4.1. Sensor Summary Information
illustrates the annual number of Landsat ARD granules over the CONUS for each sensor over the 36-year study period. From 1999 to 2011 there were significantly more Landsat ARD granules over the CONUS than in other years because both the Landsat 5 TM and Landsat 7 ETM+ sensors were operating in this period. The annual maximum number of ARD granules (48,511) was in 2001. The annual minimum number of granules (1,517) was in 1983 when only Landsat 4 was on orbit. In Figure 2
a small number of Landsat 4 TM granules in 1984 (52) and Landsat 5 TM granules in 2012 (70) are not apparent because of the barplot y-axis scale. The Landsat 4 TM data were temporally quite discontinuous. In 1983, Landsat 4 lost the use of two of its four solar panels and both of its direct downlink transmitters, which caused a gap in observations until 1987, when the Tracking and Data Relay Satellite System (TDRSS) became operational and was used to relay Landsat 4 TM data [6
illustrates the seasonal number of Landsat ARD granules over the CONUS for each sensor over the 36-year study period. More ARD granules occur typically in the summer followed by the spring, fall and winter seasons. The cause of this seasonality is not immediately obvious as, nominally, every Landsat 5 TM and Landsat 7 ETM+ overpass of the CONUS is acquired. It is due to the ARD geolocation accuracy requirement and CONUS cloud seasonality. Only Landsat images that can be geometrically corrected to a geodetic accuracy ≤12 m RMSE are used to generate the ARD [2
]. At the time of Landsat overpass the CONUS is on average cloudier in the winter and progressively less cloudy in the fall, spring and summer [8
]. In the cloudy seasons there are a reduced relative number of ground control points used in the Landsat geolocation [24
] and so fewer ARD granules. There are some exceptions to this seasonal pattern because of Landsat sensor and/or acquisition abnormalities. For example, the May 2003 Landsat 7 ETM+ Scan Line Corrector (SLC) failure [26
] resulted in fewer Landsat acquisitions in the following 2003 summer months.
summarizes, for each sensor, the dates of the first and last days with one or more CONUS ARD granules and the sensor lifetime
and number of sensing days,
over the 36-year study period. The Landsat 5 TM provided the longest operating Earth remote sensing satellite mission in history [27
] with a 28.156 year sensor lifetime
. In Table 1
Landsat 7 ETM+ had an 18.518 year sensor lifetime
but continued to sense beyond the end of the 31 December 2017 study period. Landsat 4 TM had a 10.685 sensor life time
but as noted above, was not always operational. The discontinuous acquisition and availability of Landsat data over the CONUS, particularly for Landsat 4 TM, is evident by the difference between the sensor lifetime
and the number of sensing days
. In particular, the Landsat 4 TM sensor was in orbit for 3,900 days but acquired CONUS data to make ARD for only 387 days, that is, for only 9.92% of the sensor lifetime
. In contrast, the Landsat 5 TM and 7 ETM+ sensors acquired data to make ARD for 93.17% and 98.15% of their sensor lifetime
values respectively. Over the 36 year study period there were 12,191 days (>33 years) when there was at least one sensor acquiring CONUS data that were used to make ARD.
4.2. Pixel-Level Summary Information
shows, for the three sensors together, the total number of (a) non-fill observations, (b) non-fill and non-cloudy observations and (c) non-fill, non-cloudy and non-shadow observations, at each CONUS ARD 30 m pixel location over the 36 year study period. Many of the CONUS border ARD tiles contain pixels with
= 0 (shown as black). Discarding pixel locations with
= 0, the mean CONUS total number of observations over the 36 year study period derived from Figure 4
a–c is 936.2, 658.4 and 628.0 respectively, demonstrating the effect of cloud and shadow in reducing the number of surface observations.
Across the CONUS there is a regional variation in the number of observations. Notably, without consideration of the cloud or shadow status (Figure 4
a) there are fewer observations in the north-east, north-west and also in certain ARD tiles that are predominantly over water. This is due to the ARD geolocation accuracy requirement – only Landsat images that can be geometrically corrected to a geodetic accuracy ≤12 m RMSE are used to generate the ARD [2
]. Consequently, in coastal and cloudy regions, which typically have a reduced relative number of ground control points used in the Landsat geolocation [24
], there are fewer ARD granules.
b shows the total number of non-fill non-cloudy observations, that is, it shows the same information as Figure 4
a but without the cloudy observations. There are fewer observations due to cloud cover at the time of Landsat overpass, which previous CONUS Landsat cloud studies have shown predominates in the north-east and north-west, with more observations in the arid south-west [4
]. The ARD cloud mask is imperfect and even at national scale, some cloud detection issues are evident. For example, the extensive highly reflective sands of White Sands, New Mexico, have a
reduced number of observations (blue spot near the center of ARD tile h10v14) which is a previously documented Landsat cloud mask commission error [12
c shows the total number of non-fill, non-cloudy and non-shadow observations, that is, it shows the same information as Figure 4
b but without the shadow observations. There are fewer observations due to shadows across the CONUS. This is particularly evident over mountainous regions, for example, over the Rocky mountains, from Colorado to Idaho, where shadows due to relief and not just clouds are common. Pixels were considered shadow contaminated if they were labelled as Cloud Shadow in the ARD PIXELQA band [2
]. However, at local scale, the ARD shadow mask can be quite unreliable. For example, we have found examples where it fails to mask shadows or falsely labels non-shadow observations as shadow. Efforts by other researchers are underway to improve the shadow mask [28
]. Therefore, in the remainder of the paper we do not consider the shadow state as it is not reliably labelled in the current ARD version.
The approximately north-south oriented stripes in Figure 4
occur because spatially adjacent Landsat orbits overlap in the across-scan direction, with a greater number of observations (red tones) evident in the overlap regions. The overlap regions are wider in the north than in the south because the Landsat orbit paths converge further poleward [4
]. The ARD tiles are defined in the Albers equal area projection. Consequently, the change in the relative orientation of the ARD tile boundaries to the fixed Landsat orbit inclination is evident, with an approximately top to bottom relative orientation in the eastern CONUS and a more diagonal relative orientation in the western CONUS. As a result, many tiles in the western CONUS contain a larger number of overlapping Landsat observations than in the eastern CONUS.
The Landsat orbits drift slightly at annual scale, for example, considering three years of global Landsat 5 TM and Landsat 7 ETM+ data the mean drift was several kilometers east-west and less than a kilometer north-south direction [4
]. Over the satellite lifetime, the Landsat orbits are maintained by station keeping maneuvers (orbit burns) but for Landsat 5, the orbit was allowed to drift considerably when the satellite was operated commercially from 1985 to 2001 [13
]. The orbits drift spatially, which blurs the number of observations over long time periods and this is apparent when the Figure 4
data are examined in detail. As an example, Figure 5
shows a spatial subset of Figure 4
a for six ARD tiles.
shows the average annual numbers of non-cloudy and non-fill observations (
, Equation (1)) at each 30 m ARD pixel location. The same geographic patterns as Figure 4
b are apparent, including the stripes of greater observations where spatially adjacent Landsat orbits overlap.
As described above, nominally, there are a maximum of 22 or 23 Landsat overpasses per year per sensor and much of the less cloudy CONUS regions have
similar to this maximum. The
values are often smaller than the
values, particularly in the across-scan orbit overlap regions. This is because of the May 2003 Landsat 7 ETM+ Scan Line Corrector (SLC) failure that resulted in a 22% data loss occurring in a wedge-shaped pattern increasing in width at scan angles further from nadir [26
have the smallest values (Figure 6
d) and even with 387 sensing days
), full CONUS Landsat 4 TM coverage did not occur. In Figure 6
d, there are regions with no observations and there are ARD pixels with significantly reduced
values due to clouds. The CONUS-wide mean
values, considering only pixels with
> 0 and pixels in ARD tiles covering at least 50% of the CONUS, are 4.85 (Landsat 4 TM), 16.41 (Landsat 5 TM), 15.03 (Landsat 7 ETM+) and 21.22 (all three Landsat sensors).
shows the interquartile range (
, Equation (2)) of the annual number of non-fill non-cloudy observations at each CONUS ARD 30 m pixel location. It provides a non-parametric measure of the annual variation in the number of non-fill non-cloudy observations among years. It is zero valued (shown grey) if the lower and upper quartiles are the same, for example, if the annual number of non-fill non-cloudy observations is the same in every year. The combined sensor values (IQRL457
, Figure 7
a) are generally greater, that is, there is more variation among years, than for the individual sensors. This is expected because 23 of the 36 years had simultaneously two sensors acquiring data (Figure 2
) and this results in a greater annual variation over the 36 years than over the lifetimes of the individual sensors. The IQRL5
values (Figure 7
c) are generally greater than the IQRL7
values (Figure 7
b). This is because the Landsat 5 TM acquisition strategy was less consistently maintained, particularly in the commercial operational period from 1985 to 2001 [13
]. The IQRL4
values (Figure 7
d) are generally lower than for the other sensors because the annual number of Landsat 4 TM observations are low (Figure 6
) and because for several years there were no or few Landsat 4 TM ARD granules (Figure 2
). The stripes of high interquartile range values evident across the CONUS are expected as they occur where the adjacent orbits overlap (Figure 4
, Figure 5
and Figure 6
) resulting in a greater absolute variation in the number of observations among years.
shows histograms of the temporal difference between consecutive observations of a single ARD pixel location. The results are shown for the most observed CONUS 30 m ARD pixel, that is, the pixel location with the greatest number of non-fill and non-cloudy Landsat 4 TM, 5 TM and 7 ETM+ observations over the 36 year study period. The most observed CONUS ARD pixel location is in a region of southern California where adjacent Landsat orbits overlap at 32°47′34.09″ N, 114°55′47.60″ W (pixel row 3656 and column 2154 in ARD tile h05v13). At this pixel location there were 24, 940 and 570 (total 1534), non-fill non-cloudy Landsat 4 TM, 5 TM and 7 ETM+ observations over the 36 years, respectively. The results illustrate that when considering each sensor independently (Figure 8
b–d), the most common temporal differences between consecutive pixel observations are 7, 9 and 16 days. The 7 and 9 day differences reflect how adjacent orbits laterally overlap at the Landsat swath edges and the 16 day difference is due to the 16 day Landsat repeat cycle. Other temporal differences are due to cloud obscuration between consecutive observations and periods of missing observations due to sensor and/or acquisition abnormalities. The maximum consecutive temporal differences (not plotted in Figure 8
) are 57, 176 and 1975 for the Landsat 7 ETM+, 5 TM and 4 TM observations respectively. Similar results are evident considering the three sensors combined (Figure 8
a) but with a greater variety of differences and notably one and eight day differences that can occur between successive Landsat 7 ETM+ and 5 TM observations [29
The results in Figure 8
are for the single most observed non-fill non-cloudy CONUS ARD pixel. At other ARD pixel locations greater temporal differences between consecutive non-cloudy Landsat observations occur more frequently. These differences are not straightforward to summarize at every CONUS ARD pixel because they vary considerably. In certain years (Figure 2
) and in the non-summer months (Figure 3
), there may be relatively fewer observations of an ARD pixel and at some pixel locations there may be no or only a small number of non-fill non-cloudy observations over the sensor lifetime (Figure 6
4.3. ARD Tile-Level Summary Information
shows the per tile annual mean number of non-fill non-cloudy observations (
, Equation (3)). Considering each sensor separately and all three sensors combined, reveals similar patterns. The most frequently observed tiles are in the south-west, particularly over the Mojave and Sonora deserts and also in mountain rain shadow regions of the Cascade Range (Columbia basin and High Desert), Sierra Nevada Range (Great Basin Desert), Coast Range (Central Valley) and the Rocky Mountains (Great Plains). The least observed tiles tend to be in the north-east, around the Great Lakes and along parts of the north-west coast. As noted above, the coastal ARD tiles have low
values as there are fewer ARD granules for the geolocation processing reasons discussed at the beginning of Section 4.2
In Figure 9
the pixel-level Landsat orbit swath overlap patterns (Figure 4
and Figure 6
) are less evident because of the tile level averaging. However, the change in the relative orientation of the ARD tile locations to the Landsat orbits is still apparent, with more tiles in the western CONUS having greater
values than in the eastern CONUS due to this phenomenon and also due to reduced cloudiness.
shows histograms of the Figure 9
data. The histograms are approximately normally distributed, except for the Landsat 4 TM histogram that is positively skewed (i.e., a greater frequency of small values). Across the CONUS, the most common
values are 18 to 20 (23.4% of tiles) for all three sensors, 12 to 14 (30.7% of tiles) for the Landsat 7 ETM+, 14 to 16 (26.2% of tiles) for the Landsat 5 TM and 2 to 4 (36.5% of tiles) for the Landsat 4 TM.
The ten CONUS ARD tiles with the highest
values are summarized in Table 2
and highlighted in Figure 9
. The tiles are clustered over the south-west and with the exception of Landsat 4 TM, have
> 20. Considering Landsat 7 ETM+, Landsat 5 TM and the three sensors combined, eight of the ten most frequency observed ARD tiles are common. Southern California ARD tiles h05v13 (centered 33°06′56.45″ N, 114°53′38.52″ W) and h04v11 (centered 35°27′32.82″ N, 117°11′13.44″ W) have the highest
values for these sensors. The top ten Landsat 4 TM ARD tiles have a different geographic distribution compared to the other sensors, due to the spatially irregular number of observations (Figure 6
). Southern Arizona ARD tile h06v13 (centered 33°22′12.09″ N, 113°18′11.39″ W) has the highest
value for the Landsat 4 TM.
The ten CONUS ARD tiles with the lowest
values are summarized in Table 3
and highlighted in Figure 9
. They are located predominantly in north-west coastal regions and in tiles surrounding the Great Lakes. For all three sensors and also Landsat 5 TM, the least observed tile is h28v04 that encompasses northern New York state and parts of southern Canada (centered 44°58′01.15″ N, 74°06′31.86″ W). For Landsat 7 ETM+, the least observed tile is h25v07 that is over the state boundary of Ohio and Pennsylvania (centered 41°47′42.71″ N, 80°40′9.54″ W). Notably, these least observed tiles have approximately half the
values compared to the most observed tile
values for these sensors (Table 2
). The least observed Landsat 4 TM ARD tile is h22v08 that is over northern Indiana (centered 40°59′40.72″ N, 86°17′1.36″ W) and has a very low
(0.526) that is, in this tile ARD pixels have on average less than one non-fill non-cloudy observation per year.