Application of back-propagation networks in debris flow prediction

Debris flows have caused serious loss of human lives and a lot of damage to properties in Taiwan over the past decades. Moreover, debris flows have brought massive mud causing water pollution in reservoirs and resulted in water shortage for daily life locally and affected agricultural irrigation and industrial usages seriously. A number of methods for prediction of debris flows have been studied. However, the successful prediction ratio of debris flows cannot always maintain a stable and reliable level. The objective of this study is to present a stable and reliable analytical model for occurrence predictions of debris flows. This study proposes an Artificial Neural Networks (ANN) model that was constructed by seven significant factors using back-propagation (BP) algorithm. These seven factors include (1) length of creek, (2) average slope, (3) effective watershed area, (4) shape coefficient, (5) median size of soil grain, (6) effective cumulative rainfall, and (7) effective rainfall intensity. A total of 178 potential cases of debris flows collected in eastern Taiwan were fed into the ANN model for training and testing. The average ratio of successful prediction reaching 93.82% demonstrates that the presented ANN model with seven significant factors can provide a stable and reliable result for the prediction of debris flows in hazard mitigation and guarding systems.     

Recognition of debris flow,debris flood and flood hazard through watershed morphometrics

Debris flows, debris floods and floods in mountainous areas are responsible for loss of life and damage to infrastructure, making it important to recognize these hazards in the early stage of planning land developments. Detailed terrain information is seldom available and basic watershed morphometrics must be used for hazard identification. An existing model uses watershed area and relief (the Melton ratio) to differentiate watersheds prone to flooding from those subject to debris flows and debris floods. However, the hazards related to debris flows and debris floods are not the same, requiring further differentiation. Here, we demonstrate that a model using watershed length combined with the Melton ratio can be used to differentiate debris-flow and debris-flood prone watersheds. This model was tested on 65 alluvial and colluvial fans in west central British Columbia, Canada, that were examined in the field. The model correctly identified 92% of the debris-flow, 83% of the debris-flood, and 88% of the flood watersheds. With adaptation for different regional conditions, the use of basic watershed morphometrics could assist land managers, scientists, and engineers with the identification of hydrogeomorphic hazards on fans elsewhere.     

基于神经网络的泥石流危险度区划

本文探讨了用于泥石流危险度区划的神经网络模型,阐述了基本原理,并结合实例应用表明此方法是可行的和有效的     

河南省泥石流危险度分区

在分析历史资料和防治现状的基础上引入城市重要性作为判别准则之一,以18个地级市为研究对象利用层次分析法对河南省泥石流危险度分区。依据泥石流危险度权重的不同将河南省分为5个区间,其中危险度大的城市1个,危险度中等的城市3个,危险度小的城市6个,危险度极小城市2个,无危险度城市6个。以期为河南省泥石流灾害的防治提供有益的参考。     

Modeling the cyclic swelling pressure of mudrock using artificial neural networks

The stochastic nature of the cyclic swelling behavior of mudrock and its dependence on a large number of interdependent parameters was modeled using Time Delay Neural Networks (TDNNs). This method has facilitated predicting cyclic swelling pressure with an acceptable level of accuracy where developing a general mathematical model is almost impossible. A number of total pressure cells between shotcrete and concrete walls of the powerhouse cavern at Masjed–Soleiman Hydroelectric Powerhouse Project, South of Iran, where mudrock outcrops, confirmed a cyclic swelling pressure on the lining since 1999. In several locations, small cracks are generated which has raised doubts about long term stability of the powerhouse structure. This necessitated a study for predicting future swelling pressure. Considering the complexity of the interdependent parameters in this problem, TDNNs proved to be a powerful tool. The results of this modeling are presented in this paper.     

Slurry-flow deposits in the Britannia Formation (Lower Cretaceous), North Sea: a new perspective on the turbidity current and debris flow problem

The Lower Cretaceous Britannia Formation (North Sea) includes an assemblage of sandstone beds interpreted here to be the deposits of turbidity currents, debris flows and a spectrum of intermediate flow types termed slurry flows. The term ‘slurry flow’ is used here to refer to watery flows transitional between turbidity currents, in which particles are supported primarily by flow turbulence, and debris flows, in which particles are supported by flow strength. Thick, clean, dish‐structured sandstones and associated thin‐bedded sandstones showing Bouma T_b–e divisions were deposited by high‐ and low‐density turbidity currents respectively. Debris flow deposits are marked by deformed, intraformational mudstone and sandstone masses suspended within a sand‐rich mudstone matrix. Most Britannia slurry‐flow deposits contain 10–35% detrital mud matrix and are grain supported. Individual beds vary in thickness from a few centimetres to over 30 m. Seven sedimentary structure division types are recognized in slurry‐flow beds: (M₁) current structured and massive divisions; (M₂) banded units; (M₃) wispy laminated sandstone; (M₄) dish‐structured divisions; (M₅) fine‐grained, microbanded to flat‐laminated units; (M₆) foundered and mixed layers that were originally laminated to microbanded; and (M₇) vertically water‐escape structured divisions. Water‐escape structures are abundant in slurry‐flow deposits, including a variety of vertical to subvertical pipe‐ and sheet‐like fluid‐escape conduits, dish structures and load structures. Structuring of Britannia slurry‐flow beds suggests that most flows began deposition as turbidity currents: fully turbulent flows characterized by turbulent grain suspension and, commonly, bed‐load transport and deposition (M₁). Mud was apparently transported largely as hydrodynamically silt‐ to sand‐sized grains. As the flows waned, both mud and mineral grains settled, increasing near‐bed grain concentration and flow density. Low‐density mud grains settling into the denser near‐bed layers were trapped because of their reduced settling velocities, whereas denser quartz and feldspar continued settling to the bed. The result of this kinetic sieving was an increasing mud content and particle concentration in the near‐bed layers. Disaggregation of mud grains in the near‐bed zone as a result of intense shear and abrasion against rigid mineral grains caused a rapid increase in effective clay surface area and, hence, near‐bed cohesion, shear resistance and viscosity. Eventually, turbulence was suppressed in a layer immediately adjacent to the bed, which was transformed into a cohesion‐dominated viscous sublayer. The banding and lamination in M₂ are thought to reflect the formation, evolution and deposition of such cohesion‐dominated sublayers. More rapid fallout from suspension in less muddy flows resulted in the development of thin, short‐lived viscous sublayers to form wispy laminated divisions (M₃) and, in the least muddy flows with the highest suspended‐load fallout rates, direct suspension sedimentation formed dish‐structured M₄ divisions. Markov chain analysis indicates that these divisions are stacked to form a range of bed types: (I) dish‐structured beds; (II) dish‐structured and wispy laminated beds; (III) banded, wispy laminated and/or dish‐structured beds; (IV) predominantly banded beds; and (V) thickly banded and mixed slurried beds. These different bed types form mainly in response to the varying mud contents of the depositing flows and the influence of mud on suspended‐load fallout rates. The Britannia sandstones provide a remarkable and perhaps unique window on the mechanics of sediment‐gravity flows transitional between turbidity currents and debris flows and the textures and structuring of their deposits.     

The application of genetic algorithm in debris flows prediction

Debris flows caused serious loss of human lives and damages to properties in Taiwan for the past decades. A number of methods for prediction of debris flows have been studied including numerical method, statistic method, experiment method and neural network method in recent years. This study proposed a genetic algorithm (GA) model for occurrence prediction of debris flows. A total of 154 potential cases of debris flows collected in eastern Taiwan were fed into the GA for training and testing. The average ratio of successful prediction reaching 90.4% demonstrates that the presented GA model can provide a stable and reliable result for prediction of debris flows in the hazard mitigation and guarding system.     

Estimation of soil compaction parameters by using statistical analyses and artificial neural networks

This study presents the application of different methods (simple–multiple analysis and artificial neural networks) for the estimation of the compaction parameters (maximum dry unit weight and optimum moisture content) from classification properties of the soils. Compaction parameters can only be defined experimentally by Proctor tests. The data collected from the dams in some areas of Nigde (Turkey) were used for the estimation of soil compaction parameters. Regression analysis and artificial neural network estimation indicated strong correlations (r ² = 0.70–0.95) between the compaction parameters and soil classification properties. It has been shown that the correlation equations obtained as a result of regression analyses are in satisfactory agreement with the test results. It is recommended that the proposed correlations will be useful for a preliminary design of a project where there is a financial limitation and limited time.     

自适应BP算法及其在河道洪水预报上的应用

提出一种改进的BP算法,即自适应BP算法。该方法采用两种策略:一是在权重修改公式中加动量项;二是学习率随总误差的变化作自适应调整,亦即总误差增加时,学习率将减小,反之学习率增大。以上两种策略能有效的抑制网络陷于局部极小并缩短了学习时间。实例研究表明,该算法用于河道洪水的预报,能取得令人满意的结果。     

Artificial neural networks approach for zeta potential of Montmorillonite in the presence of different cations

In this study, the zeta potential of montmorillonite in the presence of different chemical solutions was modeled by means of artificial neural networks (ANNs). Zeta potential of the montmorillonite was measured in the presence of salt cations, Na⁺, Li⁺ and Ca²⁺ and metals Zn²⁺, Pb²⁺, Cu²⁺, and Al³⁺ at different pH values, and observed values pointed to a different behavior for this mineral in the presence of salt and heavy metal cations. Artificial neural networks were successfully developed for the prediction of the zeta potential of montmorillonite in the presence of salt and heavy metal cations at different pH values and ionic strengths. Resulting zeta potential of montmorillonite shows different behavior in the presence of salt and heavy metal cations, and two ANN models were developed in order to be compared with experimental results. The ANNs results were found to be close to experimentally measured zeta potential values. The performance indices such as coefficient of determination, root mean square error, mean absolute error, and variance account for were used to control the performance of the prediction capacity of the models developed in this study. These indices obtained make it clear that the predictive models constructed are quite powerful. The constructed ANN models exhibited a high performance according to the performance indices. This performance has also shown that the ANNs seem to be a useful tool to minimize the uncertainties encountered during the soil engineering projects. For this reason, the use of ANNs may provide new approaches and methodologies.     

The Zymoetz River landslide, British Columbia, Canada: description and dynamic analysis of a rock slide–debris flow

The Zymoetz River landslide is a recent example of an extremely mobile type of landslide known as a rock slide–debris flow. It began as a failure of 900,000 m³ of bedrock, which mobilized an additional 500,000 m³ of surficial material in its path, transforming into a large debris flow that traveled over 4 km from its source. Seasonal snow and meltwater in the proximal part of the path were important factors. A recently developed dynamic model that accounts for material entrainment, DAN3D, was used to back-analyze this event. The two distinct phases of motion were modeled using different basal rheologies: a frictional model in the proximal path and a Voellmy model in the distal path, following the initiation of significant entrainment. Very good agreement between the observed and simulated results was achieved, suggesting that entrainment capabilities are essential for the successful simulation of this type of landslide.     

用遗传神经网络分析泥石流活动性

泥石流是我国山区的主要地质灾害之一。影响泥石流活动性的因素十分复杂，并且具有随机性和模糊性。遗传神经网络结合了神经网络和遗传算法的优点，可以模拟学习和进化之间的交互作用，很适合用于分析泥石流活动性。文章简要讨论了遗传神经网络的原理，建立了泥石流活动性分析的遗传神经网络模型，并将该模型用于川藏公路沿线30条泥石流沟的活动性分析。网络的拓扑结构为(9，6，4，3)，即输入节点(评价指标)、第l隐含层、第2隐含层和输出接点(分析结果)分别为9、6、4、3。首先以其中25条泥石流沟作为样本对网络进行训练，训练时网络的连接权采用遗传算法进行自适应演化，待模型稳定后将其余5条泥石流沟的数据输入模型，计算它们的活动性，计算结果与实际观测基本相符，证明模型是可行的，各个参数的选取也是合适的。     

Applicability of artificial neural networks for obtaining velocity models from synthetic seismic data

Seismic velocity analysis is a crucial part of seismic data processing and interpretation which has been practiced using different methods. In contrast to time consuming and complicated numerical methods, artificial neural networks (ANNs) are found to be of potential applicability. ANN ability to establish a relationship between an input and output space is considered to be appropriate for mapping seismic velocity corresponding to travel times picked from seismograms. Accordingly a preliminary attempt is made to evaluate the applicability of ANNs to determine velocity and dips of dipping layered earth models corresponding to travel time data. The study is based on synthetic data generated using inverse modeling approach for three earth models. The models include a three-layer structure with same dips and same directions, a three-layer model with different dips and same directions, as well as a two-layer model with different dips and directions. An ANN structure is designed in three layers, namely, input, output, and hidden ones. The training and testing process of the ANN is successfully accomplished using the synthetic data. The evaluation of the applicability of the trained ANN to unknown data sets indicates that the ANN can satisfactorily compute velocity and dips corresponding to travel times. The error intervals between the desired and calculated velocity and dips are shown to be acceptably small in all cases. The applicability of the trained ANN in extrapolating is also evaluated using a number of data outside of the range already known to ANN. The results indicate that the trained ANN acceptably approximates the velocity and dips. Furthermore, the trained ANN is also evaluated in terms of capability of handling deficiency in input data where acceptable results were also achieved in velocity and dip calculations. Generally, this study shows that velocity analysis using ANNs can promisingly tackle the challenge of retrieving an initial velocity model from the travel time hyperbolas of seismic data.     

A direct inversion scheme for deep resistivity sounding data using artificial neural networks

Initialization of model parameters is crucial in the conventional 1D inversion of DC electrical data, since a poor guess may result in undesired parameter estimations. In the present work, we investigate the performance of neural networks in the direct inversion of DC sounding data, without the need ofa priori information. We introduce a two-step network approach where the first network identifies the curve type, followed by the model parameter estimation using the second network. This approach provides the flexibility to accommodate all the characteristic sounding curve types with a wide range of resistivity and thickness. Here we realize a three layer feed-forward neural network with fast back propagation learning algorithms performing well. The basic data sets for training and testing were simulated on the basis of available deep resistivity sounding (DRS) data from the crystalline terrains of south India. The optimum network parameters and performance were decided as a function of the testing error convergence with respect to the network training error. On adequate training, the final weights simulate faithfully to recover resistivity and thickness on new data. The small discrepancies noticed, however, are well within the resolvability of resistivity sounding curve interpretations.     

神经网络在泥石流活跃程度评判中的应用

泥石流活跃程度的评判结果对保护当地人民生命财产安全和经济建设的发展及地质灾害的防治工程布置有着很大的影响。然而以往的评判方法多以定性评判为主。由于每个人的知识水平、工作经验及评判问题的思维方式的差异，从而使评判结果或多或少存在一定的误差。论文旨在寻求一种新的方法来实现对泥石流活跃程度的定量分析，以便尽可能的减少人为误差。人工神经网络是一种具有学习、记忆、计算、仿真等功能的网络结构。BP网络是目前工程上运用最为广泛的一种误差反传的人工神经网络。它可以模拟任意复杂的非线形映射关系。应用神经网络对泥石流活跃程度进行定量分析评判，可以在一定程度上减少定性评判中的人为因素影响，提高评判的准确性。论文简要介绍了BP神经网络的基本原理、训练过程，以及如何利用MATLAB软件中的神经网络工具箱来创建、训练和应用评判泥石流活跃程度BP网络。在BP网络模型建立时采用了对研究区泥石流活跃程度影响最主要的8个参数作为输入层，并选取了研究区的20个样本对网络进行训练。最后用训练好的网络对研究区的10条支沟分别进行计算。计算结果与实际情况相符，说明利用BP神经网络来评判泥石流活跃程度具有很好的实用价值。     

模糊优选神经网络在边坡稳定性分析中的应用

本文应用模糊优选神经网络理论,建立了边坡稳定性评价模型,综合分析了影响边坡稳定性的各种因素,根据它们作用程度的不同,赋予不同的权值,通过对收集到的边坡稳定性实例进行学习,提出了与优属度有关的函数关系式。可以看出:该方法具有精度高、收敛速度快、权值调整模型好等优点。     

Kuskonook Creek, an example of a debris flow analysis

Two debris flows occurred on Kuskonook Creek in British Columbia, Canada, in August and September 2004. The initiation factors included a major forest fire in the catchment in 2003, in association with relatively small rainfall events and the accumulation of sediment in the creek channels since the last large debris flow event. Previous regional studies and morphometric comparisons with other similar catchments indicate that Kuskonook Creek has characteristics predisposed to debris flows, even without the affects of a forest fire. Based on the investigations and analyses, a magnitude/frequency relationship for future debris events on Kuskonook Creek was developed, and this information was used to carry out a partial risk assessment. It is suggested that for design purposes, a 1/50-year return period and the corresponding debris flow magnitude of 15,000 m³ would provide conservative protection to the users of the provincial highway at the mouth of the creek.     

Some observations on the prediction of the dynamic parameters of debris flows in pyroclastic deposits in the Campania region of Italy

Over the last 20 years, many tools have been developed for the prediction of the post-failure behaviour of rapid landslides. However, as pointed out by several researchers, knowledge may be improved by the performance of back-analyses using different models and the evaluation of their reliability. This paper reports the back-analysis, conducted using numerical models, of 57 rapid landslides that have occurred in the Campania region. The back-analysis has been performed using the 2-D DAN_W code (version 2003) with two different rheological models: the Voellmy and the frictional models. The latter has been immediately discarded because it did not match the observed data. Instead, using the Voellmy model, the best-fit values of the parameters (friction μ and turbulence ξ) for different types of flow (channelled, un-channelled and mixed flows) have been researched. With these values a parametric study has been carried out on four representative slope profiles of the Campania region, enabling the prediction of runout, velocity and depth of flow (dynamic parameters) of potential debris flows.

Neural network modeling for regional hazard assessment of debris flow in Lake Qionghai Watershed, China

This paper presents a neural network (NN) based model to assess the regional hazard degree of debris flows in Lake Qionghai Watershed, China. The NN model was used as an alternative for the more conventional linear model MFCAM (multi-factor composite assessment model) in order to effectively handle the nonlinearity and uncertainty inherent in the debris flow hazard analysis. The NN model was configured using a three layer structure with eight input nodes and one output node, and the number of nodes in the hidden layer was determined through an iterative process of varying the number of nodes in the hidden layer until an optimal performance was achieved. The eight variables used to represent the eight input nodes include density of debris flow gully, degree of weathering of rocks, active fault density, area percentage of slope land greater than 25° of the total land (APL25), frequency of flooding hazards, average covariance of monthly precipitation by 10 years (ACMP10), average days with rainfall >25 mm by 10 years (25D10Y), and percentage of cultivated land with slope land greater than 25° of the total cultivated land (PCL25). The output node represents the hazard-degree ranks (HDR). The model was trained with the 35 sets of data obtained from previous researches reported in literatures, and an explicit uncertainty analysis was undertaken to address the uncertainty in model training and prediction. Before the NN model is extrapolated to Lake Qionghai Watershed, a validation case, different from the above data, is conducted. In addition, the performances of the NN model and the MFCAM were compared. The NN model predicted that the HDRs of the five sub-watersheds in the Lake Qionghai Watershed were IV, IV, III, III, and IV–V, indicating that the study area covers normal hazard and severe hazard areas. Based on the NN model results, debris flow management and economic development strategies in the study are proposed for each sub-watershed.     

Prediction of ground subsidence in Samcheok City,Korea using artificial neural networks and GIS

This study shows the construction of a hazard map for presumptive ground subsidence around abandoned underground coal mines (AUCMs) at Samcheok City in Korea using an artificial neural network, with a geographic information system (GIS). To evaluate the factors governing ground subsidence, an image database was constructed from a topographical map, geological map, mining tunnel map, global positioning system (GPS) data, land use map, digital elevation model (DEM) data, and borehole data. An attribute database was also constructed by employing field investigations and reinforcement working reports for the existing ground subsidence areas at the study site. Seven major factors controlling ground subsidence were determined from the probability analysis of the existing ground subsidence area. Depth of drift from the mining tunnel map, DEM and slope gradient obtained from the topographical map, groundwater level and permeability from borehole data, geology and land use. These factors were employed by with artificial neural networks to analyze ground subsidence hazard. Each factor’s weight was determined by the back-propagation training method. Then the ground subsidence hazard indices were calculated using the trained back-propagation weights, and the ground subsidence hazard map was created by GIS. Ground subsidence locations were used to verify results of the ground subsidence hazard map and the verification results showed 96.06% accuracy. The verification results exhibited sufficient agreement between the presumptive hazard map and the existing data on ground subsidence area. An erratum to this article can be found at