automating, optimizing & simplifying - Image Processing Laboratory (IPL)

From model simulations towards vegetation properties mapping:

automating, optimizing & simplifying 0.7 7 0.6

6.2 5.5 4.7

0.4

3.9 lai

Directional reflectance

0.5

3.1

0.3

2.4 0.2 1.6 0.84

0.1

0.076 0 400

600

800

1000

1200

1400 1600 (nm)

1800

2000

2200

2400

1.1609 20.827340.493860.160279.8267 Cab

J. Verrelst, J.P. Rivera & J. Moreno ISSI Workshop – 21-25 Nov2016

1/55

Background

AUTOMATING RTMs • RTMs • ARTMO/forward • Retrieval toolboxes

OPTIMIZING exploiting spectroscopy data • Retrieval • Band selection • Dimensionality & sample reduction

SIMPLIFYING RTMs

• Global sensitivity analysis • Emulation • Retrieval

2/55

AUTOMATING • RTMs • ARTMO/forward • Retrieval toolboxes

3/55

Radiative transfer models (RTMs) Canopy RTMs

Leaf RTMs

layers

Compact spheres

N-fluxes

Ray tracing

Turbid medium

Hybrid

Geometric

Volumetric

A diversity of RTMs exist with different complexity.

4/55

RTMs are important tools in EO research, but not always easy to use. Only very few of them offer user-friendly interfaces. Which RTM to choose?

Only very few offer a GUI.

(Chuvieco & Prado, 2005)

• No interface exists that brings multiple RTMs together in one GUI. • None of existing (publicly available) GUIs provide post-processing tools. 5/55

Toolbox for EO applications:

Automated Radiative Transfer Models Operator

http://ipl.uv.es/artmo/

6/55

ARTMO

7/55

ARTMO v. 3: modular design

File

Models

Forward

Retrieval

Tools

Help

Load Project New Project DB adminstration Settings Model inputs

Leaf Canopy

Leaf Canopy Combined

Spectral Indices MLRA LUT-based Inversion

Sensor Graphics Spectral resample GSA Emulator

Show Log

Save Load

Combined

New DB Change DB Delete Update LUT class Project Database

PROSPECT 4 PROSPECT 5 DLM LIBERTY Fluspect-B

User’s manual Installation guide Disclaimer GSA configuration GSA results

4SAIL FLIGHT INFORM SCOPE

8/55

Conceptual architecture ARTMO Project Management

Radiative Transfer Models

Data Storage

Retrieval

Tools

9/55

Forward

10/55

RTM outputs only a few clicks away…

Data flow

11/55

ARTMO’s leaf models

PROSPECT-4

DLM

LIBERTY

PROSPECT-5

Fluspect-B

12/55

ARTMO’s canopy models SAIL

INFORM

FLIGHT

13/55

ARTMO’s combined models: SCOPE

14/55

Sensor Simulations can be generated according to band settings of a selected sensor.

• New sensor settings can be imported by clicking on the ‘Import’ button in the top bar. • Existing band settings can be modified or new ones can be added by clicking on the ‘Edit’ button. • Also a spectral filter of a sensor can be imported or viewed by clicking on the ‘Spectral Filter’ button.

Default sensors: • Landsat 7 TM • Landsat 7 ETM+ • SPOT-4 VMI • SPOT-4 HRVIR • CHRIS Mode-3 • MODIS • MERIS

• • • • • •

Sentinel-2 Sentinel-3 OLCI Sentinel-3 SLSTR Landsat 8 Pleiades-1A Quickbird 15/55

Graphics

Visualization options: Export:

Spectral data

Associated metadata

16/55

Retrieval Model

17/55

Retrieval families

Parametric regression

Spectral relationships that are sensitive to specific vegetation properties

Non-parametric regression

Advanced techniques that search for relationships between spectral data and biophysical variables

RTM inversion

Models that simulate interactions between vegetation and radiation leaf

Normalized Difference Vegetation Index canopy

Methods of these different families can be combined: hybrid methods

18/55

ARTMO’s retrieval toolboxes: Spectral indices toolbox

Machine learning regression algorithm toolbox

LUT-based inversion toolbox

Optimizing and generating maps of vegetation properties only a few clicks away…

19/55

General structure: Database [spectral+ variables]

T/V % Training data/ % validation data

Training Spectral data

Validation

Biophysical variables

Noise

Spectral data

Retrieval algorithm

Developed model

Biophysical variables

Goodness-offit statistics

Spectral data

Optimized model

To account for natural variability

Estimated biophysical variable

Biophysical variable map

20/55

OPTIMIZING • Retrieval: parametric/non-parametric/inversion • Band selection • Dimensionality & sample reduction 8 7

LAI Estimated

6 5 4 3 2 1 0 0

1

2

3

4 5 LAI Measured

6

7

8

21/55

1/2

Spectral indices toolbox: Properties:

• Calculates all possible band combinations. • For index formulations with up to 10-band indices (#b10, for a 10-band sensor that would be 10 billion combinations)

• Includes multiple fitting functions (linear, exponential, logarithmic, power, polynomial)

• Noise & Cross-validation options • Results stored in MySQL • Top-performing indices per formulation and fitting function are given. • Can process both image or individual spectra. SPARC – HyMap - LAI

R2

5 s. Best-performing index can be applied to an image.

22/55

SPARC dataset (Barrax Spain); HyMap data

2/2

SR and ND for different fitting functions

Best performing model.

• Which VI method is most correct? • Why restricting to a few bands only?

23/55

1/15

Machine learning regression algorithm toolbox Properties:

About 15 MLRAs implemented Single-output & multi-output Noise & Cross-validation options Dimensionality reduction options Results stored in MySQL GPR properties: band relevance & uncertainties • Can process both images or individual spectra. • Active learning, GPR-BAT, dim. reduction • • • • • •

Simpler to execute than SI: no band selection needed.

Non-parametric models: • SimpleR [Camps-Valls et al., 2013] • http://www.uv.es/gcamps/code/simpleR.html

Also: • Bagging trees (BAGTREE) • Boosting trees (BOOST) • Neural networks (NN) • Extreme Learning Machines (ELM) • Support Vector Regression (SVR) • Relevance Vector Machine (RVM) • Variational Heteroscedastic Gaussian Process Regression (VHGPR)

GPR in Bayesian framework also provides: • Band relevance • Uncertainty estimates

24/55

CHRIS

GPR maps LCC [µg/cm2]

St Dev LCC [µg/cm2]

2/15

CASI1500 - pixel size: 1.4m - 288 bands Same GP model was applied

Uncertainty 25/55

3/15

LCC [µg/cm2]

26/55

LCC

[µg/cm2]

St Dev LCC [µg/cm2]

4/15

RGB CASI

• Uncertainty maps provide additional info which may be hidden on the images. • Implausible estimations are detected.

27/55

LCC [µg/cm2]

St Dev LCC [µg/cm2]

5/15

RGB CASI

• •

In turn, despite low uncertainties also good estimations. No impact on recently irrigated areas (other methods have difficulties with wet soils). For operational applications, of interest to flag/mask regions with high uncertainties. 28/55

GPR to S2 GPR Retrievals with uncertainties