Daily wildfire prediction and key factor analysis throughout the year using machine learning: a case study from Gang Won, Korea

Our research methods include three stages: data collection, model training and evaluation, analysis and visualization. First, we collected wildfire occurrence data and independent variables (e.g. weather, forest-related, socioeconomic factors) at target sites from 2013-2023, and then preprocessed the data before entering it into a machine learning model. Here, interpolation was used to assign weather variables estimated days per Si/gun, and all variables in daily units where static or annual variables (such as CFRT, CLRT, FPDN) were positively filled to represent constant daily values throughout the year. This process ultimately created a dataset in the form of “(Si/Gun Per/Gun) x (by date).)”. Five machine learning models (i.e., logistic regression, XGBoost, LightGBM, Random Forest, and additional trees) were then trained on data from 2013-2022 to optimize performance through parameter tuning. Finally, we evaluated the model's prediction accuracy as a test dataset using previously isolated data in 2023. The evaluation process not only compared quantitative metrics such as accuracy and recall, but also used Shapley Additive Description (SHAP) values to improve the interpretability of the model. Here, we used machine learning packages such as Scikit-Learn in a Python 3.9 anaconda environment. The overall flow of the research can be found in Figure 2.

Data Preprocessing

The data collected here are categorized into weather, forest-related socioeconomic data, and wildfire occurrence records that require different pretreatment due to the various collection methods and units of each data. In particular, for weather data, there are some cases where direct observations are missing due to errors in collection devices and uneven distribution of stations across the region. Here, the average values of data collected at nearby stations were used to replace missing measurement data in a given area and interpolate missing values for a given date by reflecting trends over consecutive days. Regarding precipitation, KMA provided data on daily cumulative precipitation (MM/day), but preliminary analysis revealed that precipitation values are not dependent on wildfire predictions when applied to the model. Here we coded four consecutive levels: 3 (rainy day), 2 (next day), 1 (after 2 days), and 0 (other day). This coding reflects a decrease in the effect of precipitation on fire risk as time passes since the rain event. Conversely, this study took into account various variables such as weather, forest-based, and demographic data, so each variable has a different unit and numerical range. Therefore, standardization techniques must be introduced to prevent the model from becoming sensitive to a particular scale. Specifically, I used the StandardScaler from the Scikit-Learn package. This is converted to mean 0, with a standard deviation of 1 (Equation 1).

Model selection and development

We compared and analyzed representative machine learning techniques to predict daily wildfire outbreaks. For model selection, we considered the following: Logistic regression is traditionally used frequently because of its relatively simple explanatory power, random forests.³²a tree-based ensemble model, xgboost³³extra tree³⁴and lightgbm³⁵. These models predict the likelihood of wildfires as binary classifications (0: non-outbreak, 1: outbreak), and differ in prediction accuracy and ability to interpret characteristics.

Logistic regression is a traditional method for estimating the probability of wildfire occurrence via linear combinations, and used in Korea to identify wildfire risk indexes with meteorological factors as input variables³⁶. Unlike normal linear regression, we use a sigmoid function to output probability values between 0 and 1, and interpret the result in the form of \(\text {p} = 1/(1+{\text {e}}^{ – \text {z}})\). here, \(\ text {z} \) It is a linear combination of independent variables and is the sign of the regression coefficient (\(\ upbeta \)) can be applied to relatively infer the direction of wildfire occurrence from variables.

Xgboost and LightGBM are both ensemble models of the boost family, employing a method in which previous predictors weigh the parts where the classifier is incorrect. xgboost³³ To compensate for the error, we assign weights to false predictions from the previous tree and train the tree in stages. There is a regularization technique to show fast learning speeds and prevent overfitting, showing excellent performance in processing large data sets.

Random forests and extra trees are both ensemble models by creating a large number of decision trees. Random Forest³² During the training process, some of the training data is randomly extracted and candidate variables are randomly selected for each split to construct a tree. Such bagging-based randomization reduces the risk of overfitting in a single tree and ensures stable predictive performance by investigating various variable combinations. Extra tree³⁴ Enhanced randomness when splitting nodes causes candidate node splitting criteria to be randomly selected when they are split every time. This technique allows for increased variance instead of lowering bias, so proper parameter tuning is essential. With LightGBM³⁵histogram-based data structure and leaf growth strategies can be applied, resulting in higher learning speed and accuracy. It is designed to efficiently process large and high-dimensional data, and by simply adjusting the parameters, it exhibits high predictive power in unbalanced data situations.

The aforementioned model minimized log-likelihood loss for binary classification by using independent variables preprocessed as input data during the training phase and updating the parameters to target the dependent variable, wildfire occurrence. After training, after inputting 2023 data, we equip two probability of wildfire occurrence predicted by the model at a threshold of 0.5 and compared with actual observations to assess the classification performance of the final model.

Data imbalance: Oversampling techniques

Research into wildfire occurrences is a so-called disproportionate data problem, and is a serious problem, with the number of “non-escaping (zero)” cases of wildfires being overwhelmingly higher than those of “wildfire occurrence (1)” cases.^31,37. In particular, if data is routinely organized and by SI and gun, it is capable of having a wildfire-free situation on most days, and therefore outbreak cases represent less than 5% of the total. So, if the model is trained in a very small number of one category (wildfire occurrence: 1), an error may exist. In extreme cases, the model achieves quite high accuracy, even though it predicts an outbreak (0) for all events. Furthermore, recalls drop dramatically, increasing the room for false negatives (FN) issues that models miss actual wildfire outbreaks.

To address this data imbalance, we applied the synthetic minority oversampling technique (SMOTE). This is one of the oversampling techniques.³⁸. Rather than simply replicating data from a minority class at random, Small can reduce overfitting by generating new synthetic data points using vector differences between adjacent samples within the same class (Equation 2). Given that excessive synthetic data is generated when applying Small, we set the normalized number of samples to a 1:1 level, as there is a possibility that the classification boundaries are skewed.

Hyperparameter tuning

Machine learning models have their own hyperparameters, and various settings of hyperparameters can have a major impact on predictive performance³³. For example, in a random forest, you can adjust the number of trees (N_ESTIMATORS), maximum tree depth (MAX_DEPTH), and minimum number of samples (min_samples_split). XGBoost and LightGBM support more parameters by including Learning_rate, Maximum Depth, and L1/L2 normalization coefficients (REG_ALPHA, REG_LAMBDA).

Here, we performed a random search first by setting an intuitive range, then selecting a promising combination, repeatedly narrowing the range. This process triggers the problem of MAX_DEPTH overfitting to the training data, and too many trees (N_ESTIMATORS) not only increases computational costs, but does not guarantee actual performance improvements for the model. Finally, N_ESTIMATORS is limited to the 50-300 range, while MAX_DEPTH is limited to below 20, avoiding the issue and maximizing the performance of the model. To ensure a robust evaluation of hyperparameter combinations, 5x cross-validation was utilized during this tuning process.

Model performance evaluation

The model used here serves as a kind of classifier for classifying wildfire occurrence status, and uses a combination of accuracy, recall, and area under the ROC curve (AUC), which is a representative for evaluating binary classification models. Accuracy and recall refer to the correct prediction percentage of all predicted events and actual wildfire events models, as in the equation. 3 and 4 respectively.

AUC is a metric that can comprehensively assess whether a model exceeds a certain level of classification ability, regardless of the threshold setting.^21,38. It is defined in the expression. 5. If AUC is close to 0.5, the same level of performance as the randomized classification is displayed, and if it is close to 1, the complete classification model. If a model exhibits high AUC and excellent recalls, it can generally be considered as having stable classification performance that does not lack actual occurrence cases.

One limitation of machine learning-based models is the black box nature, which makes it difficult to explain results even if they show good results.^25,26. Its practical use may be reduced if it is difficult to interpret which factors contribute significantly to the predicted outcome. Here, we use Shap (Shapley Additive Description) values to improve interpretability. Shap is born from the concept of Shapley values in game theory, and the contribution of each independent variable can be quantified into predicted results.³⁹. SHAP techniques involve calculating the contribution per variable, by averaging the difference in predicted values between exclusion and inclusion of variables between all combinations of variables, and can be expressed as shown in the equation. 6²⁷. The disadvantage of this technique is that complex calculations are required to calculate SHAP values, but the advantage is that the results can be visualized intuitively. Analysis using SHAP values was applied to four machine learning models except logistic regression.