ARTMO s new Machine Learning Regression Algorithms (MLRA) module for mapping biophysical parameters

Transcription

1 Earsel ISW 9/4/13 1/23 ARTMO s new Machine Learning Regression Algorithms (MLRA) module for mapping biophysical parameters Jochem Verrelst, Juan Pablo Rivera, Jordi Muñoz-Mari, Jose Moreno, Gustavo Camps-Valls

2 Earsel ISW 9/4/13 2/23 Outlook Background Biophysical parameter retrieval Nonparametric regression for retrieval of biophysical parameters ARTMO MLRA toolbox MLRAs MLRA settings Results tests Retrievals Multi-output Coupling with RTMs Conclusions

3 Earsel ISW 9/4/13 3/23 Basics biophysical parameter retrieval Retrieval of biophysical parameters from optical EO data always occurs through a model; e.g. through statistical models, through inversion of physicallybased radiative transfer models (RTM), or through hybrid forms. Statistical approaches Physically based RTM approaches RTM Hybrid forms Design Retrieval Evaluation VIs

4 Earsel ISW 9/4/13 4/23 Parametric Nonparametric Physicallybased inversion Parametric regression: Some constraints introduced Nonparametric regression: No constraints in developing models Physically-based approaches: Inversion of RTMs using parametric or nonparametric inversion techniques (i.e., hybrid forms). Nonparametric regression is a form of regression analysis in which the predictor does not take a predetermined form but is constructed according to information derived from the data. Nonparametric regression requires larger sample sizes than regression based on parametric models because the data must supply the model structure as well as the model estimates. Nonparametric regression based on machine learning algorithms: Linear transformations: PCR, PLS Non-linear transformations: NN, GPR, KRR, SVR.

5 Earsel ISW 9/4/13 5/23 Pros Nonparametric regression Powerful Relatively fast Additional features Multi-output Cons Portability? Black-box (for the majority) Difficult to use Operational retrieval of biophysical parameters Neural networks coupled with RTMs widely used but face limitations: black box, unstable, difficult and slow in training. MERIS SPOT-VGT Sentinel-2,-3 The ability of NN to process hyperspectral data? Towards new generation of MLRA regressors for operational use.

6 Earsel ISW 9/4/13 6/23 V3: Modular design All models and modules can be accessed from the Menu bar Simulations according to a predefined sensor setting LUTs can be configured per land cover class or defined by user. File Models Forward Modules Retrieval Load Project New Project DB adminstration Settings Model inputs Save Load New DB Change DB Delete Import DB Export DB Analyze DB Leaf Canopy Combined Leaf Canopy Combined PROSPECT 4 PROSPECT 5 FluorMODleaf 4SAIL FLIGHT FluorSAIL Sensor Graphics Spectral Indices Machine Learning LUT-based Inversion LUT class Project Database SLC SCOPE Simulations can be done for any sensor in 4-24 nm range. Input, output and metadata stored in MySQL running underneath.

7 Earsel ISW 9/4/13 7/23 MLRA toolbox Load Image & Class Input MLRA setting Test MLRAs Retrieval Tools Load Image Load Class RTM data User data Select project Edit settings Single-output Multi-output Load test New test Save Load Test database View maps Settings When loading a land cover map then retrieval strategies can be optimized per class. In Test MLRAs, run tests with the predefined strategies or load an existing test Input data can come either from RTMs, from field observations or from both. Manually set up a MLRA retrieval strategy or select an earlier evaluated strategy. In MLRA setting multiple MLRA retrieval strategies can configured, either single-output or multioutput Tools offer various options to manage the MLRA Module. Rename Delete

8 Implemented MLRAs Linear nonparametric regressors: Linear regression (LR) Partial least squares regression (PLS) Alternative regressors are planned to be added: PCR, LASSO Earsel ISW 9/4/13 8/23

9 MLRA settings If active, configure per land cover class. Data: SPARC campaign, Barrax, Spain Select a MLRA Options to add noise Option to mix RTM with field data Option to select bands (manually or automatically through mutual information: band with most information first) Field data: LCC measured with CCM-2 LAI measured with LiCor LAI-2 Spectral data: CHRIS mode 1 (62 bands; 34m) nadir spectra HyMap (5 m resolution; 125 bands ; nm Jochem.verrelst@uv.es Earsel ISW 9/4/13 9/23

10 Earsel ISW 9/4/13 1/23 training Results test Results can be organized according to land cover class, parameter, cal/val, and statistical output Overview of results. Here, best results per SI andcurve fitting Options to plot all kinds of output and export results Selected strategies appear here and can be transported to the Retrieval module. 6 1:1 Gaussians Processes Regression - LAI output 2 x 14 GPR bands Gaussians Processes Regression SNoise:2 - ParmNoise: RTMtrain:.1 - Usertrain:.65 1D results lai_user_gaussians Processes Regression.2 2D results RELRMSE_user_lai_Neural Network Estimated Measured Sigma Bands RELRMSE model t rain training model t rain spect n oise noise Case study SPARC-CHRIS

11 LCC User data (SPARC - CHRIS) LR PLS NN KRR GPR LR poor, only suitable with high training. PLS somewhat better, but not excellent performances. Needs noise. NN behaves erratic: can lead to good performances but unstable. KKR: Excellent performances, very robust. GPR: Excellent performances, robust. SI 3-bands LR: R2:.91 NRMSE: 8,25 Jochem.verrelst@uv.es Earsel ISW 9/4/13 11/23

12 LAI User data (SPARC - CHRIS) LR PLS NN KRR GPR LR poor, only acceptable with high training. PLS somewhat better, but not excellent performances. Needs noise. NN behaves erratic: can lead to good performances but unstable. KKR: Excellent performances, very robust. GPR: Excellent performances, robust. SI 3-bands LR: R2:.91 NRMSE: 6,73 Jochem.verrelst@uv.es Earsel ISW 9/4/13 12/23

13 Earsel ISW 9/4/13 13/23 Retrieval Manual options Options to select land cover class, parameter and algorithm. Options to add noise, select user/rtm and train/test data distribution. Selected strategies, from above or imported from earlier test. Plotting options Case studies: Applying GPR to images because of additional features - LCC The same SPARC-trained model has been applied to various CHRIS images. Also HyMap data was processed

14 LCC CHRIS Portability SPARC-trained GPR model Barrax Jul 3 Barrax Jul 4 Barrax Aug 9 Demmin May 6 Demmin June 6 Monegros July 4 Tablas Aug 9 Salldbury May 6

15 Relative uncertainty Absolute uncertainty σ CV: σ µ

16 Earsel ISW 9/4/13 16/23 SPARC HyMap - LCC Estimated PLS Partial least squares regression - LCC Regressor Training [%] RMSE NRMSE R2 Gaussians Processes Regression Kernel ridge Regression Neural Network Partial least squares regression Linear Regression Gaussians Processes Measured Regression - LCC GPR GPR - The lower the sigma the more relevant the band Most relevant bands (σ<3 ): 462, 58, 585, 6, 662, 134, 1461, 1475, 1747, 1759, 1771, 28, 227, 2136, 2241, 2389, 2421, 2437, Estimated Measured Note that PLS performs considerably poorer. GPR insight in relevant bands

17 Earsel ISW 9/4/13 17/23 LCC map µ σ CV: σ µ 7 6 HyMap flight line 5 m resolution 125 bands NRMSE: 3,51% GPR mean estimates, uncertainties and relative uncertainties

18 Earsel ISW 9/4/13 18/23 SPARC HyMap - LAI Estimated NN Neural Network - LAI Regressor Training [%] RMSE NRMSE R2 Neural Network Gaussians Processes Regression Kernel ridge Regression Partial least squares regression Linear Regression GPR Measured Gaussians Processes Regression - LAI GPR - The lower the sigma the more relevant the band Most relevant bands (σ<3): 478, 723, 1257, 1271, 1327, 1795, 187, 197, Estimated Measured

19 Earsel ISW 9/4/13 19/23 LAI map µ σ CV: σ µ HyMap flight line 5 m resolution 125 bands NRMSE: 5,66% Uncertainties poorer; both over vegetated and non-vegetated surfaces.

20 Earsel ISW 9/4/13 2/23 Multi-output CHRIS-SPARC LCC LAI Faster but not necessarily better.

21 Earsel ISW 9/4/13 21/23 model_train Coupling RTM (PROSAIL) with MLRAs Training: 1 random simulations; Validation: SPARC dataset (without bare soil) LR NRMSE_Full_image_LAI_Linear Regression model_train PLS NRMSE_Full_image_LAI_Partial least squares regression Sentinel-2 (1m) model_train KRR NRMSE_Full_image_LAI_Kernel ridge Regression model_train NN NRMSE_Full_image_LAI_Neural Network spect_noise spect_noise spect_noise Coupling so far unsuccessful: Difficulties to deal with large training samples Poor matching: noise needed to enable matching with real data Results poor: better configurations needed NN rather unstable GPR best performing model_train spect_noise GPR NRMSE_Full_image_LAI_Gaussians Processes Regression spect_noise

22 Earsel ISW 9/4/13 22/23 Conclusions Nonparametric regressors powerful retrieval algorithms. They easily outperform parametric regressors (e.g. VI-based). PLS not most powerful. MLRA such as NN, KRR and GPR were best evaluated. GPR a Bayesian regressor; insight in relevant bands and provides uncertainties. MLRA toolbox developed in ARTMO that guides the user through all necessary processing steps. Coupling RTMs with MLRAs possible, but further efforts needed to make it successful.

23 Thanks Earsel ISW 9/4/13 23/23

24 Availability ARTMO is work in progress - beta version Accessible at Valencia University under our supervision. Matlab programmers are encouraged to write their own apps. In turn, a copy can be given. Atmospheric models BRDF apps Temporal domain classifiers Public available after publication (will take some time so far unsuccessful) Jochem.verrelst@uv.es Earsel ISW 9/4/13 24/23