Conference Agenda

Atmospheric correction (AC) of optical satellite images is essential for quantitative remote sensing applications. The open and free data access to Earth Observation (EO) satellite missions has significantly increased the scientific interest in AC processors. Several approaches have been introduced, involving different radiative transfer models, single or multitemporal images, various algorithms to estimate aerosol properties and water vapour content, constant or diverse aerosol models, and various sources of ancillary data. These methodologies are typically validated independently by developers and users, based on specific sites with available reference data, or by comparing results from other AC processors. To comprehensively investigate all aspects and issues of AC, a benchmark Atmospheric Correction Inter-comparison eXercise (ACIX) was initiated in 2016 in the frame of the Committee on Earth Observation Satellites Working Group on Calibration & Validation (CEOS WGCV). The goal was to compare the state-of-the-art AC processors in a voluntary and open-access initiative to which every AC processor’s developer is invited to participate. The first two exercises (ACIX-I and ACIX-II) investigated the variability of AC performances over diverse atmospheric and land cover conditions using Landsat 8 and Sentinel-2A input data. In the current third implementation of ACIX-III over land, the focus is on imaging spectrometer data, i.e. hyperspectral data. 46 PRISMA (for the years 2021 to 2023) and 44 EnMAP scenes (for the years 2022 and 2023) will be analysed over a set of test sites. These sites have been selected based on the availability of ground-based surface reflectance measurements and flight campaigns data with coincident acquisitions from RadCalNet, HYPERNETS, CHIME-AVIRIS-NG and EnMAP validation campaigns. Aerosol and water vapour retrieval will be validated over AERONET sites. These measurements, both for surface reflectance and aerosol/water vapour validation, cover diverse land cover types, seasons and a global extent. The first ACIX-III workshop was held in ESA/ESRIN (Frascati, Italy) June 2022, where the key points of the implementation protocol were discussed. The submission phase began in February 2024 with nine participating AC processors, and the first results are expected in June 2024. This presentation, will cover the ACIX-III initiative, highlighting its main implementation points and latest status, including initial intercomparison results.

Multisensor remote sensing enhances environmental mapping and monitoring by integrating data from diverse Earth observation (EO) sensors. In this context, harmonized Analysis Ready Data (ARD) cubes, comprising consistently preprocessed imagery from different sensors and organized in a unified pixel grid and tiling scheme, facilitate effective multisensor data access and analysis. In the multispectral data realm, ARD cubes are well-established. Recent advancements in hyperspectral EO, with missions like EnMAP and PRISMA, are pioneering globally sampled hyperspectral data collections. These collections expand EO data archives and harness synergies between multi- and hyperspectral time series. However, integrating hyperspectral data into ARD cubes and harmonizing them with multispectral data is an underexplored area for further enhancement.

Here, we introduce a workflow for seamlessly integrating EnMAP, Landsat, and Sentinel-2 data into harmonized multisensor ARD cubes. On the one hand, our workflow leverages the free and open-source FORCE processing engine to generate combined Landsat- and Sentinel-2 ARD cubes. On the other, we utilize multiple free and open-source Python libraries (EnPT, AROSICS, GDAL) for processing, harmonization, and integration of EnMAP data into the same ARD cube. We demonstrate the efficacy of developing harmonized EnMAP, Landsat, and Sentinel-2 ARD cubes for 2023 and 2024 across study sites in California and Southern Africa. Results from three case studies focusing on the retrieval of non-photosynthetic vegetation, characterization of plant life forms, and mapping woody species show that integrating multisensor data enhances the effective use of EnMAP time series, both independently and synergistically with Landsat and Sentinel-2 data.

Our workflow establishes a framework for integrating and harmonizing hyperspectral and multispectral satellite data. Future efforts will include data from missions like PRISMA, enhancing the temporal density of hyperspectral time series. This development advances multisensor data integration, crucial for next-generation global hyperspectral satellite missions like CHIME and SBG.

Future CHIME mission, from EU/ESA Copernicus programme, and SBG-VSWIR mission from NASA will provide orbital spectroscopic mapping with regular periodicity the Earth’s land mass and coastal areas. A revolutionary new capability for monitoring Earth’s agriculture, water and geologic resources, and ecosystems. These new missions will join a growing constellation of orbital imaging spectrometers such as EMIT, PRISMA, DESIS and EnMAP. While each of these instruments is powerful independently, their combination will enable a higher cadence of observations to meet demanding requirements of precision agriculture, terrestrial ecosystems, and snow measurements. To enable this vision, the SBG-VSWIR and CHIME teams have identified several key parameters, to be described in this talk, that bear on the interoperability of the surface reflectance products.

A few key topics relate to the fundamental definition of the surface reflectance product. In these key areas, agreement is vital for interchangeability, and SBG-VSWIR and CHIME have elected to follow similar paths. For example, both missions’ Level 2 products estimate the Hemispherical Directional Reflectance Factors (HDRFs) of surfaces that are tilted with the local digital elevation model. Both missions will use the Copernicus 30m DEM for these geometric calculations, ensuring that the target quantity for each observed location is in principle identical. Another important area is the basemap for image feature-based geolocation. Both missions have decided to use the Sentinel-2 Global Reflectance Image (GRI), facilitating seamless mosaics using both products. Some other choices do not change the target quantity, but simply represent different legitimate modeling assumptions for the retrieval process. These include solar irradiance models, glint models, aerosol parameterizations, and masking strategies. We will describe the methods pursued by the two teams and the reasons for favoring alignment or plurality in different areas.

We present a novel harmonization method that we use to create an improved at-sensor radiance product (L1B) for imaging spectrometers. The unique feature of our approach is that in addition to the center wavelength and angle, the shapes of the spectral and geometric response functions of each single pixel can be individually modified. Our method can therefore be used to harmonize data between different instruments, but also to correct optical artifacts, such as keystone and smile. In this way, an L1C product can be created that resembles data from an (almost) ideal imaging spectrometer with minimal optical artifacts. For example, the output product can be designed to feature a uniform Gaussian response function with constant shape for all geometric pixels and spectral channels.As usage of such a product requires less knowledge of the intricate optical properties of the imaging spectrometer used to obtain it, it significantly lowers the bar for inexperienced users to incorporate the product into their data evaluation chains. Additionally, it can reduce the computational burden on subsequent processing steps such as atmospheric correction or matched filter algorithms. Since artifacts removable by the proposed algorithm do not necessarily need to be corrected by optical engineering, our approach fosters cost-effective instrument design and production as overly strict (and often expensive) requirements on the optical instrument performance may be avoided.The resulting L1B product is obtained by linear transformation of the radiometrically calibrated at-sensor radiance data. While it can be created for almost any sensor systems, the algorithm was designed with typical hyperspectral imaging spectrometers in mind. Its computation requires a comprehensive knowledge of an instrument’s spectral and geometric response functions and it tends to work best with sensors that oversample in the spectral and spatial directions. We briefly summarize the mathematical prerequisites underlying the method before describing how we have implemented it to compute the L1B product for two airborne hyperspectral sensors (HySpex VNIR-3000N & SWIR-384) operated at the German Aerospace Center (DLR). Further, we describe the extensive on-ground system calibration performed in our laboratory to obtain the calibration key data required for the L1B product generation. We conclude with the presentation of exemplary L1B data, demonstrating that the algorithm works as expected and is capable of delivering an improved, high-quality and easy-to-use L1B product to the scientific community.We believe that a similar L1B product could be a valuable extension to current and upcoming imaging spectrometer missions.

The development of innovative, robust, and non-invasive mineral prospecting methods is fundamental for discovering new mineral resources that aid the green energy transition. Efficient use of satellite-based hyperspectral data, which are increasingly available from recently launched and upcoming satellite missions, can assist industries to create a significantly sustainable and risk-free environment to meet the demand of critical raw materials.

Spaceborne sensors like EnMAP (from DLR), EMIT (from NASA), and PRISMA (from ASI) offer an extensive volume of hyperspectral data, but at a relatively coarser spatial resolution as compared to airborne sources. The potential of these spaceborne platforms can be enhanced with analogous use of high-resolution (but limited spatial coverage) hyperspectral data from airborne platforms. This study aims to perform an efficient, multi-scale and multi-sensor unmixing, to take advantage of the (theoretically) increased endmember purity in higher-spatial resolution data and the large spatial extent of satellite data.

First, we construct reference mineral abundance maps from high-resolution airborne hyperspectral data, using manually selected endmember spectra and the non-negative least squares (NNLS) spectral unmixing algorithm. Then, satellite hyperspectral data are processed with non-dominated sorting genetic algorithm II (NSGA-II) to select spectral combinations that correlate with those derived with airborne data in a training area where these datasets overlap. The NSGA-II algorithm primarily functions on crossover (combining parts of two solutions) and mutation (randomly altering parts of the solution) mechanisms to explore the search space (i.e. the total number of spectral bands and indices representing the band combinations). NSGA-II iteratively selects and generates new solutions to efficiently balance the maximization of prediction accuracy and maintain diversity in spectral band selection, converging towards a set of optimal solutions. The predicted mineral abundances obtained with modeled spectral bands and empirical spectral-index equations are compared with abundances computed with NNLS on satellite hyperspectral data and the reference proxy abundance maps.

Notably, this study applies NSGA-II as an approach that can balance the trade-off between maximizing accuracy of predicted mineral abundance while ensuring relevant and diverse band selections that are robust to the various sources of noise or processing artifacts typical of satellite hyperspectral data (e.g., water bands). Our overarching aim is to develop a model that selects spectral subsets from hyperspectral data with enhanced predictive power, to improve the mechanism of unmixing algorithms and enhance the interpretability of results.

.