The Environmental Mapping and Analysis Program (EnMAP) is an spaceborne imaging spectroscopy mission under development by a consortium of German Earth Observation research institutions. The DLR - Agency is the project manager, the GFZ in Potsdam has the scientific lead, OHB - Systems AG is the industrial prime for the sensor and bus, and the DLR-EOC is responsible for the ground segment. EnMAP is designed for the retrieval of bio-, geochemical and physical parameters characterising the Earth surface for applications such as agriculture, land-use, water systems, soil science, and geology. The core payload of the EnMAP satellite is a combination of two prism-based imaging spectrometers covering the 420-2450 nm spectral window with 242 contiguous spectral bands. Spectral resolution ranges between 6.5 and 10 nm in the visible and near-infrared and between 8.5 and 12 nm in the shortwave infrared, with threshold signal-to-noise ratio of 500 and 150, respectively. The ground sampling distance is 30 m and the swath is 30 km. The mission is now in phase D (construction). The expected mission lifetime is 5 years. Top-of-atmosphere radiance and geometrically-corrected surface reflectance products will be produced and distributed by the ground segment. To ensure high quality standards, additional independent product validation activities are planned. Recent preparatory activities have supported industrial developments and the final consolidation of the mission concept. A scene simulator generating EnMAP-like data under realistic conditions has been implemented. It enables the definition of optimal instrument configurations for radiometric, spectral and geometric parameters as well as the evaluation and profiling of data-processing algorithms. In terms of the exploitation of EnMAP data, the so-called EnMAP-Box, a software environment for the interactive processing of EnMAP data, is being implemented. Tools for calibration, pre-processing and the derivation of higher-level biophysical products are to be included in this software. Several research groups in Germany, including application areas like agriculture, forestry, soils, geology and water, are collaborating for the development of those retrieval algorithms to be implemented in the EnMAP-Box.

In this talk we will provide an overview of the mission status including both technical developments and preparatory activities for the implementation of the EnMAP scientific program.

Rare earth elements (REEs) are a group of 17 chemically similar elements that cover fifteen lanthanides (atomic numbers 57-71 in the periodic table), plus scandium and yttrium (atomic numbers 21 and 39, respectively). This group of elements is essential to manufacturing many high technology products, and products that belong to the so called “green technology” category (e.g. wind turbines and hybrid vehicles). While the REEs are common constituents of the earth’s crust, they do not commonly occur as economically viable ore deposits. As a result, China currently holds more than 80% of the global REE production. The combination of the high supply disruption risk and the high economic importance of the REEs has led to the European Union and the U.S. Department of Energy declaring the REEs as “critical” raw materials.

The increasing global demand for the REEs has sparked a renewed interest toward this group of metals in the research community. One research avenue has focused on the potential of hyperspectral imaging (HSI) to detect the presence and identity of different REEs in minerals. These investigations have mainly been conducted in the visible to near-infrared (VNIR; 0.4-1.3 µm) and short-wave infrared (SWIR; 1.3-3.0 µm) wavelength regions. These studies have provided a wealth of relevant information to the hyperspectral detection of the REEs, but also highlighted some problems. One of these is the masking of the REEs in the presence of other minerals, and another is the overlap of the spectral features of different REEs. Despite the increasing undestanding on the limitations and potential of using the HSI technology for the detection of the REEs in the shorter wavelengths, their spectral features are still poorly established in the long-wave infrared (LWIR; 8.0-12.0 µm) wavelength region. To partially fill this gap, we have investigated the LWIR spectroscopy of twelve REE-bearing samples (nine mineral samples and three rock samples) that encompass carbonate, silicate and phosphate minerals. These samples were formerly characterized using the electron probe microanalysis (EPMA), scanning electron microscopy (SEM), and HSI data acquired in the VNIR and SWIR wavelength regions.

Our results suggest that the LWIR spectral features of the REE-bearing minerals analyzed in this study (bastnasite, parisite, monazite, xenotime, eudialyte, mosandrite and zircon) can be attributed exclusively to the fundamental stretching and bending modes of the X-O (X = C, Si, P) groups and as such, the wavelength range offers little information either on the presence or identity of different REEs. Longer wavelengths (>14 µm) have the potential to be used for this end, but the commercial imaging spectrometers currently available in the market need to overcome the challenge of strong atmospheric attenuation in the far infrared to be widely available for mineral exploration and exploitation applications.

In the following years EU legislation will regulate the acceptable toxin thresholds permitted for edible grain usage. The main expected outcome and impact of this activity is to reduce toxin content in wheat for food production. The research intends to develop a robust, efficient and non-destructive method to determine toxin concentrations caused by Fusarium and Claviceps infestations in wheat fields.
Field campaigns: 13-14.05.2018 (I), 25-26.05.2018 (II), 14-15.06.2018 (III); aimed at measurements of hyperspectral properties of plants (ASD FieldSpec 4+ASD PlantProbe), chlorophyll content in leaves (OptiScience CCM300) and chlorophyll fluorescence (OptiScience OS1p). Simultaneously were conducted other measurements and multispectral image acquisition from airborne level using the UAV’s Tatracam Micro-MCA6).
The research was carried out on 3 winter wheat fields (variety: solehio), with various crop history and microclimate: I – after “sweet” maize, relatively wet microclimate; II (western part) – after regular maize, relatively wet microclimate; III – after oil rape and ~30 km away in drier microclimate. Each field has 5 pairs of experimental plots (6x6 m each), located in zones of differing productivity. In each pair, one plot will be treated with fungicide (as the bulk of the field), and one will not be treated.
The analysis of spectral curves and remote sensing of vegetation indicators showed statistically significant differences in the state of wheat condition between the studied periods and between each pair, one plot was treated with fungicide (as the bulk of the field), and one wasn't be treated. Differences in the range of electromagnetic spectrum between polygons were visible in the range of spectrum describing of the amount of water in the plant as well as the ratio of chlorophyll to carotenoids. Statistical differences in the spectral reflection curves during the phenological period were the most important between the first and the second measurement period, which was related to the maturation of wheat.
Based on these measurements, harvest and logistic strategies will be created to separate contaminated and noncontaminated grains using GIS-technology. For farms with severe infestations, preventive methods will be developed to avoid food contamination.
Acknowledgements
Research has been carried out under the Separating MycOtoxin-contaminated Wheat grains using Precision Farming technologies. EIT Food 2018, project ID: 18116. The analyses were verified in the frame of the COST project: Optical synergies for spatiotemporal SENsing of Scalable ECOphysiological traits (SENESCO)

In 2014, the German Aerospace Center (DLR) and the US company Teledyne Brown Engineering (TBE) decided to collaborate and develop the hyperspectral instrument "DLR Earth Sensing Imaging Spectrometer" (DESIS) for integration into the "Multi-User System for Earth Sensing" (MUSES) installed on the International Space Station (ISS). Four years later, on 29 June 2018, the instrument was launched from Cape Canaveral with a SpaceX Dragon rocket. One month later, the instrument was robot-assisted integrated into MUSES and the commissioning phase started. First light images were available after 48 hours. The commissioning phase includes checkout for functional testing, End-to-End tests, processing parameter tuning, in-flight calibration tests and product validation. The operational phase of the mission has been entered several months later. From this point on, hyperspectral data will be available to the scientific and commercial user community.

In this paper a summary of results from the commissioning phase will be presented including aspects of data quality, operational performance and first data evaluation results.

The MUSES platform provides accommodations for two large and two small hosted payloads. MUSES is attached at the ELC-4 (EXPRESS Logistics Carriers) starboard of the ISS. It is a space-based, Earth-pointing platform providing position and attitude sensing, data downlink, and other core services common for each payload. DESIS has a mass of ~88 kg and is integrated in one of the large containers. Two gimbals allow a rotation of the whole MUSES platform around two axes resulting in ±25° forward / backward view, 45° backboard (port) view and 5° starboard view. The platform is equipped with a star tracker (sampling rate 10 Hz) and a MIMU (Miniature Inertial Measurement Unit) (sampling rate 50 Hz) providing a 10 Hz attitude measurement after filtering. ISS GPS data provide position and velocity vectors and time tags (sampling rate 1 Hz). The predicted viewing capability of MUSES, when operating at the ISS orbit inclination of 51.6°, will enable the DESIS instrument to scan greater than 90% of the populated Earth with an average cadence of about one week.

The DESIS hyperspectral instrument is realized as a pushbroom imaging spectrometer and features 235 bands with 3.5-4.0 nm spectral resolution, covering a range from 400 nm to 1000 nm with 30 m spatial resolution employing a 2-dimensional back illuminated CMOS (Complementary Metal Oxide Semiconductor) detector array. The optical design is based on the Offner-type grating spectrometer widely used in hyperspectral imaging. DESIS is equipped with a Pointing Unit (POI) consisting of a rotating mirror in front of the entrance slit allowing a forward and backward viewing change up to ±15° w.r.t. the nominal (e.g. nadir) view. The POI can be operated in a static mode with 3° angle steps for the viewing direction and in a dynamic mode with up to 1.5° change in viewing direction per seconds. This allows, besides standard Earth data products, acquisition of Bidirectional Reflectance Distribution Function (BRDF) products and continuous observations of the same targets on ground (using forward motion compensation mode).

DESIS has been characterized in laboratory and will be re-calibrated after launch employing measurements of internal light sources (bank of white and colored LED lamps) and globally distributed reference sites aiming to assess radiometric, spectrometric and geometric characteristics of the DESIS hyperspectral instrument in orbit. Vicarious validation activities will be performed at regular interval to ensure that a suitable data quality is met. This includes validation over well-known calibration sites, cross-comparisons with other simultaneous acquired data sets and analysis of homogenous areas.

The Ground Segment at DLR has developed the operational processing chain to derive different types of DESIS products from tiled data takes of size 1024 x 1024 pixels (~30 x 30 km²). An identical processing chain is licensed to TBE running on a cloud based system (© Amazon Cloud) for commercial product generation. DESIS level 1A products (Earth image scenes, on-board calibration measurements, dark current measurements and experimental products) will be long-term archived together with the corresponding metadata, while level 1B products (systematically and radiometrically corrected data), level 1C products (geometrically corrected data employing global references) and level 2A products (atmospherically compensated data) will be processed on demand before being delivered to the user for further value-add product generation. For scientific purposes only, DLR can share DESIS scientific data with other scientific organizations. TBE distributes data on a commercial basis.

Originally designed as a two-year mission, and launched in October 2001, the Project for OnBoard Autonomy-1 (Proba-1) continues to support ESA’s Earthnet Programme. With orbital drift, the mission has provided both ascending and descending mode data since 2014 and the Local Time of Descending Node (LTDN) is currently 02:59. Its Compact High Resolution Imaging Spectrometer (CHRIS) can provide up to 62 channels over the 400-1050 nm spectral range, operating in five different acquisition modes. The data is acquired for specific sites using the chosen acquisition mode, which may focus on having the maximum number of channels or highest spatial resolution depending on the application that can be marine and/or terrestrial; 17 m nadir ground sampling distance is the highest-possible spatial resolution. Also, Proba-1 can acquire multiple views of the same target at up to five different viewing angles: ±55°, ±36° and the standard nadir view; with off-track also available on request.

In the 17 years of operation, the programmability of CHRIS‘s spectral and spatial sampling has given rise to many investigations and findings, aiding the definition of future hyperspectral imaging missions for both scientific and service-oriented operational applications; including, potentially, the Copernicus Expansion Sentinel candidate 'CHIME'. The acquisition nominal plan has been expanded to include the eight instrumented Committee on Earth Observation Satellites (CEOS) LANDNET Reference Test Sites (https://calval.cr.usgs.gov/rst-resources/sites_catalog/ceos-reference-sites/), with additionally, some of the pre-existing Principle Investigator (PI) sites having been systematically collected throughout the mission’s lifetime. As a result, current activities are focused on understanding the changes in radiometric calibration alongside developing improved processing techniques that take advantage of the information provided by this unique time-series.

The presentation will place CHRIS-Proba within the context of both historical and future missions, and showcase results from the on-going activities that will not only allow the PIs and broader community to make better use of this dataset but also provide an enhanced understanding of requirements related to hyperspectral satellite missions.

In this paper, we propose a novel method aiming at reducing the noise in hyperspectral images based on the use of the general version of a neural network autoencoder.
In the process of hyperspectral image acquisition, there are several sources of noise that can have a negative impact on the successive processing steps of hyperspectral images. In general, the noise produced during the acquisition phase can be connected with two main sources: photon generated sampling noise, which is dependent on the number of electrons and is relatively insensitive to temperature, and dark current noise, that on the other hand is sensitive to temperature. In VIS/NIR hyperspectral imagers, the photon noise tends to dominate dark current noise, while thermal hyperspectral imagers tend to generate both types of noise in comparable amounts and in some cases the dark noise may be the dominant component. For these reasons signal-to-noise ratio (SNR), is one of the most important parameters in the design of hyperspectral imagers because a high SNR permits to mitigate both types of noise. However, a high level of SNR can be obtained using sensors and systems that are generally expensive. In some case, artefacts such as striping, and atmospherical contribution, also have a negative impact on the quality of hyperspectral image, but cannot be removed by increasing the SNR.
In the literature, several approaches and techniques have been proposed to improve the quality of hyperspectral images. Methods based on wavelet decomposition and Principal Component analysis (PCA) are widely used [1][2] for the removal of white uncorrelated noise. However, due to the linear nature of these methods, they are not effective when applied to images with nonlinear and/or correlated noise [3]. Techniques for the de-hazing and de-striping of hyperspectral images have been proposed too [4][5], but handling all types of degrading artefacts is still an open challenge.
A possible solution to these problems is presented in [6] where a projection in a nonlinear feature space, obtained through the use of Kernel Principal Component Analysis (KPCA), permits to effectively remove most of the linear and nonlinear noise affecting an image. The main drawback of the KPCA is the non invertible mapping function into the feature space. Thus, with KPCA it is only possible to obtain an estimation of the denoised image in the original data space [7].
A nonlinear generalization of the standard PCA can be obtained also with the Nonlinear Principal Component Analysis (NLPCA). Under certain conditions KPCA and NLPCA are able to provide similar mapping functions. However, the main advantage of NLPCA over KPCA is the invertibility of the mapping function, allowing the NLPCA to provide denoised reconstructions of the original images.
NLPCA is commonly based on multi-layer perceptrons (MLP) and performed by Autoassociative Neural Network (AANN) (or autoencoder) [8]. An AANN is a conventional feedforward NN having sigmoidal activation functions in each node and trained with by the Standard Conjugated Gradient (SCG) in order to minimize the mean sum-of-squares error (MSE). Generally, the nonlinear AANN sees the use of three hidden layers, including an internal bottleneck layer of smaller dimension than either input or output. The AANN is trained to perform identity mapping, where the input X has to be equal to the output X′. This means that if the training phase finds an acceptable solution, i.e., a solution that gives an error below a predefined threshold, a good compressed representation of the input must exist in the bottleneck layer. Since there are fewer nodes in the bottle- neck layer than in the input/output, the bottleneck nodes must represent or encode the information obtained from the inputs for the subsequent layers to reconstruct the input.
In this study the proposed approach has been applied to several hyperspectral images having different characteristics and affected by different kind of noises and artefacts. A qualitative and quantitative comparison with state of the art techniques has been also carried out.
A first experiment has been performed using a data set acquired by the ROSIS sensor (0.43 to 0.86 μm with 103 bands) over the University of Pavia, Italy. The image presents several bands, characterized by uncorrelated noise. The same approach has been applied to a Hyperion image (220 hyperspectral bands from 0.4 to 2.5 μm) acquired in 2008 over the Campi Flegrei area, North-West of Naples, Italy. From the original dataset only 155 spectrally unique bands have been retained, discarding the bands without relevant information. In this case the de-noising of the image is extremely challenging because several kinds of noise are present at the same time. In particular, the image has been acquired from a satellite, meaning that the atmospheric contribution has a relevant role in the noise contribution of the image. Moreover, the image is also strongly affected by striping, a spatially correlated noise. A third experiment has been performed by analyzing the effectiveness of denoising applied to the CASI hyperspectral image used in the 2013 Data Fusion Contest. In this case the hyperspectral imagery consists of an airborne acquisition of 144 spectral bands in the 360 nm to 1050 nm spectral range and has been calibrated to at-sensor radiance. The image is strongly affected by nonlinear signal dependent noise in the first bands. A fourth experiment has been carried out using a hyperspectral image acquired in the thermal domain. The thermal hyperspectral image was acquired using the ”Hyper-Cam” air- borne LWIR hyperspectral imager consisting of 84 spectral bands in the 868 to 1280 cm−1 region (7.8 μm to 11.5 μm), at a spectral resolution of 6 cm−1. The spatial resolution is approximately 1 m. Differently from the dataset investigated in the previous experiments, in the thermal hyperspectral image the dark noise is the dominant component affecting more the last bands of the image, but at the same time the photon noise could not be neglected and is more evident in the first bands.
The obtained results demonstrated a general good performance of the NLPCA approach for noise suppression and artefact removal. In particular, the experiments proposed in this work evidenced that PCA and NLPCA permit to remove more noise from hyper spectral images than wavelet shrinkage based methods. This is because the wavelet shrinkage operates in the spatial domain while PCA and NLPCA operate in the spectral domain. However, while PCA permits to mitigate the influence of additive noise, NLPCA is able to remove both linear and nonlinear noises. Moreover, thanks to its nonlinear nature, NLPCA was also able to suppress striping artefacts.


[1] B. Rasti, J.R. Sveinsson, M.O. Ulfarsson, and J.A Benediktsson, “Hyperspectral image denoising using first order spectral roughness penalty in wavelet domain,” Selected Topics in Applied Earth Observations and Remote Sensing, IEEE Journal of, vol. 7, no. 6, pp. 2458–2467, June 2014.

[2] G. Chen and S. E. Qian, “Denoising of hyperspectral imagery using principal component analysis and wavelet shrinkage,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 3, pp. 973–980, 2011.

[3] Goossens B., Pizurica A., and Philips W., “Removal of correlated noise by modeling the signal of interest in the wavelet domain,” Image Processing, IEEE Transactions on, vol. 18, no. 6, pp. 1153–1165, June 2009.

[4] Y. Zhang and B. Guindon, “Quantitative assessment of a haze suppression methodology for satellite imagery: Effect on land cover classification performance,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 5, pp. 1082–1089, May 2003.

[5] A. Plaza, J. Benediktsson et al., "Recent advances in techniques for hyperspectral image processing", Remote Sens. Environ., vol. 113, pp. 110-122, Sep. 2009.

[6] S. Mika, B. Schlkopf, A. Smola, K. R. Muller, M. Scholz, and G.Ratsch, “Kernel pca and de-noising in feature spaces,” Advances in Neural Information Pro- cessing Systems, vol. 11, pp. 536–542, 1999.

[7] T.Y.KwokandI.W.H.Tsang,“Thepre-imageproblem in kernel methods,” IEEE Trans. Neural Networks, vol. 15, no. 6, pp. 1517–1525, 2004.

[8] C. Bishop, “Neural networks for pattern recognition,” Oxford Univ. Press, London, U.K., 1995.