Final Report Summary - FLUORBRDF (Assessing BRDF effects of sun-induced fluorescence for quantifying canopy-level photosynthetic efficiency)
The FLUORBRDF project was proposed to enforce a more robust spatially explicit retrieval of chlorophyll fluorescence. FLUORBRDF aimed to account for bidirectional reflectance distribution (BRDF) effects, which is not only a function of sun-sensor geometry but also of the anisotropic properties of vegetation cover. Challenges related to BRDF effect affecting the fluorescence signal brought us to the following objectives:
(i) to gain insights in the BRDF effects of the canopy-level emitted Fs signal on a theoretical basis using radiative transfer (RT) models;
(ii) to develop a methodology to correct for it; and
(iii) to achieve higher accuracies of airborne fluorescence measures of heterogeneous canopies.
This project was carried out at Valencia University in the Laboratory of Earth Observation (LEO), headed by Prof J. Moreno. This group is also the driver behind the Fluorescence Explorer (FLEX) mission, a European Space Agency (ESA)'s Earth Explore candidate with the task to measure vegetation chlorophyll fluorescence from Space and the exploitation of this signal to better understand the carbon cycle. Meanwhile ESA's Earth Science Advisory Committee (SAC) recommended the investigation of the FLEX concept as an in-orbit demonstrator carrying a fluorescence imaging spectrometer (FLORIS), to be flown as a tandem mission with Sentinel-3 (S-3). This tandem concept consists of a high-resolution spectrometer measuring in the O2-A and O2-B absorption features complemented by S-3 instruments.
The main results can be categorised into the development of novel methodologies and toolboxes on the one hand, and results that were generated during the FLUORBRDF project on the other hand. They are further explained below:
Development of novel methodologies and toolboxes
Since the FLEX mission is still in development phase an effective approach to gain insight in the effects of canopy structure on the fluorescence signal as measured from Space is to make use of RT models. RT modelling plays a key role for Earth observation (EO) because it is needed to design EO instruments and to develop, apply and test inversion algorithms. The inversion of a RT model is considered as a successful approach for the retrieval of biophysical parameters (e.g. chlorophyll content, leaf area index, fluorescence) because of being physically-based and generally applicable. However, to the broader community this approach is considered as laborious because of its many processing steps and expert knowledge is required to realise precise model parameterisation.
In an attempt to make these RT models more accessible to the broader community and to automate the many steps that are needed to operate these models, we have developed a RT toolbox which is named automated radiative transfer model operator (ARTMO). ARTMO is a scientific toolbox that provides in a graphical user interface (GUI) essential models and tools required for terrestrial EO applications such as sensitivity studies and model inversion. In short, the toolbox allows the user:
(i) to choose between various plant leaf and canopy RT models (e.g. models from the PROSPECT and SAIL family, FLIGHT);
(ii) to choose between spectral band settings of various air- and space-borne sensors or defining own sensor settings;
(iii) to simulate a massive amount of spectra based on a look up table (LUT) approach and storing it in a relational database;
(iv) to plot spectra of multiple models and compare them with measured spectra, and finally;
(v) to run model inversion against optical imagery given several cost options and accuracy estimates.
The ARTMO toolbox formed the basis of various studies related to the FLUORBRDF project. The toolbox is still under development and first versions have been enthusiastically received at international conferences (EARSeL Imaging Spectroscoy workshop 2011, EGU 2011, 2012, IGARS 2012). Further information can be found on a dedicated website: https://sites.google.com/site/jochemverrelst/ARTMO.
To enhance scientific progress in FLUORBRDF (and fluorescence-related research in general) special efforts were devoted to implement FluorMODleaf and FluorSAIL into ARTMO. These models have been recently developed and simulate fluorescence fluxes at leaf level and canopy level, respectively. ARTMO takes care of their coupling and enables the user to automate the running of simulations given a wide range of scenarios. A large spectral database has been generated and coupled with and atmospheric model MOMO to upscale reflectance and fluorescence fluxes to top-of-atmosphere radiance. This work called FLUSS (atmospheric corrections for fluorescence signal and surface pressure retrieval over land) formed the basis of various FLEX-related sensitivity studies in cooperation of Oxford University and the Free University of Berlin.
At the same time, the development of ARTMO initiated a new research line that is devoted to the development of more powerful retrieval algorithms. Currently two strategies are followed.
(1) The first strategy relies on the inversion of these RT models against satellite images. This is typically done through the generation of look-up tables (LUT) and applying a cost function (minimum distance) and regularisation options. Together with a mathematician from Swansea University we have identified and implemented more than 60 different cost functions. These cost functions found origin in various fields (e.g. mathematics, signal processing, bioinformatics) and the most of them are not known in the EO community. We implemented a module to evaluate these cost functions and regularisation options such as added noise prior to apply an inversion strategy to the whole image. Moreover, the LUT-based inversion strategy has been made sufficiently flexible that it can account to spectral and structural variations in different vegetation covers.
(2) The second strategy relies on machine learning regression algorithms (MLRAs). This research line was initiated by collaborating with the signal processing group at University of Valencia (Prof. G. Camps-Valls). MLRA have the potential to generate adaptive, robust relationships and, once trained by a generic LUT, they are very fast to apply to any image. MLRAs learn the relationship between the input (e.g. reflectances) and output (e.g. biophysical parameters) by fitting a flexible model directly from the data. Typically, MLRAs are able to cope with the strong nonlinearity of the functional dependence between the parameter and the observed reflectance. Currently state-of-the-art MLRAs are being implemented into ARTMO, such as kernel ridge regression (KRR), support vector regression (SVR), neural networks (NN) and Gaussian processes regression (GPR). Particularly the latter holds promise for processing remotely sensed images as GPR is transparent in terms of model development and does not only provides mean estimates but also confidence intervals.
Finally, the ARTMO toolbox is currently being prepared for implementing new types of 'modules' and 'apps'. A 'vegetation indices' module is under construction, in which automatically a wide range of vegetation indices can be evaluated that then subsequently can be used as a quick way to map biophysical parameters. Apps are small utilities that take care of a specific task. Here, efforts are underway to implement atmospheric models (e.g. 6S or MODTRAN) and the energy budget model SCOPE that takes fluorescence fluxes into account. The latter is of specific interest in further ESA-related fluorescence studies. Examples of added apps are provided.
Results that generated during the FLUORBRDF project:
The development of the ARTMO toolbox went hand in hand with the execution of various studies that took part of the FLUORBRDF project. The most important results are summarised below:
(1) In view of assessing the impact of BRDF effect on retrieval of biophysical parameters ESA's multi-angular CHRIS sensor was used to assess the retrieval of leaf area index (LAI). In comparison to nadir viewing geometry, LAI retrievals can be slightly improved when configuring the viewing angle closer to the hot spot position. In turn, when configuring in the opposite direction (forward viewing) then LAI retrievals worsened due to more effects of shadowing. This work is currently in revision in Remote Sensing journal.
(2) Since fluorescence emission is directly related to total chlorophyll content, of importance is to have spatially-explicit information of chlorophyll content available. That was the starting point of applying MLRAs to have chlorophyll content and other biophysical parameters estimated. Specifically GPR proved to be a successful regressor. We compared this novel regressor against competitive regressors in view of ESA's forthcoming Sentinel-2 and Sentinel-3 missions. GPR demonstrated to be a fast regressor and at the same delivered excellent accuracies. Results are relevant in view of having FLEX configured in the tandem with Sentinel-3. Results have been published in various scientific journals, being remote sensing of environment, IEEE Transactions in Geoscience and Remote Sensing, and in review in Journal of Selected Topics in Applied Earth Observations and Remote Sensing. Examples of GPR applied to airborne hyperspectral CASI data for chlorophyll content mapping are provided. Zoom-ins of original image, mean estimates and associated confidence intervals are provided.
(3) A large synthetic dataset was generated by ARTMO that was used for various sensitivity studies in the FLUSS project. These studies ranged from assessing effects of polarisation, BRDF effects and atmospheric parameters, optimising band configurations and developing fluorescence retrieval strategies. Specifically, a whole study was devoted the added value of having FLEX configured with Sentinel-3. It was demonstrated not only that with respect to fluorescence parameters FLEX proved to be significantly more sensitive than Sentinel-3 but also that by using the synergy of both datasets accuracies still improved. These results have been presented to ESA and to international conferences (IGARSS 2012, ESA's Sentinel-3 symposium).
(i) to gain insights in the BRDF effects of the canopy-level emitted Fs signal on a theoretical basis using radiative transfer (RT) models;
(ii) to develop a methodology to correct for it; and
(iii) to achieve higher accuracies of airborne fluorescence measures of heterogeneous canopies.
This project was carried out at Valencia University in the Laboratory of Earth Observation (LEO), headed by Prof J. Moreno. This group is also the driver behind the Fluorescence Explorer (FLEX) mission, a European Space Agency (ESA)'s Earth Explore candidate with the task to measure vegetation chlorophyll fluorescence from Space and the exploitation of this signal to better understand the carbon cycle. Meanwhile ESA's Earth Science Advisory Committee (SAC) recommended the investigation of the FLEX concept as an in-orbit demonstrator carrying a fluorescence imaging spectrometer (FLORIS), to be flown as a tandem mission with Sentinel-3 (S-3). This tandem concept consists of a high-resolution spectrometer measuring in the O2-A and O2-B absorption features complemented by S-3 instruments.
The main results can be categorised into the development of novel methodologies and toolboxes on the one hand, and results that were generated during the FLUORBRDF project on the other hand. They are further explained below:
Development of novel methodologies and toolboxes
Since the FLEX mission is still in development phase an effective approach to gain insight in the effects of canopy structure on the fluorescence signal as measured from Space is to make use of RT models. RT modelling plays a key role for Earth observation (EO) because it is needed to design EO instruments and to develop, apply and test inversion algorithms. The inversion of a RT model is considered as a successful approach for the retrieval of biophysical parameters (e.g. chlorophyll content, leaf area index, fluorescence) because of being physically-based and generally applicable. However, to the broader community this approach is considered as laborious because of its many processing steps and expert knowledge is required to realise precise model parameterisation.
In an attempt to make these RT models more accessible to the broader community and to automate the many steps that are needed to operate these models, we have developed a RT toolbox which is named automated radiative transfer model operator (ARTMO). ARTMO is a scientific toolbox that provides in a graphical user interface (GUI) essential models and tools required for terrestrial EO applications such as sensitivity studies and model inversion. In short, the toolbox allows the user:
(i) to choose between various plant leaf and canopy RT models (e.g. models from the PROSPECT and SAIL family, FLIGHT);
(ii) to choose between spectral band settings of various air- and space-borne sensors or defining own sensor settings;
(iii) to simulate a massive amount of spectra based on a look up table (LUT) approach and storing it in a relational database;
(iv) to plot spectra of multiple models and compare them with measured spectra, and finally;
(v) to run model inversion against optical imagery given several cost options and accuracy estimates.
The ARTMO toolbox formed the basis of various studies related to the FLUORBRDF project. The toolbox is still under development and first versions have been enthusiastically received at international conferences (EARSeL Imaging Spectroscoy workshop 2011, EGU 2011, 2012, IGARS 2012). Further information can be found on a dedicated website: https://sites.google.com/site/jochemverrelst/ARTMO.
To enhance scientific progress in FLUORBRDF (and fluorescence-related research in general) special efforts were devoted to implement FluorMODleaf and FluorSAIL into ARTMO. These models have been recently developed and simulate fluorescence fluxes at leaf level and canopy level, respectively. ARTMO takes care of their coupling and enables the user to automate the running of simulations given a wide range of scenarios. A large spectral database has been generated and coupled with and atmospheric model MOMO to upscale reflectance and fluorescence fluxes to top-of-atmosphere radiance. This work called FLUSS (atmospheric corrections for fluorescence signal and surface pressure retrieval over land) formed the basis of various FLEX-related sensitivity studies in cooperation of Oxford University and the Free University of Berlin.
At the same time, the development of ARTMO initiated a new research line that is devoted to the development of more powerful retrieval algorithms. Currently two strategies are followed.
(1) The first strategy relies on the inversion of these RT models against satellite images. This is typically done through the generation of look-up tables (LUT) and applying a cost function (minimum distance) and regularisation options. Together with a mathematician from Swansea University we have identified and implemented more than 60 different cost functions. These cost functions found origin in various fields (e.g. mathematics, signal processing, bioinformatics) and the most of them are not known in the EO community. We implemented a module to evaluate these cost functions and regularisation options such as added noise prior to apply an inversion strategy to the whole image. Moreover, the LUT-based inversion strategy has been made sufficiently flexible that it can account to spectral and structural variations in different vegetation covers.
(2) The second strategy relies on machine learning regression algorithms (MLRAs). This research line was initiated by collaborating with the signal processing group at University of Valencia (Prof. G. Camps-Valls). MLRA have the potential to generate adaptive, robust relationships and, once trained by a generic LUT, they are very fast to apply to any image. MLRAs learn the relationship between the input (e.g. reflectances) and output (e.g. biophysical parameters) by fitting a flexible model directly from the data. Typically, MLRAs are able to cope with the strong nonlinearity of the functional dependence between the parameter and the observed reflectance. Currently state-of-the-art MLRAs are being implemented into ARTMO, such as kernel ridge regression (KRR), support vector regression (SVR), neural networks (NN) and Gaussian processes regression (GPR). Particularly the latter holds promise for processing remotely sensed images as GPR is transparent in terms of model development and does not only provides mean estimates but also confidence intervals.
Finally, the ARTMO toolbox is currently being prepared for implementing new types of 'modules' and 'apps'. A 'vegetation indices' module is under construction, in which automatically a wide range of vegetation indices can be evaluated that then subsequently can be used as a quick way to map biophysical parameters. Apps are small utilities that take care of a specific task. Here, efforts are underway to implement atmospheric models (e.g. 6S or MODTRAN) and the energy budget model SCOPE that takes fluorescence fluxes into account. The latter is of specific interest in further ESA-related fluorescence studies. Examples of added apps are provided.
Results that generated during the FLUORBRDF project:
The development of the ARTMO toolbox went hand in hand with the execution of various studies that took part of the FLUORBRDF project. The most important results are summarised below:
(1) In view of assessing the impact of BRDF effect on retrieval of biophysical parameters ESA's multi-angular CHRIS sensor was used to assess the retrieval of leaf area index (LAI). In comparison to nadir viewing geometry, LAI retrievals can be slightly improved when configuring the viewing angle closer to the hot spot position. In turn, when configuring in the opposite direction (forward viewing) then LAI retrievals worsened due to more effects of shadowing. This work is currently in revision in Remote Sensing journal.
(2) Since fluorescence emission is directly related to total chlorophyll content, of importance is to have spatially-explicit information of chlorophyll content available. That was the starting point of applying MLRAs to have chlorophyll content and other biophysical parameters estimated. Specifically GPR proved to be a successful regressor. We compared this novel regressor against competitive regressors in view of ESA's forthcoming Sentinel-2 and Sentinel-3 missions. GPR demonstrated to be a fast regressor and at the same delivered excellent accuracies. Results are relevant in view of having FLEX configured in the tandem with Sentinel-3. Results have been published in various scientific journals, being remote sensing of environment, IEEE Transactions in Geoscience and Remote Sensing, and in review in Journal of Selected Topics in Applied Earth Observations and Remote Sensing. Examples of GPR applied to airborne hyperspectral CASI data for chlorophyll content mapping are provided. Zoom-ins of original image, mean estimates and associated confidence intervals are provided.
(3) A large synthetic dataset was generated by ARTMO that was used for various sensitivity studies in the FLUSS project. These studies ranged from assessing effects of polarisation, BRDF effects and atmospheric parameters, optimising band configurations and developing fluorescence retrieval strategies. Specifically, a whole study was devoted the added value of having FLEX configured with Sentinel-3. It was demonstrated not only that with respect to fluorescence parameters FLEX proved to be significantly more sensitive than Sentinel-3 but also that by using the synergy of both datasets accuracies still improved. These results have been presented to ESA and to international conferences (IGARSS 2012, ESA's Sentinel-3 symposium).
