Performance of Support Vector Machine for Classifying Land Cover in Optical Satellite Images: A Case Study in Delaware River Port Area

Abstract

The availability of high-resolution satellite images provides precise information without direct observation of the research target. Korea Multi-Purpose Satellite (KOMPSAT), also known as the Arirang satellite, has been developed and utilized for earth observation. The machine learning model was continuously proven as a good classifier in classifying remotely sensed images. This study aimed to compare the performance of the support vector machine (SVM) model in classifying the land cover of the Delaware River port area on high and medium-resolution images. Three optical images, which are KOMPSAT-2, KOMPSAT-3A, and Sentinel-2B, were classified into six land cover classes, including water, road, vegetation, building, vacant, and shadow. The KOMPSAT images are provided by Korea Aerospace Research Institute (KARI), and the Sentinel-2B image was provided by the European Space Agency (ESA). The training samples were manually digitized for each land cover class and considered the reference image. The predicted images were compared to the actual data to obtain the accuracy assessment using a confusion matrix analysis. In addition, the time-consuming training and classifying were recorded to evaluate the model performance. The results showed that the KOMPSAT-3A image has the highest overall accuracy and followed by KOMPSAT-2 and Sentinel-2B results. On the contrary, the model took a long time to classify the higher-resolution image compared to the lower resolution. For that reason, we can conclude that the SVM model performed better in the higher resolution image with the consequence of the longer time-consuming training and classifying data. Thus, this finding might provide consideration for related researchers when selecting satellite imagery for effective and accurate image classification.

1. Introduction

Remotely sensed image, produced by optical or synthetic aperture radar satellites, has played an importantrole in earth observation.Remote sensing has become an essential technique in providing repeated information over large and inaccessible areas (Foody and Mathur, 2004; Kim, 2016). Satellite sensors are continuously developing and providing a higher-resolution image. The main advantages of the high-resolution image are the reduction of mixed pixels and offer more detailed spatial information (Lu and Weng, 2005).The high-resolution images often used included WorldViewII, GeoEyeI, Pleiades, and Korea Multi-Purpose Satellite (KOMPSAT) images, also known as Arirang. KOMPSAT images provided by Korea AerospaceResearch Institute (KARI) have become one of the most advanced technology since providing 70 cm resolution KOMPSAT-3, KOMPSAT-5 radar satellite that observes the earth all day and regardless of weather conditions, and 55 cm optical and infrared (IR)KOMPSAT-A.KOMPSATimageshavecontributed to various fields, especially in classification, such as mapping ofland cover(Acharya et al., 2016) and urban areas (Youn and Jeong, 2020), digital elevation model (DEM) generation (Rhee et al., 2011), and assessing wildfire impact (Nur et al., 2020).

A land cover map is essential information that represents the relationship between humans and the surrounding physical environment. Identification of land cover has given benefits in land use planning (Foody and Mathur, 2004), biodiversity conservation (Gonzalez-Perez et al., 2022), and studying changesin land surface and environment (Srivastava et al., 2012; Zhou et al., 2008).Varioussatellite imageries have been used forland coveridentification, even optical imagery, radar data, and multi-source approach by combining both radar and optical data. The availability of high-resolution images, such as WorldViewII and KOMPSAT images, with a resolution of less than 5 meters, have enabled a more detailed classification of urban features, which are usually mixed with other features on lower resolution images, such as Landsat and Sentinel product (McGlinchy et al., 2019). This advantage should affect the classification results and improve the accuracy assessment.

The accurate classification result is determined by the image’sresolution and technique for classifying the images. Various machine learning models have been widely used in image classification using remotely sensed data and provided a high accuracy. Previous studies reported that a support vector machine (SVM) is a great model for multiclass and binary image classification (Mathur and Foody, 2008). Furthermore, the SVM model has become an alternative model which provides good accuracy in addition to the random forest and maximum likelihood models (Morgan et al., 2022). In this study, we aimed to compare the performance of the SVM model on the high-resolution KOMPSAT images and medium-resolution Sentinel imagesin classifying the land cover of the Delaware River Port area, Philadelphia, USA. This comparative study provides new insight into the performance of the SVM model in identifying multiclass high-resolution and medium-resolution optical images.Therefore,related researchers can refer to these resultsin selecting appropriate satellite imagery according to the case in obtaining effective and accurate classification results.

2. Study area

Delaware River islocated in the Catskill Mountains and flows along theAtlantic coast of the United States (Moore, 2021). The river is useful for people living because it providestap water as drinking waterfor New York City and New Jersey, where 7 million people live (Kauffman Jr., 2010). The export and import activities have contributed to handling and guaranteeing the needs of raw and manufactured goods for more than 300 years(Almaz andAltiok, 2012).At least 400 ports are located along the river, making it one of the busiest commercial maritime routes in the United States. Moreover, as the third dense population in the United States, where about 90 million people live within 500 miles of the Philadelphia port, the river and its port facilities have contributed to the nation’s economy (Altiok et al., 2012).Consequently, the area around the river will be densely populated, represented by many residences and terminals filled with stacks of cargo, bulk, container, and vehicle vessels. For that reason, observing the land cover of the Delaware River affecting many populations and logistics is required, and suitable satellite images are needed for such monitoring.

The study area is nearthe city of Philadelphia, which is located downstream of the wide Delaware River that handles a large amount of vessel traffic. Many supported logistics facilities, such as a CSX Rail Yard Philadelphia, a railway logisticsfacility, and a container yard, are located in the study area. In addition, refinery facilities and related equipment are widely arranged in the center of the study area. The eastern and southern parts of the research area are residential areas such as Gloucester City, where single houses with large yards, small yards, or multi-family houses are located. Therefore, this research area, including rivers and surrounding environments, issuitable forstudying land cover classification.

Fig. 1. The study area used to classify land cover based on the SVM model. (a) Location of Delaware River Port, Philadelphia and (b) the KOMPSAT-2 images show the land cover of Delaware River port area.

3. Materials and methods

1) Data

To investigate the performance of the SVM model for image classification, we used KOMPSAT-2 and KOMPSAT-3A images representing high-resolution and Sentinel-2 images as medium-resolution optical images.TheKOMPSATimage contains a panchromatic band and four multi-spectral bands, including Blue (B1), Green (B2), Red (B3), and Near-Infrared (NIR) (B4).TheKOMPSAT-2 image, acquired on 17February 2018, is a multi-spectral camera that consists of panchromatic and multi-spectral imagery with 1 m and 4 m resolution, respectively. Wavelengths of the KOMPSAT-2 image are 450–520 nm for Blue, 520–600 nm for Green, 630–690 nm forRed, 760–900 nm for the NIR band, and the panchromatic image has 500–900 nm of wavelength. The KOMPSAT-3A acquired on 27 February 2018 provides a 55 cm panchromatic image, a 2.2 m multi-spectral image, and top-level quality infrared sensor images. The wavelength of R, G, B, and NIR KOMPSAT-3A bands are the same as the KOMPSAT-2, while the panchromatic band has 450–900 nm of wavelength. For the comparison, the Copernicus Sentinel-2 image acquired on 5 February 2018 contains 13 bands, including four bands with 10 m resolution, six bands with 20 m resolution, and three bands with 60 m resolution, which was used in this comparative study.

In preprocessing of image classification, the natural color view was generated by combining R, G, and B bandsin the optical images.The provided images were cropped to focus on the area where all images overlap. A total of six classes of samples, including water, road, vegetation, building, vacant, and shadow, were generated and sequentially labeled by considering the land use map of Philadelphia (Philadelphia City Planning Commission, 2014) and New Jersey (New Jersey Department of Environmental ProtectionBureau of GIS, 2015).

The water class denoted by blue was defined as all water contained in the study area, including rivers, swimming pools, fishponds, and puddles. Road class represented the highways, residential streets, parking lots, railroads, and transportation-relates places. The location of grasses, forests, plants, and parks or open spaces was considered as vegetation.The building class represented residential, industrial building, container, and human-built materials. At the same time, vacant class denoted the location of abandoned or unused lands. The last class, indicated by black color, showed the existing shadow due to the angle of the satellite view.The samples were labeled according to the classes and divided into training and validation data. The training data were used to train the model in generating classified images, and the validation data was used to evaluate the results.

2) Methodology

The utilization of machine learning models is increasingly approved as an image classifier. One of the machine learning-based classifiers that are widely used is the support vector machine. SVM is a nonparametric supervised classifier recently used for image segmentation and provides a high classification accuracy when separating complex and nonlinear data (Morgan et al., 2022; Sheykhmousa et al., 2020). The SVM goal is to find an optimal hyperplane with a maximum distance between the support vectors that will clearly separate two classes and minimize misclassification (Foody and Mathur, 2004). Our SVM model adopted a nonlinear radial basis function (RBF) using a 2D grid approach with optimal parameter pairs, C and gamma (Baek and Jung, 2021).TheRBF kernel has been used widely in remote sensing and provided better accuracy for land cover classification than othertraditional methods (Talukdar et al., 2020). The C and gamma parameters were set automatically using a cross-validation technique in which a subset of training pixels is retained to validate them and select the optimal hyperparameter values (Hamilton et al., 2021). The average gamma values were 28.87, while the average C values were 29,508 for this case study.

The classification process was started by collecting the opticalimages,including high-resolution KOMPSAT-3A and KOMPSAT-2 and medium-resolution Sentinel-2B images. The images are composed of several bands, and R, G, and B bands were selected to use in this study. The RGB or natural color image was utilized by considering the advantages ofimage resolution and the characteristics of our study area. The high spatial resolution should provide detailed information even using natural color. The study area was the port area which is identical to the existence of human-made material and water. In this case, we preferred to use the blue band instead of the NIR band. The blue band usually identifies water and manufactured creations such as buildings, roads, and transportation facilities. Meanwhile, the NIRband has advantagesin identifying the boundary of land and water in a wetland area and can penetrate smoky and hazy (Eid et al., 2020). This study investigated the advantages of high-resolution images compared to other lower-resolution images in clear conditions, and the natural color was considered to classify the land cover of the Delaware port area.The overlapping area of three optical images was cropped and determined as a research area.

The selection of training samples and the representativeness of the samples to each class are essential factors for generating a proper image classification. Various strategies have been implemented to collect training samples, such as single pixel, seed, and polygon generation (Chen and Stow, 2002). The strategies can be selected by considering the purpose of the study, the availability ofreference data, the spatial resolution of the observed image, and the complexity of landscapes in the study area (Lu and Weng, 2007). Chen and Stow (2002) discovered that the single-pixel training samples preferred to train spectrally homogeneous classes, while the polygon sample was suitable for spatially heterogeneous classes. The polygon training approach is useful for extracting spectral and spatial information and reducing time-consuming for high-resolution images. For that reason, we considered creating training samples by generating some polygons with irregular shapes and labeling each polygon according to defined classes. We first divided an optical image into two parts, where 70% of the image was where the training data was generated, and the validation points were distributed in the remaining 30% of the image. Over 35 random polygon samples were manually digitized for each land cover class. The training samples were manually created and labeled by considering the reference data. The polygon training strategy causes the training dataset of each image to have a different number of pixels depending on the image resolution. The same training dataset was used to train the SVM model in the same scale of high-resolution KOMPSAT-3A and KOMPSAT-2 and medium-resolution Sentinel-2B images (Nur et al., 2020).Atotal of 100 stratified random points, different from the training data location, were distributed in each class to validate the classification result by comparing the predicted and the reference images (Nur et al., 2020; Syifa et al., 2020). The confusion matrix table summarized the performance of the SVM model in classifying the land cover of the Delaware River Port area.Summarymetrics, consisting of producer’s accuracy (PA), user’s accuracy (UC), overall accuracy (OA), and Kappa coefficient (KC), have been used to assess the model accuracy in optical image classification. The consumed time for training dan classifying processes in each optical image were recorded and summed up to conclude the time-consuming of the SVM model. The detailed workflow of this study is described in Fig. 2.

Fig. 2. Procedure to classify optical images and compare the classification results.

The accuracy of classification results was assessed based on the confusion matrix analysis by comparing the actual data with the predicted image and producing a confusion matrix table.The structure of the confusion matrix is shown in Table 1.

Table 1. The structure of a confusion matrix for evaluating the model performance (modified from Acharya et al., 2016)

Stratified random points and a confusion matrix table were generated and used to compute the multiclass summary metrics. PA described how many correct pixels of a class in a classified image compared with the actual data, while UA explained the ratio between the number of pixels correctly classified as a certain class and the total pixels of that class in the predicted map generated from the algorithm used.The OA value shows the pixels correctly labeled relative to the total pixels of the map (Maxwell et al., 2021). However, these accuracies do not consider the possibility of chance agreement between the datasets. Hence, a Kappa coefficient (KC) which evaluatesthe difference between a real agreement in the confusion matrix and the chance agreement, was used to ensure the accuracy assessment. The formulas of the summary accuracy metrics and Kappa coefficient are described in equations (1)–(4) (Acharya et al., 2016).

\(\begin{aligned}User’s\; accuracy=\frac{n_{i i}}{n_{i+}}\\\end{aligned}\)      (1)

\(\begin{aligned}Producer's\; accuracy=\frac{n_{i i}}{n_{+i}}\\\end{aligned}\)       (2)

\(\begin{aligned}Overall\; accuracy=\frac{\sum_{i=1}^{m} n_{i i}}{N}\\\end{aligned}\)       (3)

\(\begin{aligned}Kappa\;coefficient=\frac{N \sum_{i=1}^{m} n_{i i}-\sum_{i=1}^{m} n_{i+} n_{+i}}{\mathrm{~N}^{2}-\sum_{i=1}^{m} n_{i+} n_{+i}}\\\end{aligned}\)      (4)

where n_ii is an element in the i-th row and i-th column, n_i+ is the row total, n_+i is the column total, and N is the number of testing points.The accuracy of the classified image and the time required forthe entire classification process were considered to conclude the performance of the SVM model in this case study.

4. Results and discussion

This study compares the performance of the SVM model on the KOMPSAT-2, KOMPSAT-3A, and Sentinel-2B imagesfor land cover classification in the Delaware River port area. The optical images were classified into six classes: water, road, vegetation, building, vacant, and shadow, using the pixel-based-SVM model. The results show that the SVM model can identify six classes in each image with good performance. Some differences emerged between the classification results of high-resolution and medium-resolution optical images. The differences are shown in Fig. 3, and the KOMPSAT-3Aimage provided a better-classified image than KOMPSAT-2 and Sentinel 2B.

Fig. 3. Natural color images and classification results of (a) and (b) KOMPSAT-2, (c) and (d) KOMPSAT-3A, and (e) and (f) Sentinel-2B images using the SVM model. White box 1 and 2 in the southwestern and the northwestern area show the obvious difference between the three classification results.

We marked the most noticeable differences between the three classified images with the white boxesin Fig. 3. We then focused on the areas in the white boxes to clarify the differences and compared them with the ground truth data (Fig. 4). The official land use map of the study area and observation in google earth images were considered asthe ground truth data to evaluate the classification results.

Fig. 4. Comparison between classified images of KOMPSAT-2, KOMPSAT-3A, and Sentinel-2B using the SVM model and the reference images based on the official land use map and google earth images in (a) the southwestern and (b) the northwestern study area (Fig. 3).

The classification result of KOMPSAT-2 represented the misclassification of the road in the vegetation area compared to the reference image (Fig. 4(a)). The difference between building classin the actual land use map and classified images was caused by a different class definition forthe KOMPSAT-2 result. Meanwhile, Sentinel-2B classified image showed that the SVM model failed to identify water in the northwestern and southwestern study areas(Fig. 4(a) and (b)).The mixed pixel in the lower-resolution images was considered to be the reason that the image could not be adequately classified (Lu and Weng, 2005). On the contrary, the detailed information in each pixel in a high-resolution image assiststhe classifierin recognizing the feature of each class and producing better classification results.

The performance of the SVM model in each optical image was evaluated using a confusion matrix analysis. The classified image of KOMPSAT-3A produced the best result with an OA value of 92% and KC value of 89.83%, followed by the KOMPSAT-2 result with OA of 86% and KC 82.46% and Sentinel-2B classified image with OAand KC85% and 81.16%,respectively. SVM has performed well in classifying land coverfrom the KOMPSAT-3Aimage based on the overall accuracy and Kappa coefficient. The accuracy assessment of each classified image is shown in Tables 2–4.

Table 2. Confusion matrix of KOMPSAT-2 classified image and summary accuracy metrics

Table 3. Confusion matrix of KOMPSAT-3A classified image and summary accuracy metrics

Table 4. Confusion matrix of Sentinel-2B classified image and summary accuracy metrics

In KOMPSAT-2 images, SVM correctly classified all vacant,shadow, and most of the water and road.The model struggled to distinguish water and vegetation or shadow due to the similar pixel value inmossy orturbid water. Some building areas were misclassified as vacant or road, producing a PA value of 76.47% and a UA value of 92.85%. The false pixels were caused by the areas labeled as building abandoned or lost, and the classifier made some errors.

Similar to the KOMPSAT-2 result, the model was difficult to identify building correctly on the KOMPSAT-3A image. Building class was mostly misclassified and interpreted asroad and vacant as well. For that reason, the PA value of the building class was only 52.94%. As a result, the UA value of the vacant class has fallen to 50%.

The confusion matrix of the Sentinel-2B classified image denoted the same misclassification in the building class and produced a PA value of 52.94%. In addition, some areas misclassified water as a shadow in the northwestern study area (Fig. 4(b)). Moreover, the SVM model also struggled to distinguish vegetation with water or vacant due to insignificant pixel value differences on mossy water. This error caused the PA value to decrease to 81.25%.

In contrast with the accuracy results, the SVM model consumed a long time to train and classify the highest resolution KOMPSAT-3A image, which was 17 minutes 16 seconds, while the KOMPSAT-2 image consumed 11 minutes 51 seconds, and the Sentinel-2 image only took 4 minutes 14 seconds. Using the same training dataset, the results expressed that the model requires more time to process an image with a detailed feature and a higher number of pixels, such as those contained in a high-resolution image. The comparison of model performance determined based on the accuracy and time-consuming of training data and classifying images is shown in Fig. 5.

Fig. 5. The performance of SVM models in classifying KOMPSAT-2, KOMPSAT-3A, and Sentinel-2B images.

5. Conclusion

The SVM model was applied to identify land cover in the Delaware River port using optical satellite images. The model performances in classifying high-resolution and medium-resolution images were compared based on the accuracy and time-consuming in the classification process.The SVM model performed well in detecting six land cover classes on optical images represented by the confusion matrix analysis. SVM produced a good result with overall accuracy ranging from 85% to 92% and a Kappa coefficient from 81.16 % to 89.83% for both high and medium-resolution images. The best performance was shown when SVM classified the KOMPSAT-3A image, followed by KOMPSAT-2 and Sentinel-2B images.

The results were influenced by mixed pixel values of the images, causing some identified misclassifications on lower-resolution images. On the contrary, the complex environment and detailed features of high-resolution images that would generate a more varied intra-class on the image caused difficulty in identifying the classes correctly. The utilization of different bands can be investigated to overcome the limitations of this study in identifying land cover classes for further work. Regarding the time-consuming, the SVM model took longerto train and classify high-resolution imagesthan medium-resolution.A general conclusion can be drawn that the SVM model performed well in classifying land cover on higher-resolution images for the case study of the Delaware River port area with long time-consuming consequences.The resultmight provide information for related researchers when selecting satellite imagery for an effective and accurate image classification study.