Evolution Characteristics Of Multiple Physics-Based Parameters And Data Mining Technology For Progressive Landslides

In China, landslide disasters are widespread and occur most frequency because of its vast mountainous area and hilly terrains. Landslide movements and failure have large, potentially catastrophic societal and economic consequences. Thus, preventing such a catastrophe is a central task for scientific community and local government.A landslide can be thought of as a process that occurs on temporal and spatial scales and may involve the progressive deformation of a sliding body, propagation and extension of surface fissures and eventual failure. Studies of landslide evolution are the foundation for the assessment of landslide risk and predictions of runout and time of failure as well as early warning systems.Studies of the evolution of landslides are generally based on measurements from in situ instrumentation, such as conventional topographic surveys, GPS, clinometers, extensometers, and distometers. The measurements are usually collected at selected points in the monitored area (e.g., landslide crown, depletion zone, accumulation zone) that may not be representative of the entire unstable area. An area-sensing approach that allows for the collection of detailed and spatially extensive information can overcome these limits. The evolution processes of landslides have been documented from morphological changes derived from aerial photographs, photogrammetric techniques and Synthetic Aperture Radar (SAR) interferometry.However, most studies have focused on a single parameter that provides only a limited view of the complexity of the phenomena, which could be better characterized with multiple physics-based parameters. Acquiring multiple types of information is thus essential for fully understanding the mechanisms of such disruptive events.In this regard, the purpose of first part for the present study is to utilize model test to analyze the mechanisms of landslide failure. Two 1-g landslide model test was conducted to obtain and analyze multiple types of information that are associated with landslide evolution. The information included deformation, lateral soil pressure and surface infrared radiation temperature data, which were obtained over the entire model test area via a multiple-monitoring system that consisted of 3D laser scanning technology, particle image velocimetry (PIV) analysis, earth pressure cells and a thermal infrared (TIR) imager. Analyses were performed based on the measurements that were collected during the model test. Four stages of evolution (i.e., initial, uniform, accelerated and failure stages) were distinguished from the monitored displacements. The evolution of the deformation, lateral force and surface temperature were investigated during each stage of evolution.Various conventional stability analysis could be performed to evaluate the actual safety and deformation based on the monitoring data. However, both methods are not exclusive. They are actually, in author’s opinion, complementary. The stability analysis provides a snapshot of the safety of the landslide at a particular time but requires specific data that sometimes is not readily available. For example, how a heavy rainfall immediately affects the piezometric level inside the landslide or how changes in elevation of the reservoir surface may have a (delayed) effect on stability. Further, when the number of landslides is large and potentially de-stabilizing factors change rapidly, e.g. heavy rainfall, quick changes in reservoir level, etc. the input values for a precise stability calculation may not be readily available, in particular when the factors of safety are already marginal. Yet, it is desirable to have mechanisms to quickly estimate the likelihood of landslide instability, since such phenomenon can have major consequences to property and lives.On the one hand, the availability of a total of 3,200 monitoring and warning systems with advanced instrumentation in the Three Gorges Reservoir area is potentially very useful for the analysis of relationships between hydrological causes and landslide movements. On the other hand, advances in monitoring instruments have resulted in huge amounts of data. For example, at the end of 2006, the database at Wushan Demonstration Monitoring Center held more than 2,850,000 records, which are continually updated. Although the conventional approach of relating field data with hydrology is still a viable mechanism for exploring the causal factors of landslide movement, it may not be very effective when the data volume and complexity increase. Consequently, transforming the vast data intelligently into useful information and knowledge from assorted sources is critical for effective and successful analysis of hazard-related incidences. In this regard, data mining is an important and effective technique in the field of knowledge discovery to find patterns and extract information from large datasets. A few landslide-related analysis methods have integrated data mining algorithms in different forms. However, current research mostly focuses on the application of data mining methods to the assessment of landslide susceptibility, and few methods have been used to analyze causes of landslide movements.The purpose of the second part for the present study is to utilize data mining technology to analyze the cause-and-effect relationship between hydrological elements and their effect on reservoir landslide movements. The Majiagou and Zhujiadian landslide in the Three Gorges Reservoir have been selected as a case study to explore the usefulness of the proposed data mining approach. Rainfall, reservoir water level and displacement data were mined to discover association rules. The rules extracted were used to assess the contributions of specific hydrological causes to the reservoir landslide movements.Multiple forms of global, regional or local rainfall threshold for rainfall-triggered landslides were reported by many researchers, such as total event rainfall (R) thresholds, rainfall event-duration (E-D) thresholds, and rainfall event-intensity (E-I) thresholds. However, those empirical thresholds are restricted, if it is applied at regional scale or to a single landslide. Moreover, exiting threshold models remain limited when taking consideration of water level fluctuations in the reservoir.Significant developments have been produced on the impacts of water level fluctuations on landslides. Numerical models have attracted considerable study on the topic of water levels fluctuations adjacent to a landslide. Physical model experiments can provide a good understanding of the failure modes and mechanism associated with changes in water level adjacent to a landslide. However, few such studies have been performed due to the high cost. Through a comparative analysis of field measurements involving landslide displacements, reservoir level and rainfall intensity, the conventional approach helps detect the deformation response to rainfall and the water level fluctuations qualitatively. However, a reasonable approach of landslide early warning should be based on quantitative models of landslide evolution processes.In the third part for the present study, we present a hybrid data mining approach to build quantitative models of landslide evolution processes for the Majiagou and Zhujiadian landslide, reservoir landslides in the Three Gorges Reservoir area. Historical records of landslide movements, precipitation and reservoir water level were analyzed more thoroughly using this novel hybrid data mining approach. Moreover, the thresholds of hydrological causes induced landslide movements were studied.The main results are as follows:(1) A conceptual model for progressive failure of landslides with strain-softening behavior has been presented. The specific features and stages of a progressive failure in a landslide are highlighted. ①It has been proposed that the following stages of down-hill progressive landslide formation may be defined as follows:in situ stage, disturbance phase, local failure phase, instability initiation phase, dynamic phase, actual slide event phase. ②It has been proposed that the following stages of up-hill progressive landslide formation may be defined as follows:in situ stage, disturbance phase, local failure phase, instability initiation phase, dynamic phase, global retrogressive failure phase. ③Based on data collection and interpretation, the landslide displacements have been divided into eight categories:steady growth, exponential growth, approximate monotonic growth, step-like, convergent growth, setback-like, fluctuation-like and mixed type. ④The cracks on the ground surface will form a complete crack system gradually with the increase of the landslide deformation in space.(2) A multiple monitoring system that consisted of a 3D laser scanner, PIV imager, soil pressure cells and thermal imaging was designed. ①The test system consisted of a high-speed multi-channel data acquisition system (DT80G), a hydraulic power unit (HPU, Model 505.60), a 3D laser scanner (RIEGL VZ-400), a video camera, an earth pressure cell (Model XTR-2030) and a TIR imager (Model FLIR SC660 with a temperature sensitivity of 0.030℃). ②3D laser scanning technology is useful for monitoring landslide deformation because it allows for the rapid collection of field topographical data with high accuracy and resolution. The error evaluation model of the point positional accuracy was derived. The point density in laser scanning was analyzed theoretically. ③In order to assess the effectiveness of the measurement methods applied to measure displacement in the laser scanning and evaluate their performances, a validation simulation experiment has been carried out. Researches show that:the error evaluation model and point density model provide a theoretical basis for the evaluation of measurement achievement and the optimal designs for the measurement scheme. The cloud to cloud comparison method and scanning data collection techniques are area based measurements, while barycenter method and the benchmark method are point based. With point based measurement small deformations can be detected in only a few selected positions due to the largely manual measurement process. The area based methods give a good approximation of the displacement amplitudes and provide the whole deformation of the slope surface. The application of the 3D laser scanning technology to the landslide model test has the advantage of combining both the point based and area based methods. It provides the whole deformation while maintaining the high precision for the selected positions.④The data processing and error sources for PIV and TIR analysis and have been presented.(3) A 1-g model test was used to investigate the temporal-spatial evolution of landslide failure. A multiple monitoring system that consisted of a 3D laser scanner, PIV imager, soil pressure cells and thermal imaging was designed and implemented to characterize the deformation, lateral soil pressure and surface temperature evolution during slip. The following conclusions were reached: ① The entire landslide movement was successfully recorded by the multiple monitoring system, which allowed for a qualitative interpretation of the spatial-temporal evolution of the landslide failure. The positive results demonstrate that the novel monitoring system can be used to determine the initiation of sliding and detect the mass movement and landslide area. ②Four stages of evolution (i.e., initial, uniform, accelerated and failure stages) were identified using the measured displacements. ③The soil movement had a homogeneous velocity pattern at the rear of the model during the initial deformation stage. During the uniform deformation stage, the displacements concentrated within the crack zone that formed as the load increased. With further loading, the landslide entered the accelerated deformation stage, where the displacements were concentrated near the landslide toe. ④The horizontal soil pressures showed that the soil stress along a vertical profile is not constant but changes during landslide deformation. The depth of the maximum soil pressure increases with landslide movement. ⑤The surface temperature was significantly higher in the landslide area than in the stable area. This observation may be used as an index for identifying landslide areas. The average change in surface temperature (ΔAIRT) in the landslide area increased sharply and subsequently decreased prior to landslide failure.(4) A 1-g model test stabilized with anti-slide piles was used to investigate the temporal-spatial evolution of landslide failure. The following conclusions were reached: ①The variance (D(X)) of the surface displacement field at different loading levels is calculated and taken as the characteristics variable of the surface displacement field. According to the variation of D(X) value, whether D(X) value changes slightly or it increases sharply, the whole deformation and failure process of the slope can be divided into several stages. Experimental results show that a statistic-the variance (D(X)) of the surface displacement field can be used to describe the formation and penetrating of the crack and indicate the failure of the slope. The severe deformation localization is occurred on the slope surface and develops to crack. Thus, the variance (D(X)) will sharply increase when the deformation localization occurs. ②The dynamic parameter of slope displacement field and the fractals-Hurst index of variance of surface displacement field is determined by means of R/S analysis. The correlation between the Hurst index of displacement field variance and the stability of the slope is carried out. It is found that the Hurst index of the surface displacement field variance was at a trend of low value-recovery-dropping during the process of landslide transform. The Hurst index of the surface displacement field variance decreases before the destabilization of the slope. ③The test results show that the trust behind the pile is a parabolic distribution and the resultant force point is in half of free part above the sliding surface. Since the built-in piles take the most part of pushing force, the soil resistance in front of it is more stable and smaller.(5) Two hybrid data mining approaches have been presented in this study. ① Transforming the vast data intelligently into useful information and knowledge from assorted sources is critical for effective and successful analysis of hazard-related incidences. Data mining is an important and effective technique in the field of knowledge discovery to find patterns and extract information from large datasets. ② A data mining approach combining the two-step cluster method and Apriori algorithm, both implemented in the software Clementine 12.0, was used to interpret landslide field monitoring data. The approach proved useful and made it possible to estimate the cause-and-effect relationship between hydrological parameters and landslide movements. ③Quantitative model of landslide evolution processes in the Three Gorges Reservoir area were established in our study, using the proposed hybrid data mining approach with the two-step cluster method and decision tree C5.0.(6) Two reservoir landslides in the Three Gorges Reservoir have been selected as case study to explore the usefulness of the proposed data mining approach. Rainfall, reservoir water level and displacement data were mined to discover association rules and quantitative model of landslide evolution processes. ①The approach proved useful and made it possible to estimate the cause-and-effect relationship between hydrological parameters and landslide movements, and establish the quantitative model of landslide evolution processes. ②Based on the data mining approach, twenty strong rules were found that relate hydrological and reservoir data with movement of the Majiagou landslide. The association rules show that rapid drawdown of the reservoir and prolonged heavy rainfall are the two main hydrological causes of movement in the Majiagou landslide. It has been found that the landslide movement is especially active when the reservoir level declines. The Majiagou landslide is a retrogressive landslide consisting of multiple sliding surfaces. The primary slide occurred along a weak mudstone interlayer at a depth of approximately 35 m. A second failure surface was found at the contact between the soil deposit and bedrock. The Majiagou landslide is unstable and continues to deform. The landslide movements have developed progressively from the front to the back. Therefore, long-term monitoring of this landslide is critical. Given the successful use of data mining algorithms to identify cause and effect for the movement of the Majiagou landslide, it seems that such methodology could be useful to identify the cause (s) for the instability of other landslides. More specifically, the rules found in this paper relating rainfall and reservoir level could be applicable to other cases. ③Seven quantitative model of landslide evolution processes that relate hydrological causes and landslide movements were discovered. Our results show that the water level fluctuations and long-duration rainfall are the top two triggering factors of the quantitative model for the Zhujiadian landslide evolution processes. It has also been found that landslide movement is more active when there is a rapid drawdown of the reservoir. Moreover, the threshold of long-duration rainfall intensity induced landslide movement is 85.5 mm per month. The threshold of water level fluctuations triggered fast landslide movement is 7.06 m per month. These findings are significant to improve the model for predicting deformation response to rainfall and water level fluctuations of reservoir landslides. Given the successful use of hybrid data mining algorithms to identify causal factors and built quantitative model of landslide evolution processes for the movement of the Zhujiadian landslide, it seems that such methodology could be useful to identify the reasons for the instability of other landslides and built the quantitative model of landslide evolution processes. More specifically, the quantitative rules found in this paper relating rainfall and reservoir level could be applicable to other cases, more specifically in the Three Gorges Reservoir area.The main innovation of this paper are summarized as follows:(1) A conceptual model for progressive failure of landslides with strain-softening behavior has been presented. The specific features and stages of a progressive failure in a landslide are highlighted.(2) Two 1-g model test were used to investigate the physics-based parameters of landslide failure with multiple systems of monitoring data. The information included deformation, lateral soil pressure and surface infrared radiation temperature data, which were obtained over the entire model test area via a multiple-monitoring system that consisted of 3D laser scanner, particle image velocimeter (PIV), earth pressure cells and a thermal infrared (TIR) imager. Analyses were performed based on the measurements that were collected during the model test.(3) Two hybrid data mining approaches have been presented in this study. A data mining approach combining the two-step cluster method and Apriori algorithm, both implemented in the software Clementine 12.0, was used to interpret landslide field monitoring data. The approach proved useful and made it possible to estimate the cause-and-effect relationship between hydrological parameters and landslide movements. A hybrid data mining approach with the two-step cluster method and decision tree C5.0 was used to establish the quantitative model of landslide evolution processes. Two reservoir landslides in the Three Gorges Reservoir have been selected as case study to explore the usefulness of the proposed data mining approach. Rainfall, reservoir water level and displacement data were mined to discover association rules and quantitative model of landslide evolution processes.