Forecasting the moisture content of live forest fuels with a transformer machine learning model and Sentinel-2 satellite data

Ivan Kotzur , Matthias Boer, Vitus Benson, Rachael Nolan

Live fuel moisture content (LFMC) provides an indicator of vegetation flammability, the potential for canopy dieback from drought, and habitat quality for arboreal fauna. Recent advances in artificial intelligence have increased the potential in meteorological forecasts for Earth surface forecasting. Combining weather forecasts with remotely sensed LFMC offers a promising avenue for short and medium-term LFMC forecasting (i.e. weeks to months). We adapted a transformer-based surface forecasting model to predict LFMC across the forest regions of New South Wales, Australia, up to 50 days in advance. Using a multi-modal spatiotemporal transformer (Contextformer) as a hind-cast model of LFMC with ‘minicubes’ of weather, soil, vegetation structure, elevation and Sentinel-2 reflectance data, we tested four variants to evaluate if theory-guided biophysical predictors improve performance compared to the base model. Explainable AI methods were applied to demystify model reliance, using Integrated Gradients attribution. Incorporating biophysical theory in predictor selection produced the lowest error (Global RMSE=15.68% dry matter), whereas adding only spectral bands to the base model was less successful (partly due to the structure of the forecast architecture). Within the LFMC forecast, explainable AI revealed mixed patterns in importance amongst temperature and rainfall variables due to the transformer’s ability to model multiple eco-physiological pathways to vegetation hydration. Dynamic weather variations dominate forecasted LFMC temporal trends, over spectral bands. Crucially, regression-to-the-mean effects in the LFMC forecasts caused poorer local performance (median R2≈0.21) compared to global evaluation (R2≈0.69, i.e. poorer within minicubes than amongst minicubes). When converting this hind-cast model to an operational model (i.e. true forecast), best performance could be utilised by running rolling short-term forecasts (5 to 15 days) to capture localised fuel hazard peaks and canopy dieback in the future.