Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OpFML: Pipeline for ML-based Operational Forecasting

Shahbaz Alvi, Giusy Fedele, Gabriele Accarino, Italo Epicoco, Ilenia Manco, Pasquale Schiano

TL;DR

OpFML addresses the gap between ML model development and operational wildfire risk forecasting by introducing a configurable pipeline for periodic ML-based forecasts. The method combines a TOML-configured DataStore, on-the-fly PreProcessing, and a ConvLSTM-based predictor to produce daily Fire Danger Index estimates, demonstrated across Southern Italy and Central Portugal. Key contributions include a flexible, robust, container-ready framework that can interchange ML models and data sources and supports EO and non-EO predictors with weather inputs from WRF_2km@CMCC. The work offers a practical, reusable blueprint for operational ML forecasting in climate and Earth sciences, with potential extension to other pilot sites or variables.

Abstract

Machine learning is finding its application in a multitude of areas in science and research, and Climate and Earth Sciences is no exception to this trend. Operational forecasting systems based on data-driven approaches and machine learning methods deploy models for periodic forecasting. Wildfire danger assessment using machine learning has garnered significant interest in the last decade, as conventional methods often overestimate the risk of wildfires. In this work, we present the code OpFML: Operational Forecasting with Machine Learning. OpFML is a configurable and adaptable pipeline that can be utilized to serve a machine learning model for periodic forecasting. We further demonstrate the capabilities of the pipeline through its application to daily Fire Danger Index forecasting and outline its various features.

OpFML: Pipeline for ML-based Operational Forecasting

TL;DR

OpFML addresses the gap between ML model development and operational wildfire risk forecasting by introducing a configurable pipeline for periodic ML-based forecasts. The method combines a TOML-configured DataStore, on-the-fly PreProcessing, and a ConvLSTM-based predictor to produce daily Fire Danger Index estimates, demonstrated across Southern Italy and Central Portugal. Key contributions include a flexible, robust, container-ready framework that can interchange ML models and data sources and supports EO and non-EO predictors with weather inputs from WRF_2km@CMCC. The work offers a practical, reusable blueprint for operational ML forecasting in climate and Earth sciences, with potential extension to other pilot sites or variables.

Abstract

Machine learning is finding its application in a multitude of areas in science and research, and Climate and Earth Sciences is no exception to this trend. Operational forecasting systems based on data-driven approaches and machine learning methods deploy models for periodic forecasting. Wildfire danger assessment using machine learning has garnered significant interest in the last decade, as conventional methods often overestimate the risk of wildfires. In this work, we present the code OpFML: Operational Forecasting with Machine Learning. OpFML is a configurable and adaptable pipeline that can be utilized to serve a machine learning model for periodic forecasting. We further demonstrate the capabilities of the pipeline through its application to daily Fire Danger Index forecasting and outline its various features.
Paper Structure (33 sections, 1 equation, 20 figures, 3 tables)

This paper contains 33 sections, 1 equation, 20 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Pipeline Layout
  3. Motivation
  4. Structure of the pipeline
  5. Layout
  6. Data Consumption
  7. Data pre-processing
  8. Robustness of the pipeline
  9. The daily FDI use case
  10. Geographic region
  11. Southern Italy
  12. Central Portugal
  13. ML architecture and training dataset
  14. Fire predictors
  15. Data scaling
  16. ...and 18 more sections

Figures (20)

  • Figure 1: Cumulative burned areas (top) and cumulative number of wires (bottom) in 2025 (until September) compared to the average from 2006 to 2024. Source EFFIS fire alert system CopernicusEFFIS_SeasonalTrend
  • Figure 2: Schematic flow chart of the pipeline.
  • Figure 3: TOML configuration for the consumption of the NDVI variable. The TOML subsection called "gathering" describes the request to be submitted to the DDS. In the "processing" section, the transformations to be applied to the variable are described together with the input required by each function.
  • Figure 4: A section of the pilot TOML file describing the data store class and processing class to be used for the pilot site. Additionally, parameters can also be defined depending on the use case.
  • Figure 5: Code snippet of the transformation function in PreProcessingUtils class for computing the wind speed using library xclim, and function uas_vas_2sfcwind. The input xarray Dataset is expected to contain the wind speed components.
  • ...and 15 more figures