mean, max, std	mean, max, std	mean, max, std	0.71	0.58	0.66
Case 5	mean, max, min, std				0.73	0.68	0.66
Case 6		mean, max, min, std	mean, max, min, std	mean, max, min, std	0.72	0.76	0.69
Case 7	mean, max, min, std	mean, max, min, std			0.69	0.77	0.71
Case 8	mean, max, min, std		mean, max, min, std		0.71	0.80	0.68
Case 9	mean, max, min, std			mean, max, min, std	0.83	0.75	0.73
Case 10	mean, max, min, std	mean, max, min, std		mean, max, min, std	0.73	0.81	0.72

) to check which composition would provide good accuracy. Overall, the support vector machine worked very well compared to other models. The results may change according to the variables which are used for classification.

Comparing the results of cases 2, 3, 4 from Table 6, using the only average did not appropriate to classify the urban types. Even SVM performed very poorly in case 4, and the more variables were applied, in general, the better result was delivered in cases 3 and 4. When cases 5 and 6 were analyzed, the accuracies were particularly lower than other combinations of case 1. The commercial buildings cannot detect well, and other buildings were completely misclassified when only satellite-based indices were used with SVM and KNN. Case 9 and case 10 gave higher accuracy than case 7 and case 8. This mean, using building height with NDVI, NDWI could not be classified well, similarly to classifying with building height with NDBI.

The building height used for classification was extracted from the Sentinel-1 SAR, objective 1. According to this result, using only Sentinel-1 SAR and Sentinel-2A alone could not yield good accuracy. [9] Also declared that using Sentinel-1 data (56.01%) only gave the worst result and the best result from the integration of Sentinel-1 and Sentinel-2A data (91.07%) for urban building classification and LULC mapping in the Eastern Brazilian Amazon. Furthermore, the building height with indices was combined separately (case 7, 8, 9); among three combinations, the building heights with NDBI provided good accuracy because several buildings were in the study area. Even the overall accuracy of combining building heights with NDVI and NDBI was lower than that of building heights with NDBI. Overall, the building heights and all satellite-based indices have the highest accuracy among all the combinations. All the variables are considered as the mean, minimum, maximum, and standard deviation of each parameter (case 1).

The mean, maximum, minimum, and standard deviation of satellite-based indices were used to classify the urban types. These values were calculated according to their respective study zones. Using maximum, mean, and standard deviation yielded higher accuracy than using them with the minimum since the minimum values of height indices were zero and minus values, respectively, in all study zone. When classifying building heights with NDVI, NDBI yielded a better result than using building height with NDWI, because the overall value, especially the minimum value, was very low. When the values were poor, the classifiers could not detect them properly.

To summarize the discussion section, the commercial building was mixed with the residential building because some buildings are used for both purposes. Moreover, open filed such as car parking can also be classified as commercial buildings. However, the classification of urban types can be done better with the integration of Sentinel-1 (building heights) and Sentinel-2A (satellite-based indices) than by using the satellite data points individually.

5. Conclusions

This study focused on two objectives; to estimate the building height and classify the urban types using Sentinel-1 and Sentinel-2A products in the Nonthaburi. Following the objectives, the building heights were estimated using VV and VH polarization in the 100 × 100 m grid samples. All the building heights produced mean height within the 500 m resolution, including buildings and non-building such as streets and parking lots. The new indicator VVH was produced by applying VV and VH to get better output and mitigate the uncertainties of the range of estimated building heights. Then, the building height model was developed at different quantile levels ranging from 5% to 100% at a step of 5%. The study had a few limitations, which would be addressed in future work. The building height in the current research was calculated based on the number of floors which can be calculated using lidar data hence making the input data more accurate. Building types with mixed characteristics were not considered and were marked as residential or commercial based on experts opinions.

Overall, the model’s results positively agree with the NOSTRA, 2012 reference building height. Furthermore, the model was evaluated to predict height in Nonthaburi and Pathum Thani, and the RMSE between the estimated and reference heights is fair. Building height information is beneficial for evaluating urban climate modeling and assessing changes in population density, energy usage, greenhouse gas emissions, and so on. For the second objective, building heights and satellite-based indices such as NDVI, NDWI, and NDBI were considered to classify urban types using machine learning algorithms such as random forest, support vector machine, and k-nearest neighbor. All classification processes were done in the 100 × 100 m, 194 sample grid zone. There were some difficulties that the reference data were from 2012, but the Sentinel-2A image was from the year 2020, so the polygons of buildings were reviewed and updated in the ArcMap.

There were 16 variables for the classes as residential, commercial, and other buildings, which include streets, parks, and vegetation. The classification was produced by applying ten scenarios. Among them, using all variables in SVM gave the best accuracy following the RF using heights and NDBI. Although the residential building is the most dominant type and the precision, recall, and accuracy of classification were good, the results of commercial and other buildings were poor and misclassified. KNN has the lowest accuracy among the models. The classification using only building heights and only satellite-based indices could not detect the urban types and produce good accuracy. Furthermore, the urban types classified map was created from the SVM model, which provided the best classification result. From the map, the percentage of the urban types was calculated to be able to help the urban growth assessment, urban types distribution, and several environmental managements.