No abstract

No abstract

Next day wildfire prediction is an open research problem with significant environmental, social, and economic impact since it can produce methods and tools directly exploitable by fire services, assisting, thus, in the prevention of fire occurrences or the mitigation of their effects. It consists in accurately predicting which areas of a territory are at higher risk of fire occurrence each next day, exploiting solely information obtained up until the previous day. The task’s requirements in spatial granularity and scale of predictions, as well as the extreme imbalance of the data distribution render it a rather demanding and difficult to accurately solve the problem. This is reflected in the current literature, where most existing works handle a simplified or limited version of the problem. Taking into account the above problem specificities, in this paper, we present a machine learning methodology that effectively (sensitivity > 90%, specificity > 65%) and efficiently performs next day fire prediction, in rather high spatial granularity and in the scale of a country. The key points of the proposed approach are summarized in: (a) the utilization of an extended set of fire driving factors (features), including topography-related, meteorology-related and Earth Observation (EO)-related features, as well as historical information of areas’ proneness to fire occurrence; (b) the deployment of a set of state-of-the-art classification algorithms that are properly tuned/optimized on the setting; (c) two alternative cross-validation schemes along with custom validation measures that allow the optimal and sound training of classification models, as well as the selection of different models, in relation to the desired trade-off between sensitivity (ratio of correctly identified fire areas) and specificity (ratio of correctly identified non-fire areas). In parallel, we discuss pitfalls, intuitions, best practices, and directions for further investigation derived from our analysis and experimental evaluation.

Forest fires in recent years are becoming increasingly devastating for ecosystems, human lives and infrastructures as they follow the climate change impact. In this context the fire monitoring and risk prediction is crucial to support Civil Protection Agencies in charge of the protection of natural ecosystems against fires. Embracing the advancements in remote sensing the fire monitoring task is more and more contributed from automated systems that exploit satellite sensors data, while in the fire risk prediction field, machine learning tends to become the most applied methodology. In this short manuscript we briefly present the development of a daily wildfire risk prediction model based on machine learning techniques and a monitoring system (Forest Fire Information System) for active fires and burn scar mapping that exploits MODIS and VIIRS remote sensing data.

<p><strong>Introduction</strong>. This abstract presents FireRisk (https://riskmap.beyond-eocenter.eu/), a web platform that produces and visualizes timely, highly granular and accurate next day fire risk predictions on a country scale. FireRisk deploys a thorough data fusion process, and a state of the art machine learning (ML) pipeline, considering a large set of fire driving factors, in order to train scalable and accurate models for next day fire prediction. On top of them, it implements a web service that supports the visualization of fire risk predictions and metadata on a user friendly, map-based web application. </p><p><strong>The FireRisk platform</strong>. The high-level architecture is depicted in the following figure. It comprises three major components.</p><p><img src="https://contentmanager.copernicus.org/fileStorageProxy.php?f=gepj.db6dd3fd5db369374533761/sdaolpUECMynit/32UGE&app=m&a=0&c=1015ea587a23a3396774dee02242f054&ct=x&pn=gepj.elif&d=1" alt=""></p><p>(a) Data fusion: This component implements the collection, preprocessing, curation and harmonization of data, leading to the generation of a rich feature set of factors that affect fire occurrence and spread. 25 fire influencing factors were considered, including topography-related, meteorology-related, Earth Observation (EO) derived variables, and historical fire occurrence information. These have been extensively documented in [1].</p><p>(b) ML model learning: This component implements a complete ML pipeline, that includes training and comparison of various ML algorithms, hyperparameter tuning and model (cross-)validation and selection. This pipeline allows the configurable production of robust ML models for fire risk prediction. It is extensively documented in [2].</p><p>(c) Web platform: This component provides an interactive daily fire risk map to users through a web interface. The user is able to view the next day fire risk predictions for the current, as well as for historical days. The predictions are depicted in a five-grade scale (from<em> very low</em> to <em>very high</em>) adopting a five-grade coloring (<em>blue</em> to <em>red</em>). The user is also able to seamlessly change the zoom level, from the whole country level, to individual fine grained areas (grid cells 500m wide), for which individual predictions are provided.<strong> </strong>Finally, the web interface can be displayed on mobile devices, where the user can additionally view their position on the map.</p><p>The risk map visualization functionality is implemented through a Web Map Service (WMS) that is configured on a GeoServer back-end installation. The daily map is stored in PostgreSQL as a raster image, using the geospatial extension PostGIS. For implementing we engage the WMS GeoServer’s capability to convert PostGIS geospatial tables to WMS.</p><p><strong>Ongoing work</strong>. Our ongoing work focuses on two directions: (a) We are adapting Deep Learning algorithms (Siamese Neural Networks and Semantic Segmentation CNNS), to better handle the extreme imbalance and the strong spatio-temporal correlations in the data. (b) We are incorporating explainability mechanisms that will allow the end user of the web application to receive simple and intuitive explanations on each individual prediction visualized on the map, based on the underlying fire driving factors.</p><p>1. Girtsou, S. et al.. A Machine Learning methodology for next day wildfire prediction. In IGARSS, 2021.</p><p>2. Apostolakis, A.; et al. Estimating Next Day’s Forest Fire Risk via a Complete Machine Learning Methodology. In Remote Sens. 2022. https://doi.org/10.3390/rs14051222</p>