Mikko Strahlendorff ^1* Anni Kröger ^1* Golda Prakasam ¹ Miriam Kosmale ¹ Mikko Moisander ¹ Heikki Ovaskainen ² Asko Poikela ²

• ¹ Finnish Meteorological Institute, Helsinki, Finland
• ² Metsäteho Oy, Vantaa, Finland

Soil wetness forecasts on a local level are needed to ensure sustainable forestry operations during summer when the soil is neither frozen nor covered with snow. Training gradient boosting models has been successful in predicting satellite observation-based products into the future using Numerical Weather Prediction (NWP) and Earth Observation (EO) climate data as inputs. The Copernicus Global Land Monitoring Service’s Soil Water Index (SWI) satellite-based observations from 2015 to 2023 at 10 000 locations in Europe were used as the predictand (target parameter) to train an artificial intelligence (AI) model to predict soil wetness with XGBoost (eXtreme Gradient Boosting) and LightGBM (Light Gradient Boosting Machine) implementations of gradient boosting algorithms. The locations were selected as a representative set of points from the Land Use/Cover Area Frame Survey (LUCAS) sites, which helped evaluate the characteristics of different locations used in fitting to represent diverse landscapes across Europe. Over 40 predictors, mainly from ERA5-Land reanalysis, were used in the final model. Over 70 predictors were tested, including the climatology of EO based predictors like SWI and Leaf-Area Index (LAI). The final model achieved a mean absolute error of 5.5% and a root mean square error of 7% for a variable values ranging from 0-100%, an accuracy sufficient for forestry use case. To further validate the model, SWI prediction was made using the 215-day seasonal forecast ensemble from April 2021, consisting of 51 members. With this, the quality could also be demonstrated in the way our forestry climate service (HarvesterSeasons.com) would use the forecasts. As soil wetness is not changing as rapidly as many weather parameters, the forecast skill appears to last longer for it than for the weather variables. The code for the machine learning work and the predictions are available as open source at https://github.com/fmidev/ml-harvesterseasons (see README-SWI2). The training data and ML models are at https://destine.data.lit.fmi.fi/soilwater/. All data used for predictions are accessible from the SmartMet server at https://desm.harvesterseasons.com/grid-gui and the work flow is available in the script https://github.com/fmidev/harvesterseasons-smartmet/blob/master/bin/get-seasonal.sh Everything is made available for ensuring reproducibility. One will need to register and use their own https://cds.climate.copernicus.eu credentials for doing so.