COMPARISON OF LANDSAT-9 AND PRISMA SATELLITE DATA FOR LAND USE / LAND COVER CLASSIFICATION

: Land use and land cover (LU/LC) detection has great significance in management of natural resources and protection of environment. Hence, monitoring LU/LC with the state-of-the-art approaches has gained importance during the recent years and free access satellite images have become valuable data source. The aim of this study is to compare classification abilities of Landsat-9 and PRISMA satellite images while applying Support Vector Machine (SVM) algorithm to distinguish different LU/LC classes. For this purpose, the study area was chosen to be of heterogeneous character that includes industrial area, roads, residential area, airport, sea, forest, vegetation and barren land. When the classification results were visually examined, it was seen that forest, industrial area and airport classes were distinguished more accurately than other classes. On the other hand, qualitative results were validated with quantitative accuracy assessment results. The overall accuracy (OA) and Kappa coefficient values were calculated as 89.33 and 0.88 for Landsat-9 satellite image and as 92.33 and 0.91 for the PRISMA satellite image, respectively. In the accuracy assessment results, although Landsat-9 and PRISMA satellite images showed similar classification performances, a slight improvement was observed by using the PRISMA satellite image. The findings indicated that although both of the Landsat-9 and PRISMA satellite images were proper data to assess the LU/LC of the complex region, a slightly more performance could be achieved by using the PRISMA satellite image.


INTRODUCTION
Assessment of the land use and land cover (LU/LC) is playing a vital role for natural resource management and environmental protection (Tan et al., 2016;Tadese et al., 2020). Mapping the LU/LC accurately by using state-of-the-art approaches has become a major focus as it has profound impact on the ecosystem (Huang et al., 2020;Ge et al., 2020). Earth observation satellite imagery have started to provide rich information with the rapid development of remote sensing technologies (Du et al., 2012;Ge et al., 2020;Tong et al., 2020). In particular, remote sensingbased images that were based on a free access policy are the major data source for assessing the spatial extents of various LU/LC classes in a cost and time effective way (Mugo et al., 2020;Nguyen et al., 2020;Tadese et al., 2020;Weigand et al., 2020).
Landsat, which is a joint mission of The National Aeronautics and Space Administration (NASA) and US Geological Survey (USGS), provides consistent multispectral images over the past five decades. Landsat-9, which was launched in September 2021, is the latest of Landsat Earth-observing satellites series. Although free and open Landsat satellite images provide opportunities for LU/LC applications, multispectral satellite images have some limited factors to classify heterogeneous regions. Thus, the need arises to discriminate complex LU/LC classes by using satellite images that have higher spectral resolution. Hyperspectral satellite images offer a great number of spectral bands that enable * Corresponding author acquiring continuous spectral reflectance of the objects on the earth (Tan et al., 2005;Velez-Reyes et al., 2006;Vangi et al., 2021). PRecursore IperSpettrale della Missione Applicativa (PRISMA), which is a mission of Italian Space Agency (ASI), is a new hyperspectral earth observation satellite provides free products and helps to improve the understanding of environment with its high spectral resolution. High accuracies were achieved by using PRISMA satellite image in different classification applications such as detection of non-photosynthetic vegetation (Pepe et al., 2020), different forest types (Vangi et al., 2021), and burned areas (Lazzeri et al., 2021) due to providing narrow spectral bands (≤12 nm).
Numerous studies about usage of previous Landsat satellite images for LU/LC classification applications have been conducted (Coulter et al., 2016;Chen et al., 2020;Deng et al., 2019;Kumar et al., 2021). Landsat-9 satellite has the same technical properties with the Landsat-8 satellite (Url-1). As Landsat-9 is a recently launched satellite, LU/LC classification applications with Landsat-9 satellite image have not yet been extensively performed. On the other hand, PRISMA satellite images were widely used in studies on discriminating different types of land cover classes (Lazzeri et al., 2021;Vangi et al., 2021). However, there is a lack of studies about discriminating different LU/LC types by using PRISMA satellite image.
Comparison of classification performances of Landsat-9 and PRISMA satellite images which bear the same spatial resolution was aimed in this study for determining different LU/LC classes.
For this purpose, hyperparameters of the Radial Basis Function (RBF) kernel were tuned and pixel-based support vector machine (SVM) which is a machine learning classifier, was applied. In addition to qualitative accuracy assessment, quantitative analysis was also conducted based on error matrices, and accuracy metrics were calculated.

Study Area
The pilot area of approximately 125.85 km 2 is located in Çanakkale Province (Turkey) was selected as the study area and depicted in Figure 1. The reason for selecting the pilot area was that it consists diverse LU/LC classes like industrial area, roads, residential area, airport, sea, forest, vegetation and barren land. The secondary reason of that selection is that acquisition of Landsat-9 and PRISMA satellite images on the same day.

Data Used
The study was based on two different types of satellite data for assessing the contribution of the spectral resolution on the accuracy of LU/LC classification results. Hence, as satellite images which were acquired on the same day and have the same spatial resolution were selected. Landsat-9 and PRISMA satellite images were acquired on 18.01.2022 and included 11 and 235 spectral bands with 30-m spatial resolution, respectively. Technical specifications of the PRISMA and Landsat-9 satellite images are shown in Table 1 (Url-2) and  Table 2. Data specifications of the Landsat-9 Operational Land Imager 2 (OLI-2) Instrument (Url-1; Url-3; Url-4).

Methodology
This research consists of several steps, which are presented in a flow chart. (Figure 2). The details of the methods are explained in this section. Since, Landsat-9 Level-2 satellite image and PRISMA Level 2D satellite images that have been atmospherically corrected were used, no further atmospheric correction was applied to the satellite images. Some of the spectral bands (1-5, 13-15, 20-21, 35-37, 56-63 of VNIR Bands; 42-49, 168-173 of SWIR Bands) of PRISMA satellite image that included high level of noise were eliminated. The composite image was generated from the selected proper spectral bands and defined as input data at the PRISMA image classification stage. On the other hand, coastal, blue, green, red, near infrared, shortwave infrared 1 and shortwave infrared 2 of Landsat-9 image were defined as input data at the Landsat-9 image classification stage. Ships and clouds were masked in the satellite images. Radial Basis Function kernel SVM (Vapnik, 1979;Vapnik, 2013) that is a supervised machine learning algorithm and also a powerful classifier (Khatami and Mountrakis, 2012;Rana and Suryanarayana, 2020) was preferred to discriminate the complex classes. As a first step training dataset was collected in the classification stage. In the next step,  (Pedregosa et al., 2011). By using optimized hyperparameters, SVM algorithm was applied to these multispectral and hyperspectral satellite images to discriminate different LU/LC classes.
In assessing the accuracies of the results that were obtained by using PRISMA and Landsat-9 satellite images, quantitative and qualitative analysis were performed. Error matrices were generated based on 300 test points to calculate accuracy assessment metrics of OA, kappa coefficient, producer's and user's accuracy, and the classification results were evaluated quantitatively. Error matrices of the classification results that were obtained by using Landsat-9 and PRISMA satellite images are shown in Table 3 and Table 4

RESULTS AND CONCLUSIONS
The LU/LC maps that were produced by applying SVM to Landsat-9 and PRISMA satellite images were shown in Figure 3 and Figure 4, respectively.  In the visual analysis ( Figure 5), Google Earth images ( Figure  5.a, 5.d, 5.g) were used as reference data. It was observed that the airport was discriminated more accurately in the classification result ( Figure 5.b), which was retrieved from the PRISMA satellite image, than the classification result ( Figure 5.c) which was retrieved from the Landsat-9 satellite image. Similarly, industrial area was distinguished more accurately by using PRISMA image ( Figure 5.h), (Figure 5.i). On the other hand, roads were classified more accurately with Landsat-9 satellite image as shown in Figure 5.f, compared to PRISMA classification result that is shown in Figure 5.e. Figure 5. Magnified images of a, d, g reference maps; b, e, h the classification results that were obtained by PRISMA satellite image; c, f, i the classification results that were obtained by Landsat-9 satellite image.
Kappa coefficient values of the SVM classification results obtained from Landsat-9 and PRISMA were calculated as 0.88 and 0.91, OAs were calculated as 89.33% and 92.33% respectively. The accuracy assessment results are given in Table  5.  Table 5. The results of the accuracy assessment metrics.

Classes
In the accuracy assessment results, similar results were obtained from these two different classifications. However, due its higher spectral resolution, the classification capability of the PRISMA satellite image was observed as slightly higher than the Landsat-9 satellite image.