Effect of Spectral Bandwidths on Linear Feature Extraction: An Evaluation of Landsat ETM+ and OLI Sensors

Hitherto there have been many studies comparing the usefulness of OLI and ETM+ sensors for linear feature extraction. However, not too much attention has been paid to the differences in the bandwidth of the two sensors. In this study, the suitability of Landsat ETM+ and OLI sensors for automatic detection of linear features by LINE algorithm was compared. In this study, eight regions in northern, central and southern parts of Iran were selected based on the diversity of lithology, the pristine status, and lack of human activities for the comparison of the two datasets. Results revealed that LINE algorithm performed better on the images with higher standard deviation. The ETM+ datasets are more suitable for linear feature extraction because ETM+ panchromatic band and first principal component analysis image (PC1 image) of ETM+ datasets have higher standard deviation compared to OLI datasets.


Introduction
Detection and extraction of linear features using digital images such as satellite or aerial images is a very important low-level operation in computer vision which has several applications. For instance, satellite imagery can be used to extract roads, railroads, rivers and other targets, and can be applied for capturing or updating data for navigation systems or for production of basic geographic maps. Remote sensing as a powerful tool to obtain information from the surface of the earth has been used widely for this purpose. In recent decades, remote sensing techniques have made progress in the earth science studies by aiding in the identification of linear features [1][2][3][4][5][6]. Several methodologies have been developed for the extraction of linear features. These include Kalman filtering, dynamic programming, multi-scale grouping and context, fuzzy-based approach, model-based approach, mathematical morphology, and image fusion. LINE algorithm in the PCI Geomatica software is the most commonly used automatic approach for linear feature extraction [7][8][9]. This algorithm employs the Canny method and detects the edge in images in three consecutive stages, namely edge detection, thresholding and curve extraction [10]. This study aims to compare the suitability of Landsat 7's ETM+ and Landsat 8's OLI data for linear feature extraction using LINE algorithm in PCI Geomatica Software. For this purpose, the performance of LINE algorithm on panchromatic bands and PC1 images of these datasets

Study Area
The study areas were chosen to contain various lithological diversities, e.g., igneous, sedimentary, and metamorphic rocks. The study areas are located in various regions in Iran, near Tabriz, Mashhad, Yasuj, Shahr-E-Kord, Yazd, Rafsanjan, Larestan, and Iranshahr cities. Yazd and Shahrekord cities are located in central Iranian block. This block is one of the most complicated structural zones in Iran. In this zone, rocks are deformed by several orogenic episodes, and metamorphism and magmatism is abundant. Shahrekord city is located in the Sanandaj-Sirjan zone. A striking feature of this zone is the presence of immense volumes of magmatic and metamorphic rocks. Mashhad city is located in Kopet Dagh. No significant sedimentary gap or volcanic formations are reported in this zone, and their sedimentary rocks were placed in their current position due to the uplift of the crust. Iran-Shahr city is located in the Makran zone and contains the ophiolites overlain by a thick sequence of sandstone, shale, and marl about 5,000 m thick [11][12][13][14]. Figure 1 shows the locations of the study areas on the structural map of Iran.

Remote Sensing Data
Enhanced Thematic Mapper Plus (ETM+) sensor onboard Landsat 7 satellite with a repeat cycle of 16 days was launched into space in July 1999. The Scan Line Corrector (SLC) instrument of the ETM+ sensor stopped working correctly in 2003, and consequently, the images collected by this sensor since May 30, 2003, have data gaps called SLC-off images. Operational Land Imager (OLI) is an instrument onboard Landsat 8 satellite, which was launched in February 2013. The spectral bands of the OLI sensor are similar to Landsat 7's ETM+ sensor except for two additional spectral bands. Band 1 is a deep blue visible channel designed for water resources and coastal zone observations, and band nine is designed for the detection of cirrus clouds. Similar to ETM+ data, Landsat 8's OLI data has 30-meters spatial resolution for bands 1 to 7 and band 9. Additionally, the band 8 is a panchromatic band with 15-meter spatial resolution. OLI scene size is approximately 170 km north-south by 183 km east-west similar to Landsat 7 data. The collection category of both ETM+ and OLI images used in this study are Tier 1 (T1). This category of Landsat satellite series is an inventory structure for Landsat Collection 1 Level-1 data products. These contain the highest quality Level 1 Precision Terrain (L1TP) data considered suitable for time-series analysis. The geo-registration of Tier 1 data is consistent, with tolerances less than 12 meters root mean square error (RMSE). The datasets used in this study were collected in the time window of October 2002 to June 2017.

Image Analysis
The failure of SLC instrument of the ETM+ sensor caused a time gap between the perfect ETM+ images collected before 2003 and OLI data collected from 2013 onwards. Considering that the ETM+ data were collected fourteen years before that of the OLI, to eliminate the effects of human artifacts like pipelines, railways, and roads on image comparison of the two datasets, relatively pristine areas devoid of human activities were chosen as areas of study. Figure 2 shows the pristine nature of the study area around Yasuj and Mashhad cities. One of the factors affecting linear feature extraction is the angle of solar radiation [15]. The change of radiation angle changes the length of shadows. Since the images of Landsat 7 and 8 were taken at the same time in Iran, this parameter did not have a considerable effect on image pair comparison. To prepare the datasets for comparison, the digital numbers (DNs) were converted to radiance using Equation (1).
(1) Where, λ is spectral radiance (W/(m 2 *sr*μm)), L is radiance multiplicative scaling factor for the band, is pixel's value in DN and L is radiance additive scaling factor for the band. Afterwards, radiance data was corrected to the surface reflectance using the Dark Object Subtraction method (DOS) [16][17][18]. Because multispectral data are often highly correlated, Principal Components Analysis was used to produces uncorrelated output bands to segregate noise components and to reduce the dimensionality of datasets [19,20]. To obtain the PC1 image bands 1 to 5, and 7 of ETM+ and bands 1 to 7 and 9 of OLI sensor were used. The suitability of PC1 of ETM+ dataset was compared to PC1 image of OLI dataset in each study area. Likewise, the panchromatic band of ETM+ dataset was compared with the panchromatic band of OLI. To determine the superiority of each dataset, the total length and number of detected linear features were compared to each other. Linear features were detected by LINE algorithm using PCI Geomatica software. This algorithm needs six input parameters and executes in three consecutive steps, namely edge detection, thresholding, and curve extraction. The LINE algorithm primarily operates on a single 8-bit image channel. If the input image is not in 8 bits, the image is first converted to an 8-bit image using a nonlinear scaling function. To consider the effect of this conversion on the LINE algorithm, all of the OLI images that were initially in 16-bit depth were also converted to 8-bit data manually before using LINE algorithm. For the manual conversion of 16-bit images to 8-bit images, both nonlinear logarithmic function and linear function were used. The linear function rescales the highest value of 16-bit data (corresponding to a DN of 2^16 -1 = 65535) to a value of 255 ( 2^8 -1) and its lowest value to 0, and the values in between are scaled proportionately using Equation (2).
Where, . is the DN value after rescaling to 8 bits, is the DN value of each pixel in 16-bit, is the maximum value in 16-bit depth (65535) and is the maximum value in 8-bit depth (255). The logarithmic function converts DN values to the logarithm in base 10 using Equation (3).
. is the re-scaled value from 16-bit to 8 bits, is the scaling constant, and is the value of each pixel in 16 bits. The logarithmic function will change the frequency chart of the DN values. In a logarithmic scale, variations of DNs of lower values are more pronounced than variations of pixels with higher DNs. Since the logarithm is not defined for 0, the value 1 must be added to the image before logarithmic conversion. The scaling constant is chosen so that the maximum output value ( The spectral bandwidth of the images and the standard deviation of images were compared to each other to find the relation between standard deviation, algorithm performance and bandwidth variations. Before computing the standard deviation of the images, the digital number values of all of the images were rescaled between 0 and 1 using Equation (5).
is the normalized value of digital number between 0 and 1, is digital number value of each pixel, and are minimum and maximum digital number values of the image. Figure 3 presents the proposed method for this comparison.

Results and Discussion
One of the most important parameters which affects the quality of satellite images is the signal-to-noise ratio (SNR). This phenomenon, which is caused by the electronic errors in the sensor's detectors, is an inevitable part of satellite systems imagery. In case of using satellite images to extract linear features, noises can be represented as short lines. Due to the possibility of occurrence of this error, the detected linear features which were shorter than the size of two pixels, were removed from results. The panchromatic bands of ETM+ and OLI sensors were given as an input to the LINE algorithm for linear feature detection without any prior conversion of OLI band to 8-bit data. The results of the operation on panchromatic bands for some study areas are given in Table 1 The PCA image for each dataset was rendered using ETM+ bands 1 to 5, and 7, and OLI bands 1 to 7, and 9, respectively. The results for execution of LINE algorithm on PC1 image derived from ETM+ and OLI datasets are tabulated in Table 2. The data indicate that in all of the study areas except for one study area (i.e. Rafsanjan study area), the LINE algorithm performed more satisfactorily on PC1 image derived from ETM+ dataset compared to PC1 image derived from OLI dataset. The LINE algorithm converts any input image to an 8-bit image by a nonlinear function before linear feature detection. This is true for OLI data which is in 16 bits. To investigate the effect of this conversion on the outcome of LINE algorithm, the OLI images used for all study areas were converted to 8-bit images with both linear and nonlinear logarithmic functions before execution of LINE algorithm using Equations (2) and (3). Figure 4 shows the histograms of one of the study areas of original OLI image and the converted versions of it to 8-bit with two mentioned functions. The results of LINE algorithm, on both types of converted images for each study area, are tabulated in Table 3.  Comparison of the standard deviations of all images in Table 4, including PC1 and panchromatic band after normalization by Equation (5) shows that logarithmic function always produced images with higher standard deviations compared to images converted by a linear function. Standard deviation is a measure of the spread of values around the mean. Lower the standard deviation more the values is clustered around the average. Standard deviation is a useful measure of image contrast. An image with a higher standard deviation shows higher contrast [21]. In all cases, the LINE algorithm gave better results on the image pair which was converted using a logarithmic function rather than the one converted by a linear function. However, execution of the LINE algorithm on 8-bit converted OLI images always gave poorer results compared to ETM+ 8-bit images. The LINE algorithm always performed better on image pairs which had a higher standard deviation (Tables 2, 3 and 4). This is the case for both panchromatic bands and PC1 images. The OLI PC1 image of the exceptional study area (i.e. Rafsanjan city) for which OLI PC1 image gave better results in LINE algorithm had higher standard deviation compared to ETM+ PC1 image. There seems to be more secured information for LINE algorithm in images with the higher standard deviation in each study area. The higher standard deviation of ETM+ bands is most probably a result of broader bandwidth of ETM+ sensor relative to the OLI sensor. The panchromatic band of the ETM+ sensor is 0.208 micrometers wider than the panchromatic band of OLI. Additionally, other bands of ETM+ sensor which record data in the reflective region of the electromagnetic spectrum are about 0.32 micrometers wider than the corresponding OLI bands. A band with wider spectral region contains more information reflected from the earth surface. When the goal is the detection of the linear phenomenon, a wider bandwidth seems to be more efficient because the edge can be composed of various materials with various spectral features. This is contrary to the usefulness of narrower bandwidths for the detection of specific spectral materials [22].Therefore, when detecting linear features by LINE algorithm, ETM+ images with higher spectral bandwidth are preferred over OLI images.

Conclusion
The LINE algorithm is the fastest and most efficient method for linear feature extraction. The resolving power of the sensor and its spatial resolution, and the bandwidth of the sensor are influential factors in linear feature detection. Images with higher standard deviation have higher contrast and give better results for linear feature extraction by LINE algorithm. Because standard deviation and image contrast are directly related to the spectral bandwidth of the sensor, the images recorded in wider spectral bandwidth are more suitable for linear feature extraction. It was shown that ETM+ panchromatic and PC1 bands almost always have higher standard deviation compared to panchromatic and PC1 bands of OLI datasets. Therefore, ETM+ datasets with higher bandwidth are more favourable for the linear feature extraction compared to OLI datasets.