Forest fire risk prediction using deep learning and machine learning algorithms

Forests are vital elements for living things in terrestrial ecosystems and ensure the integrity and sustainability of natural factors such as soil, water and climate. Forest fires are one of the disasters that cause ecosystem degradation as well as social and economic impacts. In this study, the factors affecting forest fires were determined as 21 independent variables under the headings of anthropogenic, meteorological, topographical and vegetation using RF and GIS techniques. The dependent variable, past forest fires, and the independent variables were classified and mapped at monthly temporal resolution from May to September for 7 years between 2017 and 2023. By extracting values from the produced maps, the first data set consisting of 670*22 rows*columns and the second data set consisting of 2100*22 rows*columns were created for model training and the study area geospatial data set consisting of 689419*21 rows*columns was created for risk estimation. The data sets were visualized with Exploratory Data Analysis techniques using Python language in the Google Colab interface, and correlation analyses were performed. After the model training was completed, the prediction phase was executed and risk maps were produced with random forest (0.94) and XGBoost (0.95) algorithms, which achieved the best accuracy metrics among 7 machine learning algorithms. With the sequential model, which is a Deep Learning model, a risk prediction map was produced from the trained model with a test accuracy of 0.95 using Dataset II. As a result of this study, it was revealed that more accurate and meaningful results can be obtained by increasing the size of the data set used for forest fire risk prediction with independent variables that can be associated with forest fires, and that algorithms based on ensemble models such as random forest and extreme gradient boosting are more successful in forest fire risk prediction.