PAR algorithms evaluation

Edouard Leymarie�Email :  edouard.leymarie@imev-mer.fr�

E. Leymarie, S. Bartoloni, T. Maurer, Y. Takeshita, A. Poteau, Catherine Schmechtig, H. Claustre

Models by : J. Pitarch, E. Organelli, J. Tan and R. Frouin

​

ADMT-26

Introduction

The importance of PAR measurements

PAR needed for a range of applications :

• Primary production (NPP) calculation through bio-optical models;
• Derivation of the euphotic zone depth (Zeu : depth 1% of the surface PAR)
• Chla processing: Xing et al. (2012 & 2018) & Terrats et al. (2020) => Z_IPAR15
• ….

AST-24

2

PAR models

Two models to retrieve PAR from Irradiance at 380, 443, 490 and 555nm

• Accurate estimation of photosynthetic available radiation from multispectral downwelling irradiance profiles

Pitarch, J., Leymarie, E., Vellucci, V., Massi, L., Claustre, H., Poteau, A., Antoine, D., Organelli, E., 2025. Limnology and Oceanography: Methods. https://doi.org/10.1002/lom3.10673

Code : Matlab or R

​

​

• Modeling underwater photosynthetically available radiation profiles from biogeochemical Argo floats using multi-spectral irradiance measurements

J. TAN, R. FROUIN, E. LEYMARIE, and B. G. MITCHELL. https://doi.org/10.1364/OE.566083

Code : Python

3

PAR models

Jaime Pitarch et al.

In situ Hyperspectral data

Ed 380, 443, 490 & 555 + PAR

Two layers neural network (MATLAB®)

Inputs : 4 Ed + Depth

Output : PAR

Jing Tan, Robert Frouin et al.

simulations of solar irradiance at surface (ARTDECO)

+ HYDROLIGHT to simulate Ed(λ, z)

Ocean vertically homogeneous (3320 cases) and heterogeneous (3320 cases)

General Additive Model (GAM)

Inputs : 4 Ed + Depth

Output : PAR

4

PAR models

Jaime Pitarch et al.

In situ Hyperspectral data

Ed 380, 443, 490 & 560 + PAR

Two layers neural network (MATLAB®)

Inputs : 4 Ed + Depth

Ouput : PAR, PAR_unbiased

simulations of solar irradiance at surface (ARTDECO)

+ HYDROLIGHT to simulate Ed(λ, z)

Ocean vertically homogeneous (3320 cases) and heterogeneous (3320 cases)

General Additive Model (GAM)

Inputs : 4 Ed + Depth

Ouput : PAR

Jing Tan, Robert Frouin et al.

5

PAR uncertainties

• BGC implementation plan based on manufacturer's claims : 3% (in air)
• Typical variability between sensors (in air, recent calibration) : 5% ⁽¹⁾

What uncertainty target?

1. Field Intercomparison of Radiometer Measurements for Ocean Colour Validation. https://doi.org/10.3390/rs12101587

​

6

PAR uncertainties

• BGC implementation plan based on manufacturer's claims : 3% (in air)
• Typical variability between sensors (in air, recent calibration) : 5% ⁽¹⁾
• User survey on acceptable PAR uncertainties ⁽²⁾ :

What uncertainty target?

1. Field Intercomparison of Radiometer Measurements for Ocean Colour Validation. https://doi.org/10.3390/rs12101587
2. Satellite Radiation Products for Ocean Biology and Biogeochemistry: Needs, State-of-the-Art, Gaps, Development Priorities, and Opportunities https://doi.org/10.3389/fmars.2018.00003

7

PAR uncertainties

• Single sensors use a proprietary combination of filters to match a quantum response.
• Hyperspectral sensors use direct integration

🡺 May introduce differences related to the light spectrum (surface vs. depth)

PAR measurements: Single sensor vs hyperspectral

8

Comparison Dataset

• OCR507 : 380, 443, 490, 510, 560, 665 nm + PAR
• 4 deployments : 2 in Med, 2 in Austral (Kerguelen)
• Half of the profiles come from one float in austral
• Modified models for: 380, 443, 490 & 560
• Not an Argo float (data not in DACs)

​

1- ProVal float

doi: 10.3389/fmars.2018.00437

3- SeaBASS E_d

Hyperspectral

S. Bartoloni

2- New Ramses

floats used by Jaime were excluded from the comparisons.

• Excluded data:
• No depth
• Artic
• Coastal

​

• Most of the data added was in the top 20m and from the Atlantic

​

• Blue circles show location of Seabass data that was excluded from analysis

9

Comparison on PROVAL Dataset

Float lovapm006f in Med

Both models show less noisy, and probably more accurate, PAR data at depth.

10

Comparison on PROVAL Dataset

Less good performance at depth for the second model. But data mainly from one float in Austral.

11

Comparison on RAMSES Dataset

Both models perform very well

12

Comparison on PROVAL Dataset

Estimation of the depth PAR = 15

(Depth used as input for Chla processing)

13

Comparison Dataset

• OCR507 : 380, 443, 490, 510, 560, 665 nm + PAR
• 4 deployments : 2 in Med, 2 in Austral (Kerguelen)
• Half of the profiles come from one float in austral
• Modified models for: 380, 443, 490 & 560
• Not an Argo float (data not in DACs)

​

1- ProVal float

doi: 10.3389/fmars.2018.00437

3- SeaBASS E_d

Hyperspectral

S. Bartoloni

2- New Ramses

floats used by Jaime were excluded from the comparisons.

• Excluded data:
• No depth
• Artic
• Coastal

​

• Most of the data added was in the top 20m and from the Atlantic

​

• Blue circles show location of Seabass data that was excluded from analysis

14

SeaBASS Dataset

S. Bartoloni

15

SeaBASS Dataset

S. Bartoloni

16

failure resistance testing

Simulation of a 50% reduction on a band

17

Data Management for computed PAR

Which model to choose ?

As presented uncertainties in computed PAR are within the expected accuracy of the PAR sensor field measurements, we are advocating for direct insertion into the pre-existing DOWNWELLING_PAR parameter, for floats with OCR 380,443, 490, 555.

​

This will best support the user!

• Jaime model presents slightly better results on ProVal floats but models are comparable on Ramses and SeaBASS dataset

​

• Simulated dataset used to train Robert’s model is more constraint while we have much more Ramses data to update Jaime model

​

🡺 If possible, we recommend to process both and take the mean

We can provide codes in different languages for processing: R, Matlab, Python ?

18

Comparison on RAMSES Dataset

19

Data Management for computed PAR

As presented uncertainties in computed PAR are within the expected accuracy of the PAR sensor field measurements, we are advocating for direct insertion into the pre-existing DOWNWELLING_PAR parameter, for floats with OCR 380,443, 490, 555.

​

This will best support the user!

Nonetheless, it is important for the user to be able to track WMO with computed vs measured PAR.

Thus we propose the following options:

​

1. For any WMO for which PAR is to be calculated at the DAC level using the model and inserted into DOWNWELLING_PAR, then DOWNWELLING_PAR_ADJUSTED must also be filled, where DOWNWELLING_PAR_ADJUSTED = DOWNWELLING_PAR, such that error fields, and SCIENTIFIC_CALIB_* fields can be populated.

​

• SCIENTIFIC_CALIB_COMMENT is to be filled with “PAR_MODELED : DOWNWELLING_PAR_ADJUSTED calculated from DOWN_IRRADIANCE_380, _443, _490, _555 using xxxref” where xxxref reflects the model chosen, either PITARCH25, TAN25 or MEAN_PITARCH25_TAN25

​

• DOWNWELLING_PAR_QC and DOWNWELLING_PAR_ADJUSTED_QC . Two options :
1. Filled with the highest flag of DOWN_IRRADIANCE_380, _443, _490, _555
2. Filled with ‘8’ (value estimated)

20

Thank you