Radiometric Characteristics of the Landsat Collection 1 Dataset

This study evaluates the long-term radiometric performance of the USGS new released Landsat Collection 1 archive, including the absolute calibration of each Landsat sensor as well as the relative cross-calibration among the four most popular Landsat sensors. A total of 920 Landsat Collection 1 scenes were evaluated against the corresponding Pre-Collection images over a Pseudo-Invariant Site, Railroad Valley Playa Nevada, United States (RVPN). The radiometric performance of the six Landsat solar reflective bands, in terms of both Digital Numbers (DNs) and at-sensor Top of Atmosphere (TOA) reflectance, on the sensor cross-calibration was examined. Results show that absolute radiometric calibration at DNs level was applied to the Landsat-4 and -5 TM (L4 TM and L5 TM) by −1.119% to 0.126%. For L4 TM and L5 TM, the cross-calibration decreased the radiometric measurement level by rescaling at-sensor radiance to DN values. The radiometric changes, −0.77% for L4 TM, 0.95% for L5 TM, −0.26% for L7 ETM+, and −0.01% for L8 OLI, were detected during the cross-calibration stage of converting DNs into TOA reflectance. This study has also indicated that the long-term radiometric performance for the Landsat Collection 1 archive is promising. Supports of these conclusions were demonstrated through the time-series analysis based on the Landsat Collection 1 image stack. Nevertheless, the radiometric changes across the four Landsat sensors raised concerns of the previous Landsat Pre-Collection based results. We suggest that Landsat users should pay attention to differences in results from Pre-Collection and Collection 1 time-series data sets.


Introduction
The Landsat program provides the longest continuous climate records from space. The trustworthy information preserved by the Landsat images has improved the researches of global resilience to climate change and variability. However, the inconsistency of Landsat radiometric measurements imposes the scientific values of the Landsat images. In 2016, the United States Geological Survey (USGS) started reorganizing the 40-year multi-sensor Landsat archive into a formal Collection structure. The new collection management strategy has been implemented to form the Landsat Collection 1, which was released in 2017. All Landsat data are cross-calibrated (regardless of sensor) across the full collection (Landsat Collections, https://landsat.usgs.gov/landsat-collections). The essential goal of the new Landsat product is to provide a consistent Landsat archive with improved geometric and radiometric quality. The new high quality Landsat images will finally offer the users community a Climate Data Records (CDRs) suitable for time-series analyses. Consistent radiometric measurements enable producing sustainable and scientifically defensible CDRs for environmental remote sensing studies.
The radiometric calibration procedures for the Landsat Collection Tiers represent a significant change in the implementation of creating radiometric consistent Landsat measurements. The results from this study reveal the radiometric changes on how, where and to what extent of the Landsat Collection 1 compare to Pre-Collection. The purpose of this paper is to report a preliminary study of the radiometric characteristics of the USGS latest released Landsat Collection 1 data. Following this section, this paper provides a brief review of the radiometric calibration history of the L4 TM, L5 TM, L7 ETM+, and L8 OLI and the efforts of cross-calibration between the Landsat sensors, gives the details regarding the comparison of the radiometric characteristics between Landsat Collection 1 and Pre-collection data, discusses the scientific improvements of the first radiometric consistent Landsat CDRs, and closes with a summary of the performance using Landsat Collection 1 archive for time-series analysis.

Brief Review of the Landsat Radiometric Calibration
The radiometric calibration of the L4 TM was based on the internal calibrator  [1]. For each detector, the least squares analysis was used to estimate the gain and bias against the IC lamp data. The linear coefficients were further applied to remove the residual striping [1]. However, Fischel found that detector bias of L4 TM was not constant when the scan-to-scan periods longer than scene acquisition times [2]. He developed an alternative algorithm, using the shutter data to estimate the bias and lamps to estimate the gains. The radiometric correction of L4 TM was dramatically improved using the alternative algorithm [2].
In follow-on studies, the radiometric calibration of the L4 TM were tied to L5  [8]. Nevertheless, due to the observed instrument's IC degradation a relative gain approach (lifetime gain model) was developed and implemented to all the USGS distributed L5 TM imagery [8]. Relative gain is radiometric gain of each detector relative to other detectors in the same solar reflective band. The change of relative gain could be described as a linear function of time for the L5 TM detectors. The development of the new L5 TM radiometric calibration procedure was based on period observations (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003) and was anchored to that of L7 ETM+. Helder et al. [9] thoroughly studied the L5 TM lifetime radiometric stability and concluded that the relative gain approach for the L5 TM solar reflective bands is feasible and improved. By using the relative gain model, the L5 TM radiometric calibration overcome the drawbacks of traditional histogram equalization approach, without the related degradation effects, and was tied to the cross-calibration with the L7 ETM+ [9]- [14].
Immediately after the L5 launch, Metzler and Malila started a study of cross-calibration from L4 and L5 TM [15]. The data used for this first L4 and L5 TM cross-calibration efforts were acquired during an underpass maneuver. Direct comparison of the relative radiometric responses was done between the respective solar reflective bands from L4 and L5 TM. They found that the multiplicative factors range from 0.987 to 1.145 were required to convert L4 to L5 TM data.
In addition to the lamp-based IC, the L7 ETM+ has two new on-board calibration devices for the solar reflective bands: a Full Aperture Solar Calibrator (FASC) and a Partial Aperture Solar Calibrator (PASC). Unfortunately, Markham et al. found that the PASC was unreliable [16]. The on-board calibrator lamps and diffusers indicate that the L7 ETM+ drift is generally by no more than 0.5% per year. In order to maintain the L7 ETM+ calibration accuracy and reduce the degradation effects from the on-board calibration system, the vicarious calibration approach has frequently been used to calibration the L7 ETM+. Studies show that the on-board calibrators and vicarious calibration have kept the

Landsat Image Stack
We ordered all available Landsat

Selected Landsat Scenes
Four cloud-free Landsat observations corresponding the four Landsat instruments (L4 TM, L5 TM, L7 ETM+, and L8 OLI) of the RVPN (P040/R033) were selected from the time-series Landsat image stack to examine the changes between the Landsat Collection 1 and Pre-Collection (Table 1). For purposes of comparison, 100 random points were spread across the Railroad Valley Playa, NV. The random points locations were constrained by following the procedures reported by Helder et al. [6]. The pixel values of the DN, TOA reflectance, and surface reflectance of each point were extracted from both the processed Landsat Collection 1 and Pre-Collection data.

Results
The cross-calibration changes between the Landsat Collection 1 and Pre-Collection were examined in three levels of the Landsat products (i.e. DNs, TOA reflectance, and surface reflectance).

Quantized Digital Numbers (DNs)
The quantized Landsat pixel values (Q, or digital counts) in the raw Landsat data (Level 0R or 1R) were converted to at-sensor spectral radiance (L_λ). The L_λ  (1)).

Collection1
Pre-Collection change Pre-Collection The DN values extracted from the four selected Landsat scenes and the differences between the Landsat Collection 1 and Pre-Collection are presented in Table 2. It indicated that no change had occurred in all the solar reflective bands of the Landsat Collection 1 for L7 ETM+, with less than 0.01% change in L8 OLI data, larger changes have been captured in every single solar reflective band of the L4 TM and L5 TM images (

Top-of-Atmosphere (TOA) Reflectance
As mentioned above, the DNs of Landsat scenes were converted to at-sensor spectral radiance and TOA reflectance, atmosphere correction was further applied to retrieve the Landsat surface reflectance. Conversion from the DNs  There is almost no change for the L8 OLI (less than −0.01% in Figure 1). The results indicates that the radiometric calibration of the three older Landsat sensor (i.e. L4 TM, L5 TM, and L7 ETM+) have been anchored to that of the L8 OLI.  The cross-calibration procedures applied for the Landsat Collection 1 dataset finally influences the performance of its surface reflectance. The surface reflectance generated from the USGS ESPA system has been breakdown into sensors and bands in order to reveal where the differences located. Figure

Normalized Difference Vegetation Index
When apply above changes to the terrestrial remote sensing by using the popular NDVI, the shifts of the NDVI distribution of the three older Landsat sensors (i.e. L4 TM, L5 TM, and L7 ETM+) draw concerns with the previous studies based on the Landsat Pre-Collection data. Figure 3 indicated that the NDVI have been overestimated based on the Landsat Pre-Collection data, especially for the three older Landsat sensors. Based on the Landsat Pre-Collection data, Roy et al. reported that both TOA and surface reflectance derived L7 ETM+ NDVI values are greater than that derived from the L8 OLI [31]. They found that the radiometry difference between L8 OIL and L7 ETM+ in NIR band even influence derived NDVI in 9.88%, though the differences could be reduced as 4.86% when atmospheric corrections are applied. Be noted that the area of interest (AOI) of this study is covered by homogeneous desert surface, and the scattered shrubs don't dominate the statistic distribution of the calculated NDVI. Figure    Landsat images to representative range of reflectance spectra (i.e., capturing land cover, land use, vegetation phenology and soil moisture variations) [32].

Long-Term Trend of the Landsat Collection 1
The essential goal of the new released Landsat Collection 1 dataset is to provide a radiometric consistent archive across the Landsat sensors to support time-series analysis and data stacking (https://landsat.usgs.gov/landsat-collections). Accurate radiometric calibration (consistent radiometric cross-calibration) is a critical step in developing Landsat time-series analysis ready data (ARD) with high quality to perform quantitative remote sensing. Figure 4 shows the temporal evolution of the changes of these cross-calibrations. The long-term trends showed in Figure 4 are fairly similar to both TOA and surface reflectance. It is expected that the cross-calibration for the sensors, tends to eliminate the long-term radiometric shift trend and therefore becomes more sensitive to the earth surface changes. On the other hand, the TOA reflectance changes keep the memory of the seasonal trend in each solar reflective band (left panel of Figure   4).   Even for the blue band of L8 OLI, its cross-calibrated data distribution shifts to the lower part (lower panel of Figure 5).

Conclusions
The Landsat Collection  the four Landsat sensors. Results from this study lead to two key conclusions.
The first is that cross-calibration has been implemented over the whole Landsat Collection 1 process stages. This implies two things. The first is that new absolute calibration procedures have been applied for the L4 TM and L5 TM observations. The second is that relative cross-calibration procedures have been applied to align the radiometric measurements of the four Landsat sensors during the converting DNs into TOA radiance stage. The second conclusion is that the Landsat Collection 1 radiometric reflectance has been anchored to the L8 OLI. proved radiometric calibration quality, it has achieved the best ever agreement, in terms of the on-board and vicarious calibration approaches, than the other Landsat sensors [33]. It is suitable to tie the radiometry of the four Landsat workhorse sensors with the L8 OLI.
The cross-calibration between the Landsat sensors, like L4 TM and L5 TM, L4 TM and L7 ETM, L5 TM and L7 ETM+, have been thoroughly studied and documented in past decades. This study indicates that the first four-sensor cross-calibration effort (Landsat Collection 1) is very promising. However, we suggest that Landsat users should pay attention to differences in results from Pre-Collection and Collection 1 time-series data sets.
Ultimately, the USGS new released Landsat Collection 1 provides the best ever radiometric consistent product across the four most popular Landsat sensors (i.e. L4 TM, L5 TM, L7 ETM+, and L8 OLI). It made the longest time-series CDRs possible for regional and global climate change, ecological, land-cover and land-use changes (LCLUC), and environmental remote sensing studies.

Conflicts of Interest
The authors declare no conflicts of interest regarding the publication of this paper.