For over a half-century, the Landsat program, a series of joint agency Earth observing satellite missions between the National Aeronautics and Space Administration (NASA) and the U.S. Geological Survey (USGS), has provided high-quality global land and nearshore coastal observations from a suite of medium-resolution imaging satellites (Wulder et al., 2022; Crawford et al., 2023). Upon the adoption of a collection-based archive processing and management approach in 2016 (Dwyer et al., 2018; Crawford et al., 2023), Landsat data are systematically processed, archived, and distributed by the USGS Earth Resources Observation and Science (EROS) Center located in Sioux Falls, South Dakota, USA. Through collaboration with remote sensing subject matter experts and participation from the Landsat Science Team, USGS EROS has developed and operationalized research-quality Level 1 Top of Atmosphere (TOA) calibrated reflectance and Level 2 atmospherically corrected surface reflectance and surface temperature products that can be used to map, monitor, assess, and interpret how Earth's surface has changed as a result of human influence and natural environmental conditions. These open access data products from Landsat are made publicly available at no cost (Zhu et al., 2019) through the USGS EROS Earth Explorer (EE) data portal and Machine-to-Machine (M2M) Application Programming Interface (API). USGS also offers direct access to Landsat data through the Amazon Web Services (AWS) commercial cloud environment in a “Requester Pays” (user incurs cost for data requests and downloads) bucket configuration (Crawford et al., 2023). This allows researchers, scientists, U.S. federal and state agencies, and international organizations to utilize Landsat data products for their science applications, and to facilitate informed land, natural resources, and water management decisions and policies (Wulder et al., 2019).

Landsat Level 2 science product development follows a structured process that involves iterative collaboration between principal investigator(s) (e.g., a Landsat Science Team member or a U.S. federal agency scientist) and the USGS Landsat science project to operationalize mature science algorithms. The development phases of this process (discussed in Sect. 2) include research, provisional, and operational readiness levels for the generation of science data products. Products that are considered provisional are available to the public through the EROS Science Processing Architecture (ESPA; https://espa.cr.usgs.gov, last access: 1 January 2026) on-demand interface but are actively under USGS internal evaluation and remote sensing community validation. These algorithms and the resulting product layers may undergo further modifications or improvements before being considered for operational release.

Although Landsat missions have primarily been designed for observing and monitoring land change, Landsat 8 (launched February 2013) and Landsat 9 (launched September 2021) have been used extensively for aquatic remote sensing applications (Tyler et al., 2022) due to the Operational Land Imager (OLI)'s substantial improvements in both radiometric data quality and spectral resolution compared to heritage Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) instruments (Roy et al., 2014; Pahlevan et al., 2014; Concha and Schott, 2016; Olmanson et al., 2016). Compensating for the intervening effects of atmospheric scattering and absorption between the sun, surface, and remote imaging sensor, which vary spatially and temporally, is a necessary processing step to enable reliable monitoring, characterization, and interpretation of the Earth's surface (Vermote and Kotchenova, 2008; Korkin and Lyapustin, 2023; Thompson et al., 2019a; Thompson et al., 2022; Pahlevan et al., 2017). In contrast to brighter terrestrial land surfaces, retrieving atmospherically corrected spectral reflectance information from dark aquatic targets using spaceborne imaging sensors is a major challenge because the attenuated sunlight reflected from the water is usually only a fraction of the total signal received at the top of atmosphere (Wang, 2010).

In April 2020, USGS EROS introduced a Level 2 provisional Aquatic Reflectance (AR) product for Landsat 8 OLI observations, marking the initial phase in developing a standardized global product for Landsat-derived surface water measurements. The algorithm to generate AR products for Landsat 8 (and Landsat 9 since launch in September 2021) OLI imagery was adopted from version 8.10.3 of the Level 2 Generation (l2gen) module within the SeaWiFS Data Analysis System (SeaDAS), originally developed by the NASA Ocean Biology Processing Group (OBPG). This software has been the standard processing method for several previous and ongoing NASA ocean color missions like the Coastal Zone Color Scanner (CZCS, 1978–1986), the Medium Resolution Imaging Spectrometer (MERIS, 2002–2012), the Geostationary Ocean Color Imager (GOCI, 2010–2021), the Moderate Resolution Imaging Spectroradiometer Aqua (MODIS Aqua, 2002–present), and the Visible Infrared Imaging Radiometer Suite (VIIRS, 2011–present) (Mobley et al., 2016). USGS Level 2 provisional AR products have been available to process and download from the USGS ESPA on-demand interface. These products underwent a refresh in 2022 following the release of Landsat Collection 2 and contain Level 2 AR for the visible to near-infrared (VNIR) spectral bands (OLI bands 1–5) (Fig. 1), intermediate Rayleigh-corrected reflectance (ρrc) for the visible to shortwave infrared (VSWIR) spectral bands (OLI bands 1–7), and other supporting data layers. These provisional AR products are intended for immediate, experimental use by the remote sensing community involved in water quality monitoring, seafloor classification, satellite derived bathymetry, and other surface water mapping applications so that community assessment of their suitability can be used to strengthen AR retrieval performance to operational readiness in support of applications requiring high quality measurements. Water quality surveying groups like the USGS Water Mission Area already rely on Landsat and Sentinel-2 observations to monitor U.S. national waters (Fickas et al., 2023; Stengel et al., 2023; Meyer et al., 2024), emphasizing the need for operationally generated satellite-derived data in enabling comprehensive and consistent water resource management and assessments.

Satellite-derived AR measurements are a critical asset where in situ data are scarce or costly to collect. Feedback from science applications end users ensures that data outputs are both robust and actionable, fostering trust and reliability across scientific, policy, and operational domains. The goal of USGS EROS aquatic product research and development is to enable emerging Landsat aquatic science applications and prepare for an operational processing architecture for Landsat Collection 3 in the late 2020s. The purpose of this paper is to communicate with aquatic scientists, satellite oceanographers, and the broader Earth observation community on the origins, requirements, challenges, successes, and future objectives for operationalizing global AR data products for Landsat satellite missions.

Remote sensing reflectance (Rrs) is defined as the ratio of the spectral distribution of reflected solar radiation upwelling from just beneath the water surface (Lw, W m⁻² sr⁻¹) normalized by the downwelling solar irradiance (Ed, W m⁻²) in the visible to near-infrared domain (λ= 400–900 nm, unit: steradian-1) (Lee et al., 1997; Gordon and Wang, 1994; Mobley, 1999): 1Rrs(λ)=LwλEdλ(sr-1), Rrs is the conventional measurement used in proximal, airborne, and satellite-based remote sensing to quantify the optically active, biogeochemical constituents (i.e., chlorophyll, total suspended solids, dissolved organic matter) (O'Reilly et al., 1998; Lee et al., 2001; Mishra and Mishra, 2012; Dogliotti et al., 2015) and is an essential component for the water quality analysis of lakes (Lehmann et al., 2018; Giardino et al., 2019), long term ocean color monitoring programs (Werdell et al., 2007), benthic mapping practices (Louchard et al., 2003; Dierssen et al., 2010), and optical water type classification for global water bodies (Spyrakos et al., 2018; Bi and Hieronymi, 2024).

SeaDAS, developed and maintained by NASA's OBPG, is the satellite image preprocessing software for generating aquatic Rrs image products for several ocean color missions primarily associated with global monitoring programs for over 25 years (Mobley et al., 2016). Because of this, the open source code for l2gen supports several multispectral (and hyperspectral) Earth Observation missions, including the OLI instruments onboard Landsat 8 and Landsat 9. The adaptation of l2gen processing for use with Landsat OLI data is described by Franz et al. (2015), with additional regional analyses of the impact of band selection for aerosol estimation provided by Vanhellemont et al. (2014) and Pahlevan et al. (2017).

The l2gen processing code within SeaDAS computes the Rrs for each band at each identified water pixel from the Level 1 at-sensor radiance Lt, which is assumed to be partitioned linearly into distinct physical contributions as shown below: 2Lt(λ)=[Lr(λ)+La(λ)+tdvλLwc(λ)+tdvλLw(λ)]tgv(λ)tgs(λ)fp(λ),

Lr(λ)= the radiance contribution due to Rayleigh scattering by air molecules

The l2gen atmospheric correction algorithm retrieves the water-leaving radiance Lw component of interest by estimating and subtracting the terms on the right-hand side of Eq. (2) from Lt. Of these components, the estimation of the aerosol scattering contribution La is generally the most challenging and impactful for the retrieval of Lw (outside of glint-contaminated areas, that is). While the l2gen software accepts a wide variety of processing options for aerosol radiance estimation, the parameterization most commonly used in the operational processing of supported mission data makes use of an iterative bio-optical model to satisfy a fundamental assumption of the algorithmic approach: that near-infrared water-leaving radiance is either negligible or can be accurately estimated (Bailey et al., 2010). With this assumption, the aerosol radiance in each band can be estimated via the two-band aerosol selection approach of Gordon and Wang (1994). USGS provisional AR processing uses OLI band 5 (865 nm) and band 6 (1609 nm) as the choice of bands, following the recommendation of Pahlevan et al. (2017). The value of Rrs(λ) is then computed as: 3Rrs(λ)=Lw(λ)F0(λ)fscos⁡(θs)t(θ), where: F0= extraterrestrial solar irradiance (Thuillier et al., 2003), Fs= adjustment of F0 for variation in Earth-Sun distance, and t= diffuse transmittance.

The spectral Rrs bands (in steradian) are normalized (multiplied by π) to produce dimensionless aquatic reflectance (Franz et al., 2007; Franz et al., 2015; Mobley et al., 2016): 4AquaticReflectanceARλ=Rrs(λ)⋅π, Additional details, including the full set of processing parameters used in the generation of the provisional AR products, can be found in USGS documentation (USGS, 2024).

Due to its interoperability, traceability, and availability, the l2gen algorithm in SeaDAS (SeaDAS l2gen 8.10.3) was adopted by the USGS into the EROS's Science Algorithms to Operations (SATO) process in 2018, as a baseline for developing an atmospheric correction pathway for Landsat AR. The SATO Product Maturity Matrix for USGS Landsat science products is the formal description of the development process used by USGS EROS to mature algorithms for collection processing in the operational Landsat Product Generation System (LPGS). The purpose of SATO is to enable a smooth transition of researched, developed, and matured science algorithms and prototype executables into a formally developed and maintained LPGS operational environment. The product maturity matrix for provisional Landsat science products is adopted from the National Oceanic and Atmospheric Administration (NOAA) Climate Data Record (CDR) maturity model (Bates and Privette, 2012) and is used as the template to transition select candidate science algorithms through the SATO process (Table 1).

The progression and transformation of the product follow a structured procedure, with milestones and responsibilities agreed on between the USGS Landsat science project and the algorithm principal investigator(s). Work is divided into a series of sequential phases, as follows:

Research Stage (Maturity Levels 1 and 2). During this stage, academic researchers and principal investigators lead the process. The product remains publicly restricted until it is published, because significant changes to the source code are expected. Meanwhile, principal investigators submit peer-reviewed journal articles describing the algorithmic approach.

Throughout a product's provisional lifetime, modifications to its features are expected, although the underlying algorithm to generate the product (e.g., aquatic reflectance) is unchanged. For example, algorithm ingestion into ESPA often involves modifying source code for greater processing efficiency as well as for reproducibility. Science verification at each step is conducted to ensure no anomalies are detected in the data and that any alterations or updates to the source code do not have a direct impact on the algorithm itself. Metadata standards are used to ensure product attributes are an accurate representation of the data, are understandable, and can be referenced. After verification and quality checks, the data product is released through the ESPA on-demand interface for public availability along with documentation and any known caveats published on the USGS product web page. Provisional data products are generated to enable timely scientific use and garner user feedback on quality, algorithm performance, observed uncertainties over diverse geographical regions, and community validation following early adopter feedback. It is the responsibility of USGS EROS to compile this information from the community, work with corresponding research groups, and routinely assess other candidate algorithms with potential principal investigators.

Since their release to the public in 2020, order requests for the Landsat 8/9 Level 2 provisional AR products from ESPA by the community have now surpassed 90 000 scene downloads as of the end of 30 September 2024 (Fig. 2). Maximum downloads were observed during the first year of release (and the re-release, following the availability of Collection 2), followed by downward trends with each passing fiscal year. The release of Landsat 8/9 provisional AR products allowed the opportunity to gain insights from the scientific user community on the quality and accuracy of the products. Examples of product feedback include research articles and agency reports that evaluate provisional Landsat AR products across a variety of aquatic scientific applications, including coastal ocean color mapping (Nazeer et al., 2020; Tavora et al., 2023), lake water quality monitoring (Ogashawara et al., 2020; Niroumand-Jadidi et al., 2022), and satellite-derived bathymetry (Poppenga and Danielson, 2021).

Landsat 8/9 provisional AR product limitations were recognized by the scientific community concerning (1) the omission of valid water pixels associated with the l2gen-based land/water delineation and (2) negative AR values generated primarily over inland and optically complex coastal waters (Pahlevan et al., 2019; Ilori et al., 2019; Ogashawara et al., 2020; Tavora et al., 2023). While a new water masking approach was developed for the re-release of the provisional products associated with Collection 2 to mitigate the inconsistencies associated with the l2gen-based land/water delineation, the negative values resulting from atmospheric correction remain a challenge that has been well documented in the literature across a suite of ocean colour applications (Ruddick et al., 2000; Mélin et al., 2011; Bramich et al., 2018; Wei et al., 2018; Kuhn et al., 2019; Pahlevan et al., 2021). Negative AR, which can significantly affect the accuracy of downstream water quality products, has been primarily attributed to the challenges of utilizing one or more NIR spectral bands to characterize aerosol path radiance(s) (La) over highly turbid or productive, complex case-2 type waters (Bailey et al., 2010; Werdell et al., 2010; Dash et al., 2012; Ibrahim et al., 2019; Wang et al., 2022). In these optically challenging water bodies, the traditional assumption that water-leaving radiance in the NIR portion of the electromagnetic spectrum is negligible (or effectively estimated by the assumptions of the algorithm) is not valid. Instead, such algorithms may underestimate the substantial water-leaving NIR contribution in highly turbid or productive waters, leading to overestimation of La and, consequently, dragging the downstream AR to low and even negative values (Fig. 3). This issue is intensified for inland freshwater systems, which contain varying amounts of coloured dissolved organic matter, suspended sediments, phytoplankton, and surrounding land pixels bordering the entire lake shoreline. Accurate aerosol correction in such environments is crucial for reliable water quality assessments, and addressing these limitations will be decisive for the success of Landsat AR products in future Collections. Other challenges faced by SeaDAS (and many other algorithms designed for ocean colour) include factors such as mitigating sun glint and a missing correction for adjacency effects. Increasing user awareness of these issues may explain the observed downward trend in USGS provisional AR product downloads over time. In response, the provisional product package updates that followed the release of Landsat Collection 2 also augmented the suite of data layers to include AR for the NIR band, per-pixel angle bands, intermediate auxiliary input data and Rayleigh-corrected reflectance products so that users would have supplementary information to further investigate instances when and where full atmospheric correction fails (Table 2). However, these issues must be more fully addressed for the AR product to reach operational maturity. Concurrently, comprehensive aquatic-based atmospheric correction research and applications published by a variety of authors and institutions have provided alternative approaches that may be better suited to compensate for aerosols in the atmosphere over complex water targets (Steinmetz et al., 2011; Brockmann et al., 2016; Moses et al., 2017; De Keukelaere et al., 2018; Vanhellemont, 2019); consequently, some users could be performing their own processing on Level-1 Landsat data using these alternative approaches rather than relying on the provisional AR products from USGS EROS.

The performance of the Landsat 8/9 Collection 2 Level 2 provisional AR Rrs products was evaluated using in situ Rrs measurements from AERONET-OC (naeronet=418) and GLORIA (ngloria=554) matchups against a selection of comparison metrics described in Sect. 4.2. For the AERONET-OC subset, the AR products exhibited strong agreement with AERONET-OC observations. MAD values were low across all bands, ranging from 0.0006 sr⁻¹ in the red band to 0.0014 sr⁻¹ in the coastal band (Fig. 6/Table 3). Median symmetric accuracy (ε) was below the GCOS 30 % threshold in the blue (27.6 %) and green (19.8 %) bands, while the coastal (40.7 %) and red (33.0 %) bands slightly exceeded this limit. Signed symmetric bias (β) indicated a tendency toward underestimation of the AR products in B1–B3, with the strongest bias observed in the coastal band (-23.1 %). The red band (B4) showed a slight overestimation (β= 6.6 %). In contrast, comparisons with GLORIA revealed substantially higher variation. MAD values ranged from 0.0046 sr⁻¹ (B4) to 0.0064 sr⁻¹ (B1). Values of ε exceeded the GCOS threshold in all bands, ranging from 39.6 % (green) to 68.4 % (coastal). Values of β were strongly negative across all bands (-36.8 % to -62.0 %), indicating consistent underestimation of reflectance values by the AR products relative to GLORIA observations. This generally follows the wavelength trends in the l2gen performance for the OLI sensor seen in the aquatic component of the atmospheric correction intercomparison exercise (ACIX-Aqua) (Pahlevan et al., 2021). The larger MADs seen with the GLORIA comparisons are in part due to the higher frequency of negative values in the provisional AR products over GLORIA-sampled locations. The combined dataset yielded intermediate results. MAD values ranged from 0.0028 sr⁻¹ (B4) to 0.0042 sr⁻¹ (B1). The ε values exceeded the 30 % threshold in all bands except green (29.3 %), with values ranging from 35.0 % (red) to 49.9 % (coastal). The β values remained negative across all bands, with the strongest underestimation in the coastal band (-39.8 %) and the weakest in the red band (-19.9 %).

The per-band scatter plots shown in Fig. (7) provide a closer look into the spread of OLI derived AR Rrs between each of the AERONET-OC and the GLORIA matchup datasets. Most notably, when evaluated against AERONET-OC data, the AR products demonstrated strong linear agreement, particularly in the green (R2= 0.89) and red (R2= 0.91) bands (Table 3). Moderate correlations were observed in the blue (R2= 0.76) and coastal (R2= 0.57) bands, suggesting that the AR products are generally reliable in optically simple environments. In contrast, comparisons with GLORIA revealed very weak correlations across all bands, with R2 values ranging from 0.06 (B1) to 0.29 (B4), primarily due to the substantial amount of negative AR values. The combined dataset reflected this discrepancy, with low R2 values across all bands (0.06–0.35), further emphasizing the limited predictive strength of the AR products in more complex or variable aquatic environments.

The classification of inland waters into varying optical water types is driven by the biogeochemical properties in the water column. Differences between GLORIA in situ Rrs and Landsat 8/9 Level 2 provisional AR highlight how these properties influence the sensitivity of the validation assessment. Specifically, the magnitude of the differences, reflected by ε, can vary dramatically across different water types (Fig. 8). This variability indicates that the performance of the Landsat 8/9 provisional AR retrieval is highly context-dependent – errors are minimal in optically simple waters (e.g., clear to moderately turbid coastal waters) but increase considerably in optically complex waters where factors such as elevated levels of turbidity, chlorophyll concentrations, or coloured dissolved organic matter (CDOM) dominate the water column.

Landsat 8-9 OLI Level 2 Provisional Aquatic Reflectance products can be downloaded on demand through the Earth Resources Observation and Science (EROS) Center's Science Processing Architecture (ESPA) at https://espa.cr.usgs.gov/ (last access: 1 January 2026). The validation subset used in this study can be downloaded from the USGS ScienceBase catalog (10.5066/P14MBBRM, Crawford et al., 2025b).

The development of an operational AR product for Landsat, facilitated by SeaDAS open-source code, provided a global AR processing capability for the Landsat user community. The l2gen code within SeaDAS has been the flagship processor for generating AR products for Landsat 8 and Landsat 9 OLI data, it may not be the most optimal solution as a single global processor for current, heritage (Landsat 4/5 TM Landsat 7 ETM+), and upcoming Landsat missions (Landsat Next) in terms of suitability for emerging science needs that require analysis ready data for both inland and coastal water quality mapping applications. The Landsat 8/9 provisional AR performance has shown promising results in the coastal regions, but its reflectance retrieval limitations for inland waters must be acknowledged. These limitations include challenges related to atmospheric correction processing accuracy and consistency across optically and geographically diverse water conditions. Until in situ validation campaigns are conducted on a routine basis with standard operating procedures that are community-endorsed, the combined GLORIA and AERONET-OC datasets offer an interim validation pathway for assessing the operational readiness of aquatic and/or ocean colour processing algorithms and data products Addressing these limitations will be critical for the success of Landsat AR products in future Collections. The USGS Landsat science project approach for Landsat AR algorithm research and development recognizes the importance of the SATO process and collaboration with established aquatic principal investigators. Promoting and maintaining success criteria for a global Landsat Collection 3 AR product while remaining aware of evolving mission specifications for Landsat Next is essential. Key criteria include maintaining consistency across spatial and temporal domains, ensuring interoperability with similar products from other medium-resolution multispectral and imaging spectroscopy missions (e.g., Sentinel-2, Environmental Mapping and Analysis Program [EnMAP], Copernicus Hyperspectral Imaging Mission for the Environment [CHIME]) (Pinnel et al., 2024; Dierssen et al., 2021), and balancing the trade-offs necessary to achieve optimal performance in varying atmospheric and optical water conditions. Looking ahead, the next research steps in preparing for Landsat Collection 3 AR development involves undertaking open science algorithm intercomparisons and quantitative validation that considers heritage missions and Landsat Next science readiness simultaneously. These efforts will provide a foundation for more comprehensive and reliable AR products, ultimately contributing to enhanced understanding and management of aquatic environments globally.