Deep Learning Surrogate Models of JULES-INFERNO for Wildfire Prediction on a Global Scale

Global wildfire models play a crucial role in anticipating and responding to\nchanging wildfire regimes. JULES-INFERNO is a global vegetation and fire model\nsimulating wildfire emissions and area burnt on a global scale. However,\nbecause of the high data dimensionality and system complexity, JULES-INFERNO's\ncomputational costs make it challenging to apply to fire risk forecasting with\nunseen initial conditions. Typically, running JULES-INFERNO for 30 years of\nprediction will take several hours on High Performance Computing (HPC)\nclusters. To tackle this bottleneck, two data-driven models are built in this\nwork based on Deep Learning techniques to surrogate the JULES-INFERNO model and\nspeed up global wildfire forecasting. More precisely, these machine learning\nmodels take global temperature, vegetation density, soil moisture and previous\nforecasts as inputs to predict the subsequent global area burnt on an iterative\nbasis. Average Error per Pixel (AEP) and Structural Similarity Index Measure\n(SSIM) are used as metrics to evaluate the performance of the proposed\nsurrogate models. A fine tuning strategy is also proposed in this work to\nimprove the algorithm performance for unseen scenarios. Numerical results show\na strong performance of the proposed models, in terms of both computational\nefficiency (less than 20 seconds for 30 years of prediction on a laptop CPU)\nand prediction accuracy (with AEP under 0.3\\% and SSIM over 98\\% compared to\nthe outputs of JULES-INFERNO).\n