Global Revisit Interval Analysis of Landsat-8 -9 and Sentinel-2A -2B Data for Terrestrial Monitoring

The combination of Landsat-8, Landsat-9, Sentinel-2A and Sentinel-2B data provides a new perspective in remote sensing application for terrestrial monitoring. Jointly, these four sensors together offer global 10–30-m multi-spectral data coverage at a higher temporal revisit frequency. In this study, combinations of four sensors were used to examine the revisit interval by modelled orbit swath information. To investigate different factors that could influence data availability, an analysis was carried out for one year based on daytime surface observations of Landsat-8 and Sentinel-2A -2B. We found that (i) the global median average of revisit intervals for the combination of four sensors was 2.3 days; (ii) the global mean average number of surface observations was 141.4 for the combination of Landsat-8 and Sentinel-2A -2B; (iii) the global mean average cloud-weighted number of observations for the three sensors combined was 81.9. Three different locations were selected to compare with the cloud-weighted number of observations, and the results show an appropriate accuracy. The utility of combining four sensors together and the implication for terrestrial monitoring are discussed.


Introduction
Satellite combinations of the polar-orbiting Landsat-8 (launched 2013) and Landsat-9 (proposed for launch in middle 2021) by NASA [1] as well as Sentinel-2A (launched 2015) and Sentinel-2B (launched 2017) [2] by European Space Agency (ESA) offer 10-30-m resolution multi-spectral global land coverage. This will substantially increase moderate-resolution satellite observations available for terrestrial monitoring [3]. The data availability of satellite observations is of great importance to the surface land monitoring capabilities, as more data enable more reliable land cover classification and change detection.
The data availability of satellite surface observations changes spatially and temporally and is complicated due to the fact that different factors influence data availability. Combinations of sensors, taking advantage of the different sensor acquisition patterns, could enable more observations to be collated, thus reducing the temporal revisit interval between consecutive observations. Recently, Li and Roy [4] proved that the combination between Landsat-8 and Sentinel-2A -2B could provide more observations and derive a global median average revisit interval of 2.9 days. However, to date the global revisit interval between the combination of Landsat-8 -9 and Sentinel-2A -2B four sensors together has not been investigated.
Satellite orbit swath geometry, i.e., the spatial overlap of lateral orbit swaths increases with higher latitudes [5], which enables more observations at higher latitude. Solar geometry, i.e., latitudinal and temporal variations in the highest latitude toward the North or South Pole that satellites can observe, is related to the temporal progression of the solar position over a year [6]. Acquisition strategy and For all three sensors, observations acquired in 2018 from 1 January to 31 December were selected in this study. Only daytime-acquired imageries and global land acquisitions, which even included Antarctica, were used for Landsat-8. Landsat-8 now has the capability for mapping and monitoring snow/ice and water [22,23], with improved radiometric resolution and geolocation accuracy [24][25][26]. Sentinel-2 observations were filtered by descending orbit.

Surface Reflectance Observations for Landsat-8
Landsat 8 Collection 1 atmospherically corrected the surface reflectance image covering three pixel locations: northwest of Algeria, Sahara desert (30. [20]. These three locations were selected as they have the different land cover types, different latitude/longitude and different cloud conditions. Landsat-8 Operational Land Imager (OLI) Collection 1 Surface Reflectances are generated from the Top of Atmosphere Reflectance, using the Land Surface Reflectance Code (LaSRC) [27], which produces the surface reflectance bands and pixel quality assessment band.

Global Average Revisit Intervals for Combination of Landsat-8/9 and Sentinel-2A/2B
The global average revisit interval map was derived on a global land point grid using a sinusoidal equal area projection to provide spatially unbiased sampling [28]. The grid comprised 7201 × 3601 points with a spacing of 0.05 • . This map captured the overlap of along-track and across-track swath data from Landsat, as well as orbital shifts of the sensor geometry [29,30].
To derive the average revisit interval for each land grid point, each acquisition from the four sensors was independently tested to determine whether it encompassed the land grid point. This was fulfilled by comparing the corner coordinates of an acquisition with those of the land grid point [31]. Considering the large data volume of the sensors, a pre-sorting algorithm was implemented to filter acquisitions with the central coordinates away from the land grid point by a threshold. After establishing whether the acquisition overpassed the land grid point for all four sensors, they were sorted and merged into a single acquisition queue by order of acquisition time, given they were derived from different sensors. The revisit interval dataset was determined by calculating the time difference between every two consecutive observations. The average value of the revisit interval dataset was assigned to the land grid point. A fill value was given, if there were no acquisitions. After looping through all the land grid points, a global average revisit interval map was established. Because the operation on each land grid point was independent, a multi-thread technology was used to speed up the processing of assessing the grid points. All programs were written in C language.

Global Number of Observation Maps for the Daytime Surface Observations of Landsat-8 and Sentinel-2A -2B
The global land grid points defined in sinusoidal projection were used to derive the daytime surface observations of Landsat-8 and Sentinel-2A -2B. The first step was to establish the acquisition dataset for the three sensors that overpassed each land grid point. The total number of observations were added by counting these datasets. The spacing of land grid points was set to be small enough to capture overlap between satellite observations along-track and across-track. In the along-track direction, the southern part of the overlapping area was discarded, while the northern part was retained. In the across-track direction, the overlapping areas were counted twice, as they represent different observations sensed on different dates. The average cloud cover percentage at each land grid point was derived by averaging the cloud cover percentage of each acquisition in the sensed dataset list for each land grid point. A unique fill value was given if no observations were made at a given grid point.
The cloud-obscured images clearly decrease the number of available images. In an image scene, it was assumed that all the image pixels had the same probability of being cloudy, with a value equal to the percent cloud cover in the image scene. This ensured that the number of cloud-contaminated pixels, i.e., the number of datasets lost, was proportional to the cloud cover percentage. Likewise, all useful pixel observations, i.e., those representing the clear part of the image, were proportional to the fraction of cloud cover subtracted from one. Consequently, the cloud-weighted number of observations accumulated within a given period was obtained by the probability of observations that overpassed the land grid point being clear.
Three pixel locations were selected to evaluate the accuracy of the cloud-weighted number of observations. This was fulfilled by counting the number of clear views for each of the locations through the year 2018. Pixel observations were considered as a clear view only if they were not labelled as median confidence or high confidence cloud in the pixel quality assessment band [32]. The accuracy of the cloud-weighted number of observations was compared with the cloud-weighted number of observations with the number of clear views.

Global Average
Revisit Intervals for Combination of Landsat-8/9 and Sentinel-2A/2B Figure 1 shows the average revisit intervals derived for each global land grid point for Landsat-8 and Landsat-9 and four sensors combined. The average revisit intervals for Sentinel-2A and Sentinel-2B, reported by [4], are shown for comparison. Given the wider swath width of Sentinel-2, but its longer repeat cycle compared with Landsat, the combinations of Sentinel-2A/2B had a shorter average revisit interval than Landsats-8/9 (Figure 1a,b). As shown in Figure 2, the global revisit interval histograms are not normally distributed because of the variable overlap of the orbits of different sensors and convergence of their orbits at high latitudes. The values beyond 9.0 days were not shown in Figure 2 for the low appearance (account for 0.028%, 0.007% and 0.000% of total grid points, respectively). Table 1 summarises the global mean, median, first mode and the second mode revisit interval data for the various sensor combinations explored in this study. to the percent cloud cover in the image scene. This ensured that the number of cloud-contaminated pixels, i.e., the number of datasets lost, was proportional to the cloud cover percentage. Likewise, all useful pixel observations, i.e., those representing the clear part of the image, were proportional to the fraction of cloud cover subtracted from one. Consequently, the cloud-weighted number of observations accumulated within a given period was obtained by the probability of observations that overpassed the land grid point being clear. Three pixel locations were selected to evaluate the accuracy of the cloud-weighted number of observations. This was fulfilled by counting the number of clear views for each of the locations through the year 2018. Pixel observations were considered as a clear view only if they were not labelled as median confidence or high confidence cloud in the pixel quality assessment band [32]. The accuracy of the cloud-weighted number of observations was compared with the cloud-weighted number of observations with the number of clear views. Figure 1 shows the average revisit intervals derived for each global land grid point for Landsat-8 and Landsat-9 and four sensors combined. The average revisit intervals for Sentinel-2A and Sentinel-2B, reported by [4], are shown for comparison. Given the wider swath width of Sentinel-2, but its longer repeat cycle compared with Landsat, the combinations of Sentinel-2A/2B had a shorter average revisit interval than Landsats-8/9 (Figure 1a,b). As shown in Figure 2, the global revisit interval histograms are not normally distributed because of the variable overlap of the orbits of different sensors and convergence of their orbits at high latitudes. The values beyond 9.0 days were not shown in Figure 2 for the low appearance (account for 0.028%, 0.007% and 0.000% of total grid points, respectively). Table 1 summarises the global mean, median, first mode and the second mode revisit interval data for the various sensor combinations explored in this study.    The combination of more sensors and the utility of their orbit swaths facilitated more observations at given land grid point and decreased the revisit interval between consecutive observations, as seen in Figure 1a-c. The global median average revisit intervals were: 8 d for   The combination of more sensors and the utility of their orbit swaths facilitated more observations at given land grid point and decreased the revisit interval between consecutive observations, as seen in Figure 1a-c. The global median average revisit intervals were: 8 d for Landsat-8/9, and 3.7 d for Sentinel-2A/2B. When four sensors were combined, the utility of their different swaths decreased the median average revisit interval to about 2.3 d. Figure 3 shows the number of surface land observations for each of the land grid points for Landsat-8, Sentinel-2A, Sentinel-2B and three sensors combined in 2018. Where there were no observations, the land grid point was colored grey. To make the global map spatially explicit, country boundaries as well as latitude and longitude grids were overlapped, using a sinusoidal projection interval every 30 • .

Global Number of Observations for Landsat-8 and Sentinel-2A -2B
Sensors 2020, 20, x FOR PEER REVIEW 6 of 15 Landsat-8/9, and 3.7 d for Sentinel-2A/2B. When four sensors were combined, the utility of their different swaths decreased the median average revisit interval to about 2.3 d. Figure 3 shows the number of surface land observations for each of the land grid points for Landsat-8, Sentinel-2A, Sentinel-2B and three sensors combined in 2018. Where there were no observations, the land grid point was colored grey. To make the global map spatially explicit, country boundaries as well as latitude and longitude grids were overlapped, using a sinusoidal projection interval every 30°.  Figure 3a). Clearly, different data reception strategies and orbit geometry influenced the data availability. Most of the Landsat-8 observations were located on land, but there were several observations over oceans. This is because Landsat-8 carried out limited night imaging to monitor active volcanoes and islands worldwide, as both these targets were set to have high imaging priority [8]. Landsat-8 acquired more images at high latitudes, especially above a latitude of 60°, because its swaths overlap more at high latitudes. Combining more sensors enables more data observations, which can be seen from  Figure 4 shows the average number of Landsat-8 satellite observations over each land grid point for June (upper panel) and December (lower panel) of 2018. These two months were selected because, during the summer solstice (21 June) and winter solstice (22 December), the North Pole has its maximum and minimum tilt towards the Sun, respectively. Considering the repeat cycle of Landsat-8 is 16 d, there should be no more than two observations in any 1-mo period, but the observed number of observations may be greater than two because of the overlap of lateral swaths and the convergence of its orbit at a higher latitude. In fact, the average global number of acquisitions for Landsat-8 was 3.17 in June and 3.05 in December, respectively. The total number of surface observations of Landsat-8 derived for 2018 show a complex pattern ( Figure 3a). Clearly, different data reception strategies and orbit geometry influenced the data availability. Most of the Landsat-8 observations were located on land, but there were several observations over oceans. This is because Landsat-8 carried out limited night imaging to monitor active volcanoes and islands worldwide, as both these targets were set to have high imaging priority [8]. Landsat-8 acquired more images at high latitudes, especially above a latitude of 60 • , because its swaths overlap more at high latitudes. Combining more sensors enables more data observations, which can be seen from   Satellite orbit sensor geometry clearly influences global data availability, because the lateral swath convergence at higher latitudes produces more observations at a given grid point. Figure 5 shows the mean average total number of observations by averaging all values along a given latitude (  Given the annual progression of the solar position, the geographic coverage of the polar area varies, as the satellite track moves into darkness [30]. The maximum geographic coverage of Landsat-8 towards the south is 55.08 • S in June, while there are no observations above a latitude of 66.71 • N by Landsat-8 in December.

Global Number of Observations for Landsat-8 and Sentinel-2A -2B
Satellite orbit sensor geometry clearly influences global data availability, because the lateral swath convergence at higher latitudes produces more observations at a given grid point. Figure 5 shows the mean average total number of observations by averaging all values along a given latitude (Figure 3  Generally, the latitudinal mean average number of observations for Sentinel-2 is higher than Landsat-8 between the 60° N and 60° S due to Sentinel-2′s wider orbit swath and shorter revisit interval. Landsat-8 acquires more satellite observations on the two pole area for the image acquiring strategy [8].  Figure 6 shows the global average percent cloud cover examined over land grid points that had at least one Landsat-8 daytime observation in 2018. Typically, high cloud cover occurred over tropical rainforest areas near the equator, while desert and dryland areas typically had low cloud cover. The global mean average percent cloud cover derived from all Landsat-8 daytime observations for 2018 defined on the global equal area sinusoidal projection was 0.41. Generally, the latitudinal mean average number of observations for Sentinel-2 is higher than Landsat-8 between the 60 • N and 60 • S due to Sentinel-2 s wider orbit swath and shorter revisit interval. Landsat-8 acquires more satellite observations on the two pole area for the image acquiring strategy [8].  Figure 6 shows the global average percent cloud cover examined over land grid points that had at least one Landsat-8 daytime observation in 2018. Typically, high cloud cover occurred over tropical rainforest areas near the equator, while desert and dryland areas typically had low cloud cover. The global mean average percent cloud cover derived from all Landsat-8 daytime observations for 2018 defined on the global equal area sinusoidal projection was 0.41. Figure 7 shows histograms of the global average percent cloud cover data ( Figure 6). The data for Landsat-8 average percent cloud cover were asymmetrically distributed, with a lower limit of 0 cutting the curve. Across the global map of average percent cloud cover data, the most common values were 0.4 to 0.5 for Landsat-8, occurring at 20.75% of the global grid points.    The geographical distribution pattern of cloud-weighted observations is complex and irregular. Generally, data availability was influenced by the sensor combination, data reception strategy and the system mission constraints (Figure 3). Meanwhile, more observations were carried out at high latitudes, related to greater lateral orbit swath overlap in those areas. In addition to these factors, the cloudweighted number of observations determined the cloud contamination level of all data. Thus, areas with frequent high cloud cover, e.g., tropical rainforests, were more severely contaminated by clouds, while low cloud cover over deserts and drylands ensured a higher probability of clear view observations. Overall   Figure 7 shows histograms of the global average percent cloud cover data ( Figure 6). The data for Landsat-8 average percent cloud cover were asymmetrically distributed, with a lower limit of 0 cutting the curve. Across the global map of average percent cloud cover data, the most common values were 0.4 to 0.5 for Landsat-8, occurring at 20.75% of the global grid points.  The geographical distribution pattern of cloud-weighted observations is complex and irregular. Generally, data availability was influenced by the sensor combination, data reception strategy and the system mission constraints (Figure 3). Meanwhile, more observations were carried out at high latitudes, related to greater lateral orbit swath overlap in those areas. In addition to these factors, the cloudweighted number of observations determined the cloud contamination level of all data. Thus, areas with frequent high cloud cover, e.g., tropical rainforests, were more severely contaminated by clouds, while low cloud cover over deserts and drylands ensured a higher probability of clear view observations. Overall   The geographical distribution pattern of cloud-weighted observations is complex and irregular. Generally, data availability was influenced by the sensor combination, data reception strategy and the system mission constraints (Figure 3). Meanwhile, more observations were carried out at high latitudes, related to greater lateral orbit swath overlap in those areas. In addition to these factors, the cloud-weighted number of observations determined the cloud contamination level of all data. Thus, areas with frequent high cloud cover, e.g., tropical rainforests, were more severely contaminated by clouds, while low cloud cover over deserts and drylands ensured a higher probability of clear view observations. Overall   Table 2 summarises the number of observations, cloud-weighted observations, and clear views, as well as accuracy levels, for the three selected locations for the year 2018. The accuracy is 98.7%, 91.0% and 81.7% for Algeria, (30.0° N, 0.0°), Brazil (3.138° S, 62.180° W) and Sweden (56.842° N, 15.057° E), respectively, evaluated by comparing the number of cloud-weighted observations (42.45, 9.81, 18.93) with the number of clear views (43,9,16). Table 3 (42.45, 9.81, 18.93) with the number of clear views (43,9,16). Table 3 shows the acquisition date, path, row and cloud condition of Landsat-8 acquisition imageries covering the selected location in the northwest of Algeria (30.

Discussion
The data availability of satellite observations influences surface land monitoring capabilities. Having more observations in a given time enables more reliable time series fitting [33,34], higher precision land cover classification [35], improved stable land change detection [36] and more cloud-free composited products [37,38]. The global spatial coverage of satellite observations enables the large area monitoring, i.e., on a regional or global scale, of land cover change [39] and the mapping of burned areas [40]. The polar-orbiting Landsat-8 satellite has even acquired high latitude area observations that enable ice flow mapping [23].
The combination of the Landsat-8 -9 and Sentinel-2A -2B four sensors together was enabled to develop a dense time series, improving the ability to detect abrupt land cover changes, and monitor phenology variations at a specific time period [41,42]. Combining the four sensors could offer a higher temporal resolution, addressing the gap of the observation sample data for model training, caused by the cloud obscuration and data missing issue by system [43].
More sensors are combined to facilitate more observations, shorter revisit interval between consecutive observations will be got. With similar multi-spectral bands, Sentinel-2A -2B and Landsat-8 -9 combined together provide 10-30-m resolution global land coverage. Compared to the 2.9 days from the combinations of Landsat-8 and Sentinel-2A -2B, the global median average revisit interval for the four sensors combined are 2.3 days. This increase in revisit interval was not striking, as only one Landsat was added in and also because Sentinel-2 has a wider swath coverage than Landsat. The combination of the four moderate-resolution sensors could still advance the solution for near daily temporal coverage that can benefit for many applications, etc., drought monitoring [44] and evapotranspiration estimations [45].
The global mean average number of observations was 162.6 for the combination of Landsat-8 and Sentinel-2A -2B in 2018, derived from the orbit swath model [4]. In this study, the global mean average number of observations derived from daytime surface observations of three sensors combined was 141.40, reduced by 13.0%. The global mean average number of observations derived from the orbit swath model (162.6) considers orbit swath geometry and assumes that at each location an observation is acquired with equal opportunity, without considering data acquisition strategy, system reception ability or instrument issues. Thus, the 13.0% reduction gives a global overall estimation of the influence of date acquisition strategy and instrument issues on data availability. Effects of cloud cover on satellite images used for surface monitoring are important. 2) for Sentinel-2B, respectively. As all of the three sensors acquire observations with different probabilities at different latitudes according to their data reception strategy [8], this is not a single factor analysis. Solar geometry, i.e., the annual progression of the solar position, denotes the geographical latitudinal coverage of observations that can be acquired during the daytime. The maximum latitudinal coverage towards the south for Landsat-8 is 55.08 • S in the month of June (summer solstice), and towards the north is 66.71 • N in the month of December (winter solstice).
In this study, the surface observation availability was examined on a tile level. The cloud-weighted number of observations assumed that each pixel showing cloud in the image was the same and equal to the percentage of the image being cloudy. Consequently, the shape and exact location of clouds over the area was not clear. However, the reported results give an overall evaluation of cloud-free observation areas in each image frame. The complicated pattern of data availability related to cloud was apparent in the global-scale map.

Conclusions
This study demonstrates that sensor combination, system reception, orbit geometry, solar geometry, and cloud contamination could all influence data availability. The main findings of the research were as follows: (i) Sensor combination enabled more observations and shorter revisit intervals between consecutive observations. The global median average revisit intervals for various combinations were: 8.0 d for Landsat-8 and Landsat-9, 3.7 d for Sentinel-2A and Sentinel-2B, and only 2.3 d when all four sensors were combined; (ii) The global mean average number of surface observations for the combination of Landsat-8 and Sentinel-2A -2B is 141.4; (iii) The global mean average cloud-weighted number of observations is 81.9 for the three sensors combined; (iv) Landsat-8 surface reflectance covering three different locations was used to compare the cloud-weighted number of observations. The results show an overall accuracy of more than 80%.
Given its similar spectral and spatial characteristics as Landsat-8/9, Sentinel-2A/2B data could be combined with Landsat data to provide better moderate-resolution imaging. The modeled orbit swath data obtained from COVE was used to analyse the revisit interval between consecutive observations from combined sensors. Future work could include Sentinel-2 surface land observations combined with Landsat ones to derive the surface land observation analysis at global scales for the combination of Landsat-8/9 and Sentinel-2A/2B.