Published February 17, 2026 | Version v1
Software Open

WP4 - supplementary data - Using deep learning to assimilate sun-induced fluorescence satellite observations in the ISBA land surface model: model

  • 1. ROR icon Météo-France

Description

Here are the trained weights of the following neural network that maps LAI, Latitude, Longitude and day of year to the SIF in keras format.

The model summary is as follow:

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, 128)               640       
                                                                 
 gaussian_noise (GaussianNo  (None, 128)               0         
 ise)                                                            
                                                                 
 activation (Activation)     (None, 128)               0         
                                                                 
 batch_normalization (Batch  (None, 128)               512       
 Normalization)                                                  
                                                                 
 dense_1 (Dense)             (None, 128)               16512     
                                                                 
 batch_normalization_1 (Bat  (None, 128)               512       
 chNormalization)                                                
                                                                 
 dense_2 (Dense)             (None, 1)                 129       
                                                                 
=================================================================
Total params: 18305 (143.01 KB)
Trainable params: 17793 (139.01 KB)
Non-trainable params: 512 (4.00 KB)
_________________________________________________________________

The two first activation layers are ReLu and the last is linear.

The inputs are in order: DOY, LAT, LON, LAI

The output is the SIF_L described in the article. Create a repository to save all the three files provided.

To load the model, you have to recreate the neural network architecture summarised above and to create a new keras empty model based on this.

then use the command: your_model.load_weights('your_repository/corrected_128_mdl.ckpt')

The weights are now loaded in the new model.

Files

Files (438.2 kB)

Name Size Download all
md5:5e7dce44fd55f6f761da0f202e4f5a0c
101 Bytes Download
md5:6ba5479f4889b88caa63d0d7a056cb01
435.8 kB Download
md5:17a520a2036837a6da9b8a2fe6041b7d
2.2 kB Download

Additional details

Funding

European Commission
CORSO - CO2MVS Research on Supplementary Observations 101082194