The landslide is one of the most widely distributed and hazardous geological hazards in the world [1, 2, 3]. The Three Gorges Reservoir area in China is one of the most common areas for landslide hazards [4, 5]. The main reason for this is the dramatic change in hydrogeological conditions in the reservoir area, which is caused by the rising water level upstream due to the construction of the Three Gorges Dam [6, 7]. Whenever a landslide disaster occurs, it causes a large number of casualties and property damage [8, 9, 10]. If the landslide displacement develops slowly when it is like a time bomb, at any time may be sudden widespread sliding. Therefore, a lot of reinforcement of facilities and evacuation of people is needed when the landslide has not yet slid [11, 12]. Among the many types of landslides, there is a kind of landslide with obvious deformation steps in the landslide displacement–time monitoring curve, which is called a step-like landslide [13]. The development of step-like landslide displacements is mainly influenced by seasonal rainfall and periodic fluctuations in the reservoir or a combination of both [14, 15]. It is possible that each step in landslide displacement could be mistakenly recognized as entering the sliding phase, thus posing a major problem for landslide avoidance and control. Therefore, it is important to explore the influence of monitoring frequency on the prediction accuracy and determine the optimal monitoring frequency, so that the prediction accuracy can be improved and the stage of landslide evolution can be accurately grasped. When combined with landslide displacement prediction results and corresponding disaster prevention and mitigation measures, casualties and property damage can be greatly reduced [16, 17]. As a result, exploring the effect of displacement monitoring frequency on prediction accuracy has important scientific significance.
With the development of mathematical and statistical analysis theory, as well as the emergence of new landslide monitoring techniques and computational tools, more and more scholars at home and abroad use machine learning methods in various predictions of landslides [18, 19, 20]. Within the many machine learning methods, neural network methods have a wide range of applications [21, 22, 23]. In order to build a dynamic model for landslide displacement prediction, the recurrent neural network (RNN) is definitely a better choice than static models, such as the artificial neural network (ANN) [24]. The nodes inside the RNN model are connected recursively, so that the state of the previous moment has an impact on the next moment, which realizes the state feedback of the network [25]. Long short-term memory (LSTM) networks [26], and gated recurrent units (GRU) are all special RNN networks with obvious advantages in dealing with long sequence data prediction problems. Among them the GRU can achieve the same performance as LSTM in a shorter period of time [27]. These neutral network algorithms [28, 29] have also been proven to be reliable and practical in the field of landslide displacement prediction.
The most direct and effective way of achieving landslide displacement prediction is to analyze the landslide displacement – reservoir water – rainfall – time monitoring curve [30]. Therefore, the monitoring data of landslide influencing factors (rainfall, and reservoir water level) largely affect the forecast accuracy of displacement.
In the past, the monitoring period of landslide displacement was usually 1 month in the Three Gorges reservoir area [31]. This situation has also led many scholars [29, 32–34] to use only data with a monitoring period of months when achieving landslide displacement forecasting. However, after conducting a number of studies on landslides, some scholars found that landslide displacement development would delay for a period of time, usually less than one month, after being affected by external factors. Krkac et al. [35], Thom et al. [36], and Roering et al. [37] studied the influence of rainfall and landslide deformation and determined that the period of influence of rainfall on landslide deformation is about 10–30 days. Liu et al. [38] studied the effect of rainfall on landslide deformation in the Three Gorges Reservoir area for about 7–15 days. Although, there are detailed studies on the effects of rainfall, reservoir water, and other factors on landslide displacements. However, due to the limitation of monitoring conditions, the frequency of landslide surface monitoring is monthly, which makes the time interval of influencing factors determined when landslide displacement prediction is only monthly. Therefore, with the upgrade of landslide surface monitoring equipment and the reduction of monitoring period, it is necessary to explore the influence of monitoring frequency on the accuracy of landslide displacement prediction.
The cumulative displacements of three monitoring frequencies are decomposed into trend term displacements and periodic term displacements using the EEMD algorithm. Among these, the trend term displacement is directly predicted using the polynomial fitting, and the periodic term displacement is predicted using the LSTM, and the GRU prediction models. The rainfall, reservoir water level, and cumulative displacement variation are considered the influencing factors in the prediction of the periodic term. The prediction results of cumulative displacement are obtained by superposition of the periodic term prediction results and trend term prediction results, and then the optimal combination of the prediction model and monitoring frequency is screened. The landslide displacement prediction process is shown in Fig. 1.