Abstract
The local government's need for geospatial information becomes important when green open space is difficult to find in urban areas. The information is a land cover map, which can be obtained from the classification process of high-resolution satellite images, but it still has limited spectral information where object classification also involves spatial characteristics and textures to get high accuracy value (Kushardono, 2017). These characteristics are minimal in passive sensor data but are mostly found in active sensor data (radar/LiDAR). By knowing the potential of the two sensors (satellite imagery and LiDAR), several researchers have conducted similar studies, including Awrangjeb et al., (2013), Uzar and Yastikli (2013), and Gilani et al., (2015). In this study, there are 3 main data, namely Pleiades images, nDSM, and intensity images. nDSM is derived data from the LiDAR elevation value, while the intensity image is formed from the object's reflection value at the NIR wavelength, each of which has been corrected and then interpolated into raster data and classified. The classification process is carried out using OBIA method with the Assign Class algorithm through a segmentation process. The resulting classification scheme produces 12 subclasses (within 4 main classes) from each sensor data classification to form a composition dataset. This integration process produces 3 compositional datasets, namely dataset A (Pleiades image only), dataset B (Pleiades-intensity) and dataset C (Pleiades-nDSM). With GIS analysis, an accuracy test was carried out and the accuracy value of dataset A was 44.44% and dataset B and C both produced an accuracy value of 63.89%. The accuracy value is very low when referring to Indonesian Land Use/Land Cover Classification National's Standard, because the number of 36 sample points is disproportionate (<20%) to the total number of objects that reach thousands.
Keywords: OBIA, land cover, Pleiades, LiDAR, nDSM, intensity, integration
1. Introduction
A land cover map can be obtained from the results of land classification using various methods, one of which is remote sensing techniques (Sampurno & Thoriq, 2016; Aristalindra et al, 2020; Prayogo & Basith, 2020). In a study conducted by Alonso and Malpica (2010), a study was conducted on the synergistic combination of sensors between LiDAR and multispectral imagery in mapping the distribution of buildings. From the research that has been done, the results obtained are that the Overall Accuracy classification using satellite imagery data and LiDAR is 97.12% while the Overall Accuracy classification using satellite imagery data without LiDAR data is 80.34%.
LiDAR (Light Detection And Ranging) is a remote sensing technology with active sensors in the form of lasers that are integrated with flying vehicles/aircraft or Unmaned Aerial Vehicles (UAVs) which are capable of scanning a large area (Istarno, 2016). While active sensor vehicles (radar/LiDAR) have advantages in terms of spatial (higher geometric accuracy with very minimum relief displacement), satellites have advantages in spectral terms (wider range of spectral channels/bands in identifying characteristics of objects) (Kushardono, 1997; Santosa, 2016; Aristalindra et al, 2020).
Theoretically also according to Kushardono (1997), the difference between the two types of sensors between passive sensors and active sensors provides different output but can provide the same interpretation of certain objects based on a reference or rule. Of course this can enrich the information or characteristics of an object so that it is assumed to increase the accuracy of the classification results. This study aims to explore the potential of integrated satellite imagery with LiDAR data in classifying land cover objects in urban areas that have heterogeneous land cover objects. The research results are expected to provide answers to the hypotheses that have been made and have reliability that can be measured through accuracy tests.
2. Materials and Methods
2.1 Data and Location Study
This research activity took place in Makassar City in 2013, precisely in Tamalanrea Indah Village. This location was chosen based on the researcher's consideration of the availability of data from two data provider agencies (PT. Asi Pudjiastuti Geosurvey and LAPAN Remote Sensing Information & Data Center) at adjacent locations and times. Some of the data used in this study are as follows:
1. LiDAR point cloud data (Location: Makassar City, Time of acquisition: 1 – 22 June 2012, Average point spacing: 0.47 meters, Data format: *las, Data source: PT. Asi Pudjiastuti Geosurvey)
2. Trajectory recording of LiDAR aircraft (Location: Makassar City, Time of acquisition: 1 – 22 June 2012, Flying height: 776 – 1038 meters above sea level, Data format: *trj, Data source: PT. Asi Pudjiastuti Geosurvey)
3. Pleiades satellite image data (Location: Makassar City, Time of acquisition: August 18, 2013, Spatial resolution: 0.5 meters, Data format: *TIFF, Data source: LAPAN Remote Sensing Information & Data Center)
2.2 Methodology
2.2.1 Point Cloud Data Correction & Filtering
2.2.2 normalized Digital Surface Model (nDSM)
2.2.3 Intensity Image
2.2.4 Pleiades Satellite Image
2.2.5 Object-Based Image Analysis (OBIA) Classification
In the segmentation process, the values of the input parameters (scale, shape, and compactness) are determined by trial and error, that is, values are entered repeatedly to obtain a visually proportional segment shape. The input value also pays attention to the appearance of the segment/object in the image.
Each segmented sensor data is then classified one by one. The classification process used is by using the Assign Class algorithm, which is a simple classification algorithm in which each object is classified based on the criteria value with a threshold condition into a certain class. The Assign Class algorithm is unsupervised, where object classification does not use the training area as a reference for classification. Objects in the form of segments will be classified if the criteria values on the object meet the given threshold. These criteria are in the form of spectral criteria (layer value), spatial criteria (shape and area) and texture criteria.
2.2.6 Data Integration
After the results of the three classifications of sensor data are exported, it can be overlaid with the identify method to carry out a matching up analysis of object classes to form compositional datasets. The analysis is carried out by calculating the percentage of the number of objects that have the same matching class of all objects. The technique used to find entry objects that meet these conditions is by making SQL queries. The number of objects selected with the query is then calculated based on the total number of identified objects to get the percentage value of matched up object class, while the composition dataset that is formed can be processed into the accuracy testing stage. The calculation results that have been obtained can then be analyzed further.
2.2.7 Accuracy Test
The accuracy test process was carried out on the classification results of dataset A (Pleiades images), dataset B (Pleiades images with intensity images) and dataset C (Pleiades images with nDSM). The accuracy test is measured from the matched up of the classified land cover class with the ground truth of the Google Earth temporal image which is visually interpreted. The selected ground truth coordinate points are displayed in the form of ground check points, which are taken randomly and evenly throughout the land cover of the case study area with a total of 36 points.
3. Results and Discussions
3.1 Segmentation Results
3.1.1 Pleiades Image
3.1.2 nDSM
3.1.3 Intensity Image
3.2 Classification Results
3.2.1 Exported Result
3.2.2 Query Analysis & Overlay Result
The overlay identify process focuses on overlaying spatial data (classified objects) from two different sensor data while maintaining the attribute data of each sensor data. The overlay identify result in the number of objects from each dataset being multiplied, which is mainly caused by the number and size of objects in each sensor data during the segmentation process. In order to be able to carry out a match up analysis for object classes, information on the number of objects that meet the conditions where the KELAS_1 attribute is the same as the KELAS_2 attribute is needed. Data search based on these conditions is done by using SQL queries.
3.3 Classification Scheme
3.3.1 Pleaides Image
Based on the classification scheme, the Pleiades image is very superior in terms of spectral which has four channel bands (red, green, blue and NIR) and band ratio (RVI, NDVI and NDWI) in identifying various kinds of objects. In addition, there are also subclass orchard from the class vegetation not found in subclasses from other sensor data, which in their identification use spatial criteria and texture criteria to distinguish these subclasses from other subclasses of the class vegetation.
3.3.2 nDSM
From the nDSM classification scheme, there is a difference with the classification scheme in the Pleiades imagery. This difference is very visible in the subclass road which is included in the class open area while in the class built up area the subclass building is divided into subclass residential buildings and subclass infrastructure buildings. This is because objects in the form of roads have relatively the same elevation as open area objects compared to built-up area objects. Meanwhile, the division of building subclasses into two subclasses can be done easily based on relative elevation values. In addition, another difference is also found in the absence of subclass orchard, due to limited identification features.
3.3.3 Intensity Image
In this classification scheme, it can be seen that there are many similarities in the class arrangement with the classification scheme used in the Pleiades imagery. The most striking difference is in the subclass lawn which is included in the class open area instead of the class vegetation. This is because the object has intensity characteristics that tend to resemble open area and does not give a significant value to the first of many return image like other vegetation subclasses. In addition, another difference is also found in the absence of subclass orchard, due to limited identification features.
3.4 Accuracy Test Result
Among the 36 test points, in the dataset A, the number of objects that match the ground truth is 16 points resulting in an overall accuracy value of 44.44%. Meanwhile, with the integration of LiDAR sensor data, dataset B (Pleiades-intensity) obtained 23 accurate test points resulting in an overall accuracy value of 63.89%, while dataset C (Pleiades-nDSM) also obtained 23 accurate test points resulting in an overall accuracy value of 63.89%.
3.4.1 Dataset A
In terms of producer accuracy, there are four subclasses that are classified accurately, namely bushes, pond, lake and river, while the subclasses for lawn, swamp, bare ground and other paved surfaces have the lowest accuracy where they are not classified correctly (indicated by the omission value). Meanwhile, in terms of user accuracy, orchards are the easiest subclass to identify, while the other classes tend to be difficult to distinguish from one another (indicated by commission values).
3.4.2 Dataset B
In terms of the accuracy of the map producer, there are five subclasses that are classified accurately, namely bushes, pond, lake, river and building, while swamp, bare ground and other paved surfaces are the most inaccurate subclasses (indicated by large omission values). Furthermore, in terms of map user accuracy, there are five subclasses that are the easiest to identify, namely bushes, lake, river, bare ground and building, while there are four subclasses that are still difficult to identify, namely orchard, swamp, road and other paved surfaces.
3.4.3 Dataset C
Dataset C has good mapping accuracy for six subclasses, namely bushes, trees, pond, lake, river and bare ground which can be classified accurately. However, for lawn, swamps and other paved surfaces are still the most inaccurate subclasses (indicated by the omission value). Meanwhile, in terms of map user accuracy, dataset C has almost the same accuracy as dataset B, only with slightly lower building accuracy. Thus, the accuracy of the classification results for dataset C is still better than dataset A and is equivalent to dataset B but only differs in the distribution of the subclasses.
3.5 Object Class Matching Up Analysis Results
Based on the average value, dataset C obtained the highest match up for class objects, namely 45.78% compared to dataset B which was 36.61%, where the average percentage value produced in dataset C was quite significant when compared to dataset B. Below are some examples of mismatched classification evident with various causes each from Dataset B and C.
3.5.1 Dataset B
3.5.2 Dataset C
4. Conclusion
Based on the results and discussion of the research obtained, the following conclusions are obtained:
1) The results of the Pleiades image classification (dataset A) using the OBIA method in the case study area as a whole produce a fairly low accuracy of 44.44%.
2) The results of the Pleiades image classification that have been added to LiDAR data (intensity images and nDSM) where each forms dataset B and dataset C, both of which are able to increase the accuracy of the classification results by 19.45% or nearly 20% compared to dataset A which only utilizes a single Pleiades sensor,
3) The low value of the accuracy of the classification results in this study when referring to SNI Land Cover/Land Use (minimum 70%) can be caused by the number of sample points which are only 36 sample points, very far from the number which is proportional to the total number of classified objects which reach thousands, which should cover at least 20% of the total number of classified objects to be able to provide significant test results,
4) Dataset B and dataset C each produce a different percentage of match up for object classes, where based on the lowest and highest values in dataset B produce object match up in the range of 30% to 41%, and dataset C produces object match up in the range of 31% to 56%. It can be inferred that object class matching up does not affect the increase in the accuracy of the classification results against ground truth.
5. Acknowledgement
The authors declare no any competing interest while writing this article. You can read the original version (written in Bahasa Indonesia) in this link.
6. Reference
Achsan, A. C. (2017). Pemanfaatan Citra Landsat untuk Klasifikasi Tutupan Lahan Lanskap Perkotaan Kota Palu. E-Journal Arsitektur Lansekap, 59.
Alonso, M., & Malpica, J. (2010). Satellite Imagery Classification with Lidar Data. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Science.
Aristalindra, F., Santosa, P. B., Diyono & Subaryono. (2020). Evaluasi Pemanfaatan Citra Tegak Satelit Resolusi Tinggi untuk Percepatan Pembuatan Peta Blok Pajak Bumi dan Bangunan (PBB-P2) secara Partisipatif di Desa Triharjo, Kabupaten Bantul, DIY. Journal of Geospatial Information Science and Engineering, vol 3 no 1. https://doi.org/10.22146/jgise.55788
Awrangjeb, M., Zhang, C., & Fraser, C. (2013). Automatic extraction of building roofs using LiDAR data and multispectral Imagery. ISPRS Jorunal of Photogrammetry and Remote Sensing, 83, 1-18.
Danoedoro, P. (2006). Versatile Land-use Information for Local Planning in Indonesia. Centre for Remote Sensing and Spatial Information Science (CRSSIS): School of Geography, Planning and Architecture. The University of Queensland.
Gilani, S., Awrangjeb, M., & Lu, G. (2015). Fusion of LiDAR data and multispectral imagery for effective building detection based on graph and connected component analysis. The International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, 40, 65.
Guo, Q. H., Kelly, M., Gong, P., & Liu, D. (2007). An object-based classification approach in mapping tree mortality using high spatial resolution imagery. GIScience dan Remote Sensing.
Istarno. (2016). Buku Ajar Penginderaan Jauh Sensor Aktif Airborne Laser Scanning/Lidar. Yogyakarta: Departemen Teknik Geodesi UGM.
Jia, Y. (2015). Object-based land cover classification with orthophoto and lidar data. Stockholm: School of Architecture and the Built Environment (KTH) Stockholm, Sweden.
Kushardono, D. (1997). Model Klasifikasi Penutup Lahan Data Multisensor Radar-Optik. Majalah LAPAN no.83.
Kushardono, D. (2017). Klasifikasi Digital Penginderaan Jauh. Bogor: Penerbit IPB Press.
Mills, M. (2020, June 23). Elevation Grid Creation in Global Mapper: Creating a DTM. Retrieved from Insight from Blue Marble Geographics: https://blog.bluemarblegeo.com/2020/06/23/elevation-grid-creation-in-global-mapper-creating-a-dtm/
Prayogo, L. M. & Basith, A. (2020). Uji Performa Citra Worldview 3 dan Sentinel 2A untuk Pemetaan Kedalaman Laut Dangkal (Studi Kasus di Kepulauan Karimunjawa, Jawa Tengah). Journal of Geospatial Information Science and Engineering, vol 3 no 2. https://doi.org/10.22146/jgise.59572
Sampurno, R., & Thoriq, A. (2016). Klasifikasi Tutupan Lahan Menggunakan Citra Landsat 8 Operational Land Imager (OLI) di Kabupaten Sumedang. Jurnal Teknotan Vol.10.2.
Santosa, P. B. (2016). Evaluation of satellite image correction methods caused by differential terrain illumination. Forum Geografi, vol 30 no 1. doi: 10.23917/forgeo.v30i1.1768
SNI 7645. (2010). SNI 7645: 2010 tentang Klasifikasi Penutup Lahan. Badan Standardisasi Nasional.
Suradji, A. (2020, May 18). Webinar PT Waindo Specterra. Retrieved from Teknologi GIS Enterprise, Dashboard Operation, CSRT, airborne LiDAR dan Aplikasi Pemanfaatannya di Bidang Pertambangan: https://youtu.be/1e9LgCy1yqs
Uzar, M., & Yastikli, N. (2013). Automatic building extraction using LiDAR and aerial photographs. Boletim de Ciências Geodésicas, 153-171.
Wahdaniyat, H. (2019, August 20). Ruang Terbuka Hijau yang Masih Terpinggirkan di Indonesia. Retrieved from Direktorat Jenderal Cipta Karya, Kementerian Pekerjaan Umum dan Perumahan Rakyat: http://ciptakarya.pu.go.id/pbl/index.php/detail_berita/565/ruang-terbuka-hijau-yang-masih-terpinggirkan-di-indonesia
.jpeg&w=3840&q=75)
