Application of Machine Learning Algorithms in Analyzing the Effect of Fire on Soil Water Repellency in Tushan Watershed, Golestan Province

Introduction and Goal
In recent decades, forest fires have been recognized as one of the most important environmental threats to forest and rangeland ecosystems. In addition to destroying vegetation and habitats, this phenomenon also severely changes the physical and chemical properties of the soil, the most important of which is the occurrence of the phenomenon of soil water repellency. Soil water repellency can lead to increased surface runoff, reduced permeability, increased erosion, and ultimately reduced ecosystem stability. In recent years, the occurrence of successive and widespread fires in the Hyrcanian forests of northern Iran, particularly in Golestan Province, has caused serious concerns in the field of natural resource management. Despite numerous studies worldwide, domestic research on the effect of fire on soil water repellency, especially the use of new modeling approaches such as machine learning algorithms, is still limited. On the other hand, traditional approaches such as the Water Drop Penetration Time (WDPT) test or ethanol test, although useful for initial identification, cannot model the complex and nonlinear relationships between soil properties and water repellency intensity. Therefore, this study aimed to investigate the temporal changes in soil water repellency after fire and evaluate the power of Machine learning (ML) algorithms in predicting the WDPT index in the Tushan watershed of Golestan Province.  In addition, this research, while addressing existing scientific limitations, provided a practical tool for managing fire risks and planning post-fire recovery.
Materials and Methods
The study was conducted in the Tushan watershed in Golestan Province, part of the Hyrcanian Forests registered on the UNESCO World Heritage List. The climate  of the region is semi-humid to humid, with an average annual rainfall of 620 mm and mean temperature of 16 °C. The soils in the region are mainly loess with silty loam texture and high organic matter content, which provide conditions susceptible to the formation of soil water repellency after fire. The experimental design was factorial and was designed in a completely randomized design with two main factors including land-use type and fire treatment (burned vs. control). Soil samples was conducted at three time intervals (one day, one week, and one month after the fire) and at two depths (0–5 and 5–10 cm), and total of 96 samples were collected. Physical and chemical properties of the soil, including texture, pH, EC, OC, OM, aggregate stability, bulk density, were measured. The soil water repellency intensity was measured by the WDPT test. After entering the data into the Python environment, they were preprocessed; outliers were identified using the IQR method but not removed, normalization was performed using Z-score, and the collinearity of variables was checked using VIF. Then, the data was divided into two parts: training (70%) and testing (30%). Twelve ML algorithms were implemented for modeling, including basic, neighborhood-based, aggregate, and hybrid models. Optimization of the hyperparameters was performed using Bayesian search and five-way cross-validation. The performance of the model was evaluated with R², RMSE, MAE, NSE, and CCC indices. In addition, sensitivity analysis was performed using PFI and SHAP methods, and uncertainty analysis was performed using bootstrap and Monte Carlo simulations.
Results and Discussion
The results show that the effect of fire on soil water repellency was sever but short-lived. The average WDPT one day after the fire was 2.5 minutes, indicating a significant increase in water repellency. In the first week, this decreased to less than a minute and after a month, it was almost zero. This finding indicates the transient nature of the effect of fire on soil surface water repellency. Statistical analyses ANOVA showed that the effects of fire treatment and its interaction with time and land-use on WDPT was significant. Among ML models, the best performance was for Decision Tree (R²=0.44) and Gradient Boosting (R²=0.43), while models such as SVM and Bayesian inference were less accurate. The results of sensitivity analysis showed that silt, EC, and OM were the most important variables in predicting WDPT. SHAP results also showed that increasing silt and OM increased water repellency, while increasing EC had a decreasing effect. The uncertainty analysis indicated that simpler models such as Decision Tree were more robust to outliers, while the uncertainty of more complex models increased at extreme WDPT values. Based on these findings, soil water repellency after fire can be predicted more accurately and stably using simple or boosting-based models.
Conclusion and Suggestions
This study demonstrated that the effect of fire on soil water repellency in the Tushan watershed was severe but short-lived, and this effect completely disappeared within one month. By using machine learning algorithms, especially Decision Tree and Gradient Boosting, complex relationships between soil properties and water repellency intensity can be modeled with acceptable accuracy. The results of sensitivity and uncertainty analysis showed that variables such as silt, EC, and OM played an important role in the formation of water repellency. Based on the findings of this study, it is suggested that for post-fire management, soil water repellency should be monitored in the short-term and simple and stable models should be used to predict its changes. In addition, by combining field data and remote sensing data and developing regional models, it is possible to improve fire risk management. As a result, it is suggested that future research should examine the long-term effects of fire on other soil properties and ecosystem functions, and that deep learning algorithms be used to increase the accuracy of predictions.