UAV-spectral / artificial neural network-based seepage analysis data acquisition and applicability assessment for the prediction of landslides

Abstract

This study aims to predict the time of landslide occurrence through seepage analysis and slope stability analysis in the event of rainfall by predicting seepage analysis data (soil type, water content change, soil water characteristic curve (SWCC) parameters, saturated hydraulic conductivity) that are necessary for landslide prediction. Recently, significant research is being conducted on remote sensing and unmanned aerial vehicle (UAV) monitoring. These methods are more accessible compared to satellite images and are useful for examining wide areas as they are suitable for time-series monitoring. Therefore, this study utilizes UAVs to analyze the spectral data of soil samples obtained from mountains in 18 regions in Korea. Additionally, models are developed to estimate the soil type and hydraulic characteristics. First, spectral images of soil samples were obtained through hyperspectral experiments, and the images were analyzed at the pixel level. Artificial neural networks (ANN) were then developed to predict the soil type, change in water content, SWCC parameters, and saturated hydraulic conductivity. The ANNs were trained by classifying Korean soils into four main colors: brown, red, yellow, and gray. In the case of the ANN for soil classification, an index capable of classifying each soil was proposed through the spectral analysis of 114 soils, and this index was used as input data for the artificial neural network. In the case of water content, the factor with the strongest correlation with water content was determined through regression analysis and was used as input data for the ANN. The coefficients of determination of the developed artificial neural network were all 0.9 or higher, indicating high predictive performance. The spectral images acquired using the UAV include location information, and after being converted into an orthographic image through pre-processing, spectral data were obtained through layer stacking. Subsequently, trees were removed from the image and only the exposed sections of the topsoil were extracted. The ANN capable of estimating soil type was applied to map the sand % and the fine % of the soil in the Pyeongchang area. Here, the root mean square error (RMSE) was 5.48 for sandy soil and 0.107 for fine soil. The estimated sand % and fine %values are necessary input factors of ANN models developed to predict SWCC parameters and saturated hydraulic conductivity. Similarly, the ANN model capable of predicting water content variation was applied to sites in Gokseong and Pyeongchang. As a result, the resulting the mean square error (MSE) was 1.69 in the Gokseong area and 4.12 in the Pyeongchang area. Therefore, the model is evaluated to be highly capable of predicting water content variation in soils in the field. Furthermore, a seepage analysis factor prediction model is proposed in this study. Using this model, seepage analysis and slope stability analysis were performed by considering a rainfall event that occurred in Songnisan in August 2021. First, high-resolution topographic information of the site was acquired using a UAV- LiDAR system. Next, numerical analysis was performed using data measured at the site and the predicted seepage analysis data. Based on these results, the changes in slope safety factor under rainfall were compared and analyzed. The findings highlight the potential of the final developed predictive model to be used in the field to provide primary data for slope stability evaluations over wide areas.