Deformation Classification And Prediction Study Based On Hyperparameter-optimized Machine Learning Models

Landslides are dynamic and recurring natural geological disasters that cause huge human and economic losses worldwide every year.These geological disasters are one of the most common disasters which can lead to severe damage to infrastructure and cause a large number of fatalities.There are many factors that trigger natural disasters such as landslides,some are natural,such as heavy rain,and some are man-made,such as mining and construction activities.At present,although various research has obtained a large amount of monitoring data for landslides,there are still some limitations in the in-depth excavation of the key factors causing landslides and early warning and prediction systems for the same disaster.Therefore,in order to protect the safety of human life and property,it is urgent to conduct in-depth research and analysis on these monitoring data,dig out the inducing factors of landslide disasters,and propose prediction and early warning systems to support decision-makers in risk assessment departments.The main research work of this study is as follows:First,for the unbalanced landslide monitoring datasets collected by NASA from around the world,landslides and their triggers,size,and type are primarily classified.First,in order to deal with imbalanced datasets and improve the accuracy of landslide classification,optimized machine learning models based on evolutionary algorithms and ADASYN are proposed.The accuracy of a model significantly relies upon its chosen parameters,which draws the importance of hyperparameter tuning.In this study,Particle Swarm Optimization,Genetic Algorithm,and Bayesian optimization were used to optimize the Random Forest Classifier and XGBoost’s parameters.These methods not only reduce the time spent in determining the optimal parameters of the model but also minimize the mistakes that inexperienced researchers can make when setting model parameters.Second,the nature of landslides and their influencing factors are dynamic and change with time,but most of the current landslide predictions are based on the research and practice of static machine learning methods.Therefore,this paper proposes a hybrid dynamic model based on LSTM to solve the dynamic prediction problem of landslide disasters.LSTM is a recurrent neural network model that builds its output based on historically collected data during training,and its characteristics are suitable for dealing with dynamically changing systems such as landslides.However,the LSTM model has many parameters,and it is easy to fall into the local optimum when optimizing the model parameters using the original PSO algorithm.Therefore,this paper decomposes the landslide dataset into trend components and periodic components by comparing EMD,EEMD,CEEMDAN,and differential decomposition methods.Then,a PSOGWO-CNN-LSTM model is proposed to predict the periodic landslide component,and the polynomial fitting method is used to predict the landslide trend component to more accurately predict the landslide change.The model was applied and tested on a time series dataset of landslides collected in Northern California,USA.Experimental results show that the proposed PSO-GWO-CNN-LSTM has higher prediction accuracy than the GRU,CNN-LSTM,and Elman Networks models.Third,this paper designs and implements the Beidou high-precision positioning geological disaster monitoring system based on Io T sensors that collect data and send the data to the cloud to perform real-time dynamic prediction of landslides based on the above hybrid prediction model.The analysis results are then shared on a website in the form of graphs and articles.Aiming at the classification and prediction of landslides,this paper conducts in-depth research and development from three different aspects: model parameter optimization methods,hybrid prediction models,and geological disasters prediction systems.Decision support will contribute to the prevention and mitigation of geological disasters.