Comparing WorldView-2 and PlanetScope Imagery to Mapping Housing Types Using GEOBIA

The mapping accuracy of housing types plays a vital role in urban planning and development. Choosing the right imagery for urban geospatial analysis matters in terms of spatial or textural resolution. Here we compare the effectiveness of different satellite imagery, namely WorldView-2 (2m resolution) and PlanetScope (3m resolution) to map housing types. The segmentation algorithm employed is SNIC (Simple Non-Iterative Clustering) while SVM (Support Vector Machine) algorithm is for classification. This study assessed the performance of these satellite platforms in capturing to extract spatial and spectral elements of each housing class and differentiating between urban villages (Kampung Kota), government-based housing, and private-based gated housing classes in the Tangerang area. WorldView-2, with its high spatial resolution, provides detailed information, allowing for precise delineation of housing boundaries and distinctive features, whereas Planetscope imagery offers better textural information for the segmentation stage. Despite the coarser details, the SVM classification algorithm achieved an overall accuracy of 65.00% using PlanetScope imagery. Comparative analysis revealed that WorldView-2 imagery outperformed PlanetScope imagery in terms of overall accuracy, with an overall accuracy of 65.52%. The higher spatial resolution of WorldView-2 enables better discrimination of housing types, resulting in more accurate classification. However, PlanetScope imagery provides valuable information, particularly for large-scale urban planning applications. The findings of this study contribute to the field of remote sensing and assist urban planners in making informed decisions regarding housing development and infrastructure planning based on available satellite imagery resources, both of which have their own advantages and disadvantages.


Introduction
Accurate mapping of housing types plays a pivotal role in the realm of urban planning and development.The demand for extract thematical data for urban analysis has been increasing for several decades due to the more complex social problems, namely slums, and spatial segregation affected by the gated community [1,2].The choice of suitable imagery for urban geospatial analysis holds significant implications, particularly regarding spatial and textural resolutions.For several years, pixel-based approach for land use classification, especially for urban, has been the standard until the emergence of 1264 (2023) 012007 IOP Publishing doi:10.1088/1755-1315/1264/1/012007 2 very high-resolution satellite imagery.Various types of satellite imagery along with its specific mission create even more diverse classification results.Some of the known high resolution satellite imageries are PlanetScope and WorldView series.Comparatively, WorldView-2 has better spatial resolution and clearer display on object, rather than PlanetScope which has generalized visualization of object so that hypothetically it could provide better textureal information.Previous research has been conducted to compare the results between PlanetScope and WorldView with the highest result coming from WorldView imagery [3,4].However, this research only asess the accuracy results of the land use mapping on general classes such as builtup, vegetation, bareland, etc., not the specific classes on residential based on socioeconomic aspect.Therefore, advanced research in this topic is needed to raise the issue about the ability of commercial high-resolution satellite to extract informations needed to be incorporated in urban GEOBIA analysis.This research aims to examine the ability of PlanetScope and WorldView-2 to extract spatial and spectral elements of each housing class and compare the accuracy results through the GEOBIA approach.
The core methodology employed for this comparative analysis involves the utilization of the Simple Non-Iterative Clustering (SNIC) algorithm for image segmentation, coupled with the Support Vector Machine (SVM) algorithm for subsequent classification.The study area is in Tangerang area, where the performance of these satellite platforms is rigorously evaluated in terms of their ability to capture intricate housing features and effectively differentiate between distinct urban housing categories, such as organic local housing or Kampung, government-based subsidized housing (Perumnas), and the middle-to-higher class such as Real Estate, cluster, or Gated Community.

Study Area
The primary focus of this study encompasses the Kelapa Dua and Curug Districts within Tangerang Regency, situated in the Province of Banten, Indonesia, as shown in the Figure 1 below.To the north lies an industrial zone intertwined with residential areas, which are administratively located within Tangerang City.This region is notable for the expansive Lippo Village Township, adjacent kampung settlements, and government-supported subsidized housing (Perumnas).Astronomically, the area study is in -6.218869N --6.255935N and 106.611208E -106.573988E.

WorldView-2
WorldView-2, a satellite designed for imaging and environmental monitoring by Maxar, an American aerospace company, was launched on October 8, 2009, and continues to operate proficiently [5].WorldView-2 introduces a collection of commercially available imagery features, including panchromatic imagery at a remarkably fine resolution of 0.46 meters, as well as eight-band multispectral imagery with a resolution of 1.84 meters.This resolution stands as one of the most superior spaceborne resolutions accessible in the current market context.Positioned in a sun-synchronous orbit at an elevation of 770 kilometers, the satellite completes an orbital cycle every 100 minutes, affording a revisit interval of up to 1.1 days.WorldView-2 distinguishes itself through the provision of high-resolution imagery encompassing eight distinct spectral bands.This encompasses the conventional four-color bands (blue, green, red, and near-infrared 1) and incorporates four additional spectral bands (coastal, yellow, red edge, and near-infrared 2), enabling advanced spectral analysis [6].The imagery used in this paper was recorded on 25 August 2020 in Tangerang Regency, Banten, Indonesia.
Object-Based Image Analysis (OBIA) is a technique that segments an image by grouping pixels together into vector objects, which are then classified using their shape, size, spatial and spectral properties [7,8].Using its very high spatial resolution, WorldView-2 imagery can be used in OBIA for urban land cover classification, change detection, and damage assessment and maintain its results accuracy [9,10].In a previous paper titled "The Use of WorldView-2 Satellite Data in Urban Tree Species Mapping by Object-Based Image Analysis Technique," researchers used WorldView-2 imagery and OBIA to map urban tree species in Serdang, Selangor, Malaysia [8].The study found that WorldView-2 imagery with good radiometric, spectral, and geometric resolution (up to eight bands) produced satisfactory results in urban tree species mapping.

PlanetScope
Planetscope is a constellation of approximately 130 satellites, each of which is a CubeSat 3U form factor (10 cm by 10 cm by 30 cm) [11].The complete constellation can image nearly the entire land surface of Earth every day, providing a daily collection capacity of 200 million km²/day [11].Planetscope images have a spatial resolution of approximately 3 meters per pixel and four spectral bands (RGB and NIR) [12].Planetscope satellite imagery finds extensive application in Object-Based Image Analysis (OBIA) encompassing land cover classification, change detection, and damage assessment, as established by various sources [13,14].OBIA employs a segmentation approach that groups pixels into vector objects, subsequently subjected to classification based on attributes such as shape, size, spatial relationships, and spectral characteristics [11].With an approximate resolution of 3 meters per pixel, Planetscope images consist of four spectral bands, including RGB and near-infrared (NIR).
These capabilities of Planetscope imagery extend to the extraction of urban features, as demonstrated in building inventory information extraction and the rapid identification of urban green spaces.Nevertheless, it's noteworthy that identifying urban features in Planetscope imagery encounters challenges arising from the relatively small and patchy nature of urban areas within cities [15].To address these challenges, researchers have combined machine learning techniques with Planetscope data to discern tree species in compact urban zones [15].They advocate for the integration of multitemporal eight-band Planetscope imagery in future studies, citing its potential to significantly enhance the accuracy of urban feature detection.This proposition emphasizes the evolving landscape of Planetscope's utilization and the importance of advanced technical considerations in maximizing its effectiveness in diverse urban analyses.

GEOBIA Classification
The classification method employed in this research is the Support Vector Machine (SVM) applied within the Google Earth Engine (GEE) platform for SNIC.The SVM algorithm categorizes the imagery using training input for each class and constructs a hyperplane that aligns with the closest sample, thus earning its name as this technique [16].Derived from Simple Linear Iterative Clustering (SLIC), SNIC is a clustering algorithm that boasts numerous benefits owing to its algorithmic simplicity, computational efficiency, and capability to establish boundaries based on key variables such as color, texture, and spatial proximity [17].Overall research methods are shown in the research workflow below (Figure 2): In the Google Earth Engine platform, the process employs specific settings: compactness of 0.8, seeds/grid distance of 45 in 'hex' form, and default values for the remaining parameters.The variables utilized in the classification encompass both spatial and textural information, namely GLCM Mean, GLCM Standard Deviation, GLCM Correlation, and GLCM Entropy.Subsequently, the obtained results are downloaded for further examination of their accuracy in QGIS.For training samples, each class were extracted with minimum 30 samples to get the optimal results.After that, the classification stage were conducted in Google Earth Engine with the previous extracted features as variables using SVM algorithm.The evaluation of GEOBIA (Geographic Object-Based Image Analysis) results adopts the Area-based Accuracy Assessment technique to assess the quality of each object variable, such as Overall Quality (OQ), User's Accuracy (UA), Producer's Accuracy (PA), and Overall Accuracy (OA) [18].The equations for conducting this accuracy assessment are provided below: In this study, ¬C ∩ R refers to the Reference (R) Area that does not overlap with the Classification (C) results, C ∩ ¬R denotes the Classification Area (C) that does not intersect with the Reference Area (R), and C ∩ R represents the intersection between the Classification Area (C) and the Reference Area (R).On the other hand, C ∪ R encompasses the entire area, combining both the Classification Area and the Reference Area.For the validation process, specific locations were selected, and any shape can be used.this paper, a circular shape was used to investigate the impact of outward region growing.The accuracy based then were started by doing intersection between vectorized classification results and reference vector before the area being calculated.Those area number then being calculated further in Microsoft Excel.

Result and Analysis
Good quality of imagery results in the good analysis.This also applies for the GEOBIA analysis in which object extraction from the imagery requires better visualization to make the algorithm understands the way human differentiate the object from surroundings.Based on the imagery visualization and overlayed segmentation results in Figure 3, it is notable that for the detailed housing block visualization, the algorithm could segment the object precisely into the building through the WorldView-2 rather than Planetscope.Planetscope could only show the housing block based on the homogeinity but with higher value of reflectance.This situation could be used to cluster the housing blocks with the same characteristic based on the coarser spectral value such as Low-High Density Building, but for the desired classification based on the housing greenness, pattern, density, and association, the finer resolution and quality such as WorldView-2 are the better one to use.
Based on comparison, WorldView-2 generates detailed housing block but often results in the over segmented object such as smaller unit (building, house with brighter color, roads).The results could be used to eliminate unwanted objects.On the other hand, Planetscope-2 generates poorer vision of housing blocks and under segmented object but could be used to generalize the housing blocks into binary class, such as low-density housing (LDB) and high-density housing (HDB).Two of these results are impacted to the classification results where Planetscope has less housing blocks detected than WorldView-2.Due to the detailed results of segmentation, WorldView-2 segmentation results appears to have misclassification between each other.The misclassification could be seen in southernpart of study area (Bojong Nangka Village) in Figure 4 where 'perumnas' class is misclassified as 'Elite' housing class.The result

Figure 1 .
Figure 1.Map of the Study Area

Figure 2 .
Figure 2. Workflow of This Research