Search

Hyperspectral satellite imagery provides highly resolved spectral information for large areas and can provide vital information. However, only a few imaging spectrometer missions are currently in operation. Aiming to generate synthetic satellite-based hyperspectral imagery potentially covering any region, we explored the possibility of applying statistical learning, i.e., emulation. Based on the relationship of a Sentinel-2 (S2) scene and a hyperspectral HyPlant airborne image, this work demonstrates the possibility to emulate a hyperspectral S2-like image. We tested the role of different machine learning regression algorithms and varied the image-extracted training dataset size. We found superior performance of neural network as opposed to the other algorithms when trained with large datasets (up to 100 000 samples). The developed emulator was then applied to the L2A (bottom-of-atmosphere reflectance) S2 subset, and the obtained S2-like hyperspectral reflectance scene was evaluated. The validation of emulated against reference spectra demonstrated the potential of the technique. $R^{2}$ values between 0.75 and 0.9 and NRMSE between 2 and 5% across the full 402–2356 nm range were obtained. Moreover, epistemic uncertainty is obtained using the dropout technique, revealing spatial fidelity of the emulated scene. We obtained highest SD values of 0.05 (CV of 8%) in clouds and values below 0.01 (CV of 7%) in vegetation land covers. Finally, the emulator was applied to an entire S2 tile (5490 × 5490 pixels) to generate a hyperspectral reflectance datacube with the texture of S2 (60 Gb, at a speed of 0.14 s/10000 pixels). As the emulator can convert any S2 tile into a hyperspectral image, such scenes give perspectives how future satellite imaging spectroscopy will look like.

Accurate plant-type (PT) detection forms an important basis for sustainable land management maintaining biodiversity and ecosystem services. In this sense, Sentinel-2 satellite images of the Copernicus program offer spatial, spectral, temporal, and radiometric characteristics with great potential for mapping and monitoring PTs. In addition, the selection of a best-performing algorithm needs to be considered for obtaining PT classification as accurate as possible . To date, no freely downloadable toolbox exists that brings the diversity of the latest supervised machine-learning classification algorithms (MLCAs) together into a single intuitive user-friendly graphical user interface (GUI). To fill this gap and to facilitate and automate the usage of MLCAs, here we present a novel GUI software package that allows systematically training, validating, and applying pixel-based MLCA models to remote sensing imagery. The so-called MLCA toolbox has been integrated within ARTMO’s software framework developed in Matlab which implements most of the state-of-the-art methods in the machine learning community. To demonstrate its utility, we chose a heterogeneous case study scene, a landscape in Southwest Iran to map PTs. In this area, four main PTs were identified, consisting of shrub land, grass land, semi-shrub land, and shrub land–grass land vegetation. Having developed 21 MLCAs using the same training and validation, datasets led to varying accuracy results. Gaussian process classifier (GPC) was validated as the top-performing classifier, with an overall accuracy (OA) of 90%. GPC follows a Laplace approximation to the Gaussian likelihood under the supervised classification framework, emerging as a very competitive alternative to common MLCAs. Random forests resulted in the second-best performance with an OA of 86%. Two other types of ensemble-learning algorithms, i.e., tree-ensemble learning (bagging) and decision tree (with error-correcting output codes), yielded an OA of 83% and 82%, respectively. Following, thirteen classifiers reported OA between 70% and 80%, and the remaining four classifiers reported an OA below 70%. We conclude that GPC substantially outperformed all classifiers, and thus, provides enormous potential for the classification of a diversity of land-cover types. In addition, its probabilistic formulation provides valuable band ranking information, as well as associated predictive variance at a pixel level. Nevertheless, as these are supervised (data-driven) classifiers, performances depend on the entered training data, meaning that an assessment of all MLCAs is crucial for any application. Our analysis demonstrated the efficacy of ARTMO’s MLCA toolbox for an automated evaluation of the classifiers and subsequent thematic mapping.

Thanks to the emergence of cloud-computing platforms and the ability of machine learning methods to solve prediction problems efficiently, this work presents a workflow to automate spatiotemporal mapping of essential vegetation traits from Sentinel-3 (S3) imagery. The traits included leaf chlorophyll content (LCC), leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), and fractional vegetation cover (FVC), being fundamental for assessing photosynthetic activity on Earth. The workflow involved Gaussian process regression (GPR) algorithms trained on top-of-atmosphere (TOA) radiance simulations generated by the coupled canopy radiative transfer model (RTM) SCOPE and the atmospheric RTM 6SV. The retrieval models, named to S3-TOA-GPR-1.0, were directly implemented in Google Earth Engine (GEE) to enable the quantification of the traits from TOA data as acquired from the S3 Ocean and Land Colour Instrument (OLCI) sensor. Following good to high theoretical validation results with normalized root mean square error (NRMSE) ranging from 5% (FAPAR) to 19% (LAI), a three fold evaluation approach over diverse sites and land cover types was pursued: (1) temporal comparison against LAI and FAPAR products obtained from Moderate Resolution Imaging Spectroradiometer (MODIS) for the time window 2016–2020, (2) spatial difference mapping with Copernicus Global Land Service (CGLS) estimates, and (3) direct validation using interpolated in situ data from the VALERI network. For all three approaches, promising results were achieved. Selected sites demonstrated coherent seasonal patterns compared to LAI and FAPAR MODIS products, with differences between spatially averaged temporal patterns of only 6.59%. In respect of the spatial mapping comparison, estimates provided by the S3-TOA-GPR-1.0 models indicated highest consistency with FVC and FAPAR CGLS products. Moreover, the direct validation of our S3-TOA-GPR-1.0 models against VALERI estimates indicated good retrieval performance for LAI, FAPAR and FVC. We conclude that our retrieval workflow of spatiotemporal S3 TOA data processing into GEE opens the path towards global monitoring of fundamental vegetation traits, accessible to the whole research community.

Non-photosynthetic vegetation (NPV) biomass has been identified as a priority variable for upcoming spaceborne imaging spectroscopy missions, calling for a quantitative estimation of lignocellulosic plant material as opposed to the sole indication of surface coverage. Therefore, we propose a hybrid model for the retrieval of non-photosynthetic cropland biomass. The workflow included coupling the leaf optical model PROSPECT-PRO with the canopy reflectance model 4SAIL, which allowed us to simulate NPV biomass from carbon-based constituents (CBC) and leaf area index (LAI). PROSAIL-PRO provided a training database for a Gaussian process regression (GPR) algorithm, simulating a wide range of non-photosynthetic vegetation states. Active learning was employed to reduce and optimize the training data set. In addition, we applied spectral dimensionality reduction to condense essential information of non-photosynthetic signals. The resulting NPV-GPR model was successfully validated against soybean field data with normalized root mean square error (nRMSE) of 13.4% and a coefficient of determination (R2) of 0.85. To demonstrate mapping capability, the NPV-GPR model was tested on a PRISMA hyperspectral image acquired over agricultural areas in the North of Munich, Germany. Reliable estimates were mainly achieved over senescent vegetation areas as suggested by model uncertainties. The proposed workflow is the first step towards the quantification of non-photosynthetic cropland biomass as a next-generation product from near-term operational missions, such as CHIME.

The retrieval of sun-induced fluorescence (SIF) from hyperspectral radiance data grew to maturity with research activities around the FLuorescence EXplorer satellite mission FLEX, yet full-spectrum estimation methods such as the spectral fitting method (SFM) are computationally expensive. To bypass this computational load, this work aims to approximate the SFM-based SIF retrieval by means of statistical learning, i.e., emulation. While emulators emerged as fast surrogate models of simulators, the accuracy-speedup trade-offs are still to be analyzed when the emulation concept is applied to experimental data. We evaluated the possibility of approximating the SFM-like SIF output directly based on radiance data while minimizing the loss in precision as opposed to SFM-based SIF. To do so, we implemented a double principal component analysis (PCA) dimensionality reduction, i.e., in both input and output, to achieve emulation of multispectral SIF output based on hyperspectral radiance data. We then evaluated systematically: (1) multiple machine learning regression algorithms, (2) number of principal components, (3) number of training samples, and (4) quality of training samples. The best performing SIF emulator was then applied to a HyPlant flight line containing at sensor radiance information, and the results were compared to the SFM SIF map of the same flight line. The emulated SIF map was quasi-instantaneously generated, and a good agreement against the reference SFM map was obtained with a R2 of 0.88 and NRMSE of 3.77%. The SIF emulator was subsequently applied to 7 HyPlant flight lines to evaluate its robustness and portability, leading to a R2 between 0.68 and 0.95, and a NRMSE between 6.42% and 4.13%. Emulated SIF maps proved to be consistent while processing time was in the order of 3 min. In comparison, the original SFM needed approximately 78 min to complete the SIF processing. Our results suggest that emulation can be used to efficiently reduce computational loads of SIF retrieval methods.