1. Extreme Gradient Boosting (XGBoost)

Extreme Gradient Boosting algorithm is a known novel application of the gradient boosting machine. It was proposed by (Chen and Guestrin 2016), and it is centred on boosting. Boosting here means combining predictions obtained from a series of weak learners that aids the development of a strong learner by implementing additive training operations. The Extreme Gradient Boosting algorithm (XGBoost) aims to prevent overfitting while optimizing the computation resources. To achieve this, XGBoost performs a simplification of the objective functions by giving room for the combination of predictive and regularization terms while keeping to an optimum computation speed. In addition to this, parallel automatic calculations are executed when the model is being trained. (Fan et al. 2018).

XGBoost is used in solving supervised learning problems, and unlike random forest, where the simultaneous training of forest trees is executed, XGBoost can create ensembles by implementing an iteration-based training of the decision trees on training samples to which weights are attached. These weights are updated at each iteration to reveal the ensemble's residual error at that iteration step (La Cava et al. 2019).

To further expatiate on the working process of XGBoost, Fan et al. (2018) gave the general principle and process involved in optimizing with XGBoost (additive learning). This is given as follows;

• The starting point involves fitting the first learner to all of the input data.

• Create the next model and fit it on the residuals to deal with the weak learner's problems.

• As described in step two, the process of fitting is repeated till it reaches a stopping criterion.

The final model prediction is obtained by computing the sum of the prediction given by each learner. The general function denoting the prediction at step t is denoted by equation 4.1-1 as follows;

Equation 2.4-1

fi(t)= j=1tfjai= fit-1 + ftai

fjai = learner at step t; fi(t) = predictions at step t; fit-1= predictions at step t-1;

ai = input variable

To prevent issues with overfitting and still maintain the computational speed of the model, XGBoost uses the expression denoted by the equation for the evaluation of the model’s goodness from the original function given by equation 2.5-2 as follows;

Equation 2.4-2

L(t)= k=1nl(yi,yi)+ k=1t(fi)

Equation 2.4-3

and f=γT+ 12λ∥ω∥ ²

L(t) is the regularised objective to be minimised; l is referred to as the loss function that measures the difference between the prediction (yi) and the target (yi); is the additional term that aids the prevention of overfitting. T denotes the number of leaves, is the scores in the leaves, is the regularization parameter, and is the minimum loss required for the additional partition of the node. Detailed description, information and computational processes of XGBoost can be found in (Chen and Guestrin 2016).

Spatiotemporal Characteristics of Rainstorms in Machine learning - Present Study

The studies discussed in section 2.3 show the efficacy of using machine learning models to predict flash flood. Most of the studies used similar flash flood triggering factors and even explored different machine learning models to perform a comparative analysis of their predictive modelling accuracy. However, almost all of them did not fully explore the possibility of incorporating spatiotemporal characteristics (features) such as those extracted from the rainstorm’s dataset.

Although Alipour et. al. (2020) attempted to use a few selected spatiotemporal features, their research focused on implementing a flash flood damaged prediction model, which is not the case in this study. The inclusion of a robust number of watershed properties as predictors in their model was not considered. In addition to this, their study made no comparative analysis of the machine learning model (Random forest) explored with any other machine learning model, which is an obvious step in most studies performed in the area of flash flood prediction using machine learning. This present research seeks to address these issues and, more importantly, include a broader range of spatiotemporal storm features in classifying flash flood events using machine learning algorithms.

Hence, this present study aims to complement the wider research works conducted and are ongoing. Two machine learning algorithms, Random Forest and Extreme Gradient Boosting (XGBoost) are implemented to classify the event and a comparative analysis of their performance. A paramount aspect of the research is the analysis of the sensitivity of a wide range of flash flood triggering features (both static and dynamic (spatiotemporal characteristics extracted from the rainstorm dataset)) and interpretation of the machine learning models using a model-agonistic approach. All input variables used in this study were extracted from remotely sensed data.

Flash floods impacted each country differently during these years, and this is more evident in the analysis of the distribution of the monthly events, as shown in figure 5.1-2 and figure 5.1-3. These figures show that many of these events occurred within July, one of the key months in the monsoon in which rainfall events intensity, as indicated by the inventory sources (Mekong River Commission 2018). Such intense rainfall usually continues into the month of August and as seen from the inventory compilation, a significant number of the events captured in this study, for 2014,2015, 2016 and 2018, occurred in the month of August. It is also important to note that lesser Year Cambodia Lao PDR Thailand 2014 16 9 38 2015 11 21 2 2016 4 7 16 2017 13 15 16 2018 14 16 11 Total 58 68 83 40 events were gathered for the months of September and October, with October having the lowest number of events.

Conclusion

The study applied machine learning techniques to analyse the impact of a large range of spatiotemporal rainstorm features and static features causative to flash floods in the classification of flash floods in the Lower Mekong Basin. The input features were extracted from remotely sensed data and have shown a high level of reliability in using them in machine learning classification models for classifying flash floods. The impact of these features on the flash floods classification was determined and analysed using model-agnostic methods.

The main research question read as follows:

How important are spatiotemporal storm features in the classification of flash floods in the Lower Mekong Basin, using machine learning?

To answer the main question above, some sub-questions were constructed and answered. The answers to these sub-questions are given as follows;

• What kind of association among the variables that trigger the flash floods in the Lower Mekong Basin?

From the correlation analysis, there exists a strong association among some of the triggering variables extracted from the remotely sensed data covering the LMB. However, these associations are significantly clustered. Most spatiotemporal storm variables are highly correlated with themselves, which is the same for most static features. The storm's characteristic, such as the Volume, Total magnitude and Maximum Intensity, are highly correlated with each other, with the correlation between the total magnitude and volume is the highest. Being that the case, though feature selection before model development is not the focus in this study, it is safe to say that either the total magnitude of the volume could be selected as part of the machine learning model input variable without the need of including both features in the model. Similarly, amongst the static features, slope is highly correlated with TWI and SPI. Furthermore, a correlation threshold may serve as the basis for selecting the needed features for the model.

• Given the data, which of the machine learning algorithms applied in this study is more suited to classify the flash flood events in the Lower Mekong Basin?

The machine learning task involved in this study was the classification of flash floods in the LMB, which could involve using several machine learning algorithms (Classifiers). In this study, two machine learning algorithms were implemented. These were Random Forest and XGBoost. They were both used in the global predictive classification of flash floods. The model classification results and evaluation show that XGBoost performed better than Random forest in predicting the flash floods, despite the dataset imbalanced dataset used in this study. Both algorithms are ensembles and are known to have better performance even when the available dataset is imbalanced. This could partially explain why they both performed well. However, since XGBoost performed better than Random forest, only XGBoost was used to develop the regional models, providing accurate predictions.

• Which of the spatiotemporal storm variables (factors) mainly impact the machine learning models' performance, and are they more important than the static variables in classifying the flash floods?

From the model-agnostic interpretations performed on the global models, the rainstorm volume impacts the models' performance. The total magnitude and the maximum intensity also play a role in impacting the performance of the models. The rainstorm volume is assigned the highest permutation importance score. Should the Volume feature be removed from the list of features, it will drop the f1-score for Random Forest and XGBoost by 0.528 and 0.527, respectively. Also, SHAP feature importance gives similar results, indicating the rainstorm volume as the highest contributing feature to the model predictions.

As discussed in section 5.4.2, the resulting permutation feature importance analysis of the four regional models developed show that three out of these four regional models rank the volume as the most important making the feature “volume” the strongest predictor. These results strongly confirm that the models built mainly depend on the rainstorm volume for accurate classifying flash floods.

As shown in the permutation importance results obtained, most of the spatiotemporal storm variables have a higher impact on the model predictions when compared to the static variables. Five spatiotemporal features against twelve static features were used in the model. For every model built, three out of the five spatiotemporal features rank highest in feature importance. The variables are Volume, total magnitude and Maximum intensity of the rainstorm. For all models built, the static variables were not ranked higher than these features. But it is also important to state that the model performance (global and regional) would not have been achieved without the combined effect of the static and spatiotemporal rainstorm features, even though the models, without a doubt, mainly depend on the spatiotemporal features for highly accurate and precise predictions.

In addition to the above, some of these static variables could act as dynamic variables. This is because changes in land use and configuration of the topography, especially in developing areas of the basin, will lead to changes in how flash floods are triggered in those locations. This further confirms that though the features from rainstorms are highly important, static features cannot be ignored even though they undergo long term changes.

Hence, spatiotemporal storm variables are generally more important than static features. However, both static and spatiotemporal storm features will lead to a more precise classification of the flash floods using machine learning models.