All technical data shown has been generalized/anonymized for public representation.

Final-year project · HKU · 2023

Applying machine learning algorithms to enhance wildfire prediction in Alaska

Supervised learning with low computational power to forecast next-day natural wildfire occurrence from ERA5-Land weather data — XGBoost reached 82.9% accuracy in under two minutes.

9%XGBoost accuracy
6 sXGBoost runtime
kmSpatial resolution
2010–2020Study period

Challenge

Wildfires have tremendous impacts on global biodiversity and health, and lead to losses of lives and property [1]. In 2021, wildfires and other natural catastrophes caused 270 billion USD in loss [2]. Under climate change, the frequency of wildfires is expected to rise [3], and there is an increasing need to understand and forecast them [4]. Recently, deep-learning algorithms have been introduced for prediction [5]; however, they tend to be computationally demanding and time-consuming [6][7].

In this work, supervised machine-learning algorithms with lower computational power are implemented to predict natural wildfire occurrences in Alaska with a lead time of 1 day, based on weather conditions.

Method & Standards

🗺️ Study area & period

Fig. 1 — Spatial density of naturally occurring wildfires in Alaska from 2010 to 2020. The study area is indicated by the black rectangular box.

Fig. 2 — Frequency of wildfire occurrences by month in Alaska from 2010 to 2020

Situated at high latitude, Alaska is particularly vulnerable to climate change [8]. This work studies the region and months with the highest wildfire intensities from 2010 to 2020: 62°N–67°N, 141°W–159°W, during May to August. The target data format is gridded data with a spatial resolution of 9 km and a daily temporal resolution.

📚 Datasets

MODIS Burned Area Product discussion on the NASA Earthdata Forum — MODIS Burned Area Product. Source: NASA Earthdata Forum.

Dataset	Spatial resolution	Temporal resolution	Spatial data type
ERA5-Land	9 km	Hourly	Gridded
MODIS Burned Area Product	2.3 km	11 days	Multipolygon

ERA5-Land provides hourly humidity, temperature, precipitation, and wind speed. MODIS Burned Area Product and spatial wildfire-occurrence data for the United States supply fire history and burned-area information.

🧮 Data preprocessing

Variables d2m, t2m, skt, e, tp, and 10 m wind (w) are resampled from hourly to daily form. Burned areas are intersected with US wildfire-occurrence data to determine the start date of each fire; fire perimeter data are transformed from multipolygon to gridded data via QGIS overlap analysis with a custom Python plugin.

📈 Attribute analysis & data cube

Distribution of a weather variable in burned vs unburned scenarios — Fig. 5a — Weather variable distribution: burned vs. unburned

Structure of the data cube for model training — Fig. 6 — The structure of the data cube for model training

Before model training, attribute analysis identifies the significant values of each weather variable for distinguishing burned from unburned cells the next day, minimizing computational power. At 9 km resolution, the prediction of next day's burned value (0 or 1) is based on: min evaporation (mine), mean wind speed (meanw), max 2 m temperature (maxt2m), max skin temperature (maxskt), max 2 m dewpoint temperature (maxd2m), and mean total precipitation (meantp). As the dataset is imbalanced, random undersampling of non-burned cells is carried out to avoid overfitting — the final ratio of burned to unburned cells is 1:1. The dataset is split 60%/40% train/test and a 5-fold cross-validation is performed for each model, with random undersampling each time.

Visual / Demo

Model	Mean CV Score (%)	Accuracy (%)	Precision (%)	Recall (%)	F1 (%)	Execution Time (s)
Decision Tree	79.86	80.12	77.98	82.69	80.26	114.78
XGBoost	81.86	82.90	80.24	86.26	83.14	68.56
Support Vector Machine	71.82	71.26	67.48	79.57	73.03	59.67
Random Forest	75.80	76.57	73.13	82.34	77.46	64.90
Naive Bayes	71.44	72.59	67.91	83.29	74.82	49.33

Metrics: precision = TP/(TP+FP), recall = TP/(TP+FN), F1 = 2·precision·recall/(precision+recall), accuracy = (TP+TN)/(TP+FN+TN+FP).

Impact

• Within two minutes, simple machine-learning models — particularly XGBoost — can predict next-day wildfire occurrence with an accuracy of 82.9% using weather conditions.
• Low-power ML algorithms can serve as fast, efficient alternatives for predicting natural hazards under climate change, potentially enhancing climate resilience.
• Improvements: longer lead times, higher spatial resolution, and static variables such as slope and vegetation index to improve accuracy.

References
[1] Westerling, A.L. (2007). Climate change and wildfire in California. Climatic Change 87(S1), 231–249. doi.org/10.1007/s10584-007-9363-z
[2] Bevere, L. (2023). Natural catastrophes in 2021. Swiss Re sigma. swissre.com
[3] Ghorbanzadeh, O. (2018). Wildfire susceptibility evaluation… Advances in Science, Engineering and Technology.
[4] Taylor, S.W. (2013). Wildfire prediction to inform fire management. Statistical Science 28(4). doi.org/10.1214/13-sts451
[5] Arinta, R.R. (2019). Natural disaster application on Big Data and machine learning: A Review. ICITISEE. doi.org/10.1109/icitisee48480.2019.9003984
[6] Liang, H. (2019). A neural network model for wildfire scale prediction. IEEE Access 7. doi.org/10.1109/access.2019.2957837
[7] Jaafari, A. (2019). Hybrid artificial intelligence models… Agricultural and Forest Meteorology 266-267. doi.org/10.1016/j.agrformet.2018.12.015
[8] Markon, C. (2018). Chapter 26: Alaska. Fourth National Climate Assessment, volume II. doi.org/10.7930/nca4.2018.ch26