Supplementary data : Using artificial intelligence to monitor live fuel moisture content across France, based on a high-resolution land surface analysis

Co-authors : Pierre Vanderbecken¹, Bertrand Bonan¹, Catherine Robert², Mathieu Regimbeau³, François Pimont⁴, Kevyn Raynal⁴, Xiangzhuo Liu⁵, Remi Savazzi⁶, Moncef Garouani⁷, Josiane Mothe⁸, Nemesio Rodriguez-Fernandez⁹, Lionel Jarlan⁹, Jean-Christophe Calvet¹ ¹Météo-France, CNRS, Univ. Toulouse, CNRM, Toulouse, France²Météo-France, Direction des Opérations pour la Prévision, Coordination nationale feux de végétation, Toulouse, France³Météo-France, Direction des Services Météorologiques, Division Agrométéorologie, Toulouse, France⁴INRAE, URFM, Avignon, France ⁵INRAE, UMR1391 ISPA, Villenave d’Ornon, France⁶ONF, pôle national DFCI, Direction territoriale Midi-Méditerranée, Aix-en-Provence, France⁷IRIT, Univ. Toulouse Capitole, UMR5505 CNRS, Toulouse, France⁸IRIT, UT2J, Univ. de Toulouse, UMR5505 CNRS, Toulouse, France⁹CESBIO, University of Toulouse, IRD/CNRS/UPS/CNES, Toulouse, France Abstract : This study develops a daily Live Fuel Moisture Content (LFMC) product to aid fire danger management in France. Utilising in situ observations, the product employs a neural network for real-time vegetation stress monitoring and combines land surface model outputs with a Land Data Assimilation System. Three cross-validation methods and sensitivity analyses ensure model robustness and spatial consistency while pinpointing areas for improvement.   The repository contains all the data required to reproduce the results and illustrations displayed in the associated paper. It also contains the weights of the neural network. It contains nine files: four in CSV format, three in netCDF4 format, the weights in KERAS format, and all the LFMC maps from 2017 to 2024 (between 1 May and 30 September) in zip format. Below is a brief description of each file: 'CV_val.csv': contains scores from the FC and LOYO (cross-validation scheme), as well as RMSE, NSE, KGE and R values over time. 'Fire_occ.csv': contains the cantonal-averaged LFMC value from 10 days before to 5 days after ignition. The 2,651 fires are listed using Fire ID. 'LOSO_val.csv': contains scores from the LOSO experiment for Les Adrechs, Forêt de la Coubre and La Chapelle-Moulière.  'Site' defines the studied site, 'lfmc' is the observed LFMC value and 'lfmc_pred' is the prediction. 'Neutral.csv': contains the normalised scores from the neutralisation of the predictor experiments. All R, SSD, KGE and bias have been normalised for comparison. 'Deployment.nc': is a map of the AROME domain showing the locations of the 'Réseau Hydrique' in 2024 and the areas where observations are being added. 'ISBA_LDAS_AROME_LFMC_20220813.nc': is a map of the AROME domain showing all the predictors, LFMC predictions, standard deviation, and the difference between the analysis and open loop on 13 August 2022. 'ISBA_LDAS_AROME_LFMC_20230813.nc': is a map of the AROME domain showing all the predictors, LFMC predictions, standard deviation, and the difference between the analysis and open loop on 13 August 2023. 'Model_Article.keras': contains all the weights necessary to compute the model. The inputs are in order:   Dyn_in : ['Leaf Biomass', 'LAI/WFC', 'Fuel', 'SWI'] Met_in : ['Rugosity', 'Slope', 'SAND', 'Altitude', 'Doy_sin', 'Doy_cos'] 'netcdf_LFMC.zip': contains all LFMC maps over the AROME domain. The model summary is as follow:Model: "functional" ┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩ │ input_layer │ (None, 4) │ 0 │ - │ │ (InputLayer) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dense (Dense) │ (None, 40) │ 200 │ input_layer[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ activation │ (None, 40) │ 0 │ dense[0][0] │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout (Dropout) │ (None, 40) │ 0 │ activation[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ input_layer_1 │ (None, 6) │ 0 │ - │ │ (InputLayer) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ concatenate │ (None, 46) │ 0 │ dropout[0][0], │ │ (Concatenate) │ │ │ input_layer_1[0]… │ ──────────────────────────────────────────────────────────────────────────── │ dense_1 (Dense) │ (None, 32) │ 1,504 │ concatenate[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ activation_1 │ (None, 32) │ 0 │ dense_1[0][0] │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_1 (Dropout) │ (None, 32) │ 0 │ activation_1[0][… │ ──────────────────────────────────────────────────────────────────────────── │ dense_2 (Dense) │ (None, 16) │ 528 │ dropout_1[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ activation_2 │ (None, 16) │ 0 │ dense_2[0][0] │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_2 (Dropout) │ (None, 16) │ 0 │ activation_2[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_1 │ (None, 22) │ 0 │ dropout_2[0][0], │ │ (Concatenate) │ │ │ input_layer_1[0]… │ ──────────────────────────────────────────────────────────────────────────── │ dense_3 (Dense) │ (None, 8) │ 184 │ concatenate_1[0]… │ ──────────────────────────────────────────────────────────────────────────── │ activation_3 │ (None, 8) │ 0 │ dense_3[0][0] │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_3 (Dropout) │ (None, 8) │ 0 │ activation_3[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_2 │ (None, 14) │ 0 │ dropout_3[0][0], │ │ (Concatenate) │ │ │ input_layer_1[0]… │ ──────────────────────────────────────────────────────────────────────────── │ dense_4 (Dense) │ (None, 4) │ 60 │ concatenate_2[0]… │ ──────────────────────────────────────────────────────────────────────────── │ batch_normalization │ (None, 4) │ 16 │ dense_4[0][0] │ │ (BatchNormalizatio… │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ activation_4 │ (None, 4) │ 0 │ batch_normalizat… │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_4 (Dropout) │ (None, 4) │ 0 │ activation_4[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_3 │ (None, 12) │ 0 │ dropout_4[0][0], │ │ (Concatenate) │ │ │ dropout_3[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ dense_5 (Dense) │ (None, 8) │ 104 │ concatenate_3[0]… │ ──────────────────────────────────────────────────────────────────────────── │ batch_normalizatio… │ (None, 8) │ 32 │ dense_5[0][0] │ │ (BatchNormalizatio… │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ activation_5 │ (None, 8) │ 0 │ batch_normalizat… │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_5 (Dropout) │ (None, 8) │ 0 │ activation_5[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_4 │ (None, 24) │ 0 │ dropout_5[0][0], │ │ (Concatenate) │ │ │ dropout_2[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ dense_6 (Dense) │ (None, 16) │ 400 │ concatenate_4[0]… │ ──────────────────────────────────────────────────────────────────────────── │ batch_normalizatio… │ (None, 16) │ 64 │ dense_6[0][0] │ │ (BatchNormalizatio… │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ activation_6 │ (None, 16) │ 0 │ batch_normalizat… │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_6 (Dropout) │ (None, 16) │ 0 │ activation_6[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_5 │ (None, 22) │ 0 │ input_layer_1[0]… │ │ (Concatenate) │ │ │ dropout_6[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ dense_7 (Dense) │ (None, 32) │ 736 │ concatenate_5[0]… │ ──────────────────────────────────────────────────────────────────────────── │ batch_normalizatio… │ (None, 32) │ 128 │ dense_7[0][0] │ │ (BatchNormalizatio… │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ activation_7 │ (None, 32) │ 0 │ batch_normalizat… │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_7 (Dropout) │ (None, 32) │ 0 │ activation_7[0][… │ ──────────────────────────────────────────────────────────────────────────── │ concatenate_6 │ (None, 72) │ 0 │ dropout[0][0], │ │ (Concatenate) │ │ │ dropout_7[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ dense_8 (Dense) │ (None, 40) │ 2,920 │ concatenate_6[0]… │ ──────────────────────────────────────────────────────────────────────────── │ batch_normalizatio… │ (None, 40) │ 160 │ dense_8[0][0] │ │ (BatchNormalizatio… │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ activation_8 │ (None, 40) │ 0 │ batch_normalizat… │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── │ dropout_8 (Dropout) │ (None, 40) │ 0 │ activation_8[0][… │ ──────────────────────────────────────────────────────────────────────────── │ dense_9 (Dense) │ (None, 1) │ 41 │ dropout_8[0][0] │ ──────────────────────────────────────────────────────────────────────────── │ activation_9 │ (None, 1) │ 0 │ dense_9[0][0] │ │ (Activation) │ │ │ │ ──────────────────────────────────────────────────────────────────────────── Total params: 7,077 (27.64 KB) Trainable params: 6,877 (26.86 KB) Non-trainable params: 200 (800.00 B)The activations layers are ReLu and the last is linear.--------------------------------------------------------------------------------The PhD work of Yann Baehr was supported by the Occitanie Region as part of the Earth Observation and Territories in Transition Key Challenge (O3T) and by Météo-France. This research received funding from European Union Horizon Europethrough the CORSO project (grant agreement 101082194) and the GREENEO project (grant agreement 101183071). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or the Commission. Neither the European Union nor the granting authority can be held responsible for them. The authors sincerely thank the colleagues, Marius Cadier, Bernard Chapnik, and Laura Pavan at Météo-France for their assistance in the operational implementation of the model during summer 2025. We also acknowledge the colleagues, Valentin Durinck, Elodie Beaumont and Jean-Luc Kicin at the ONF for their support in the deployment and integration of the system into operational activities.