Assessment of Landslide Susceptibility Combining Deep Learning with Semi-Supervised Learning in Jiaohe County, Jilin Province, China

A comparative study on urban waterlogging susceptibility assessment based on multiple data-driven models

Accurate identification of urban waterlogging areas and assessing waterlogging susceptibility are crucial for preventing and controlling hazards. Data-driven models are utilized to forecast waterlogging areas by establishing intricate relationships between explanatory variables and waterlogging states. This approach tackles the constraints of mechanistic models, which are frequently complex and unable to incorporate socio-economic factors. Previous research predominantly employed single-type data-driven models to predict waterlogging locations and evaluation of their effectiveness. There is a scarcity of comprehensive performance comparisons and uncertainty analyses of different types of models, as well as a lack of interpretability analysis. The chosen study area was the central area of Beijing, which is prone to waterlogging. Given the high manpower, time, and economic costs associated with collecting waterlogging information, the waterlogging point distribution map released by the Beijing Water Affairs Bureau was selected as labeled samples. Twelve factors affecting waterlogging susceptibility were chosen as explanatory variables to construct Random Forest (RF), Support Vector Machine with Radial Basis Function (SVM-RBF), Particle Swarm Optimization-Weakly Labeled Support Vector Machine (PSO-WELLSVM), and Maximum Entropy (MaxEnt). The utilization of diverse single evaluation indicators (such as F-score, Kappa, AUC, etc.) to assess the model performance may yield conflicting results. The Distance between Indices of Simulation and Observation (DISO) was chosen as a comprehensive measure to assess the model's performance in predicting waterlogging points. PSO-WELLSVM exhibited the highest performance with a DISOtest value of 0.63, outperforming MaxEnt (0.78), which excelled in identifying areas highly susceptible to waterlogging, including extremely high susceptibility zones. The SVM-RBF and RF models demonstrated suboptimal performance and exhibited overfitting. The examination of waterlogging susceptibility distribution maps predicted by the four models revealed significant spatial differences due to variations in computational principles and input parameter complexities. The integration of four WSAMs based on logistic regression has been shown to significantly decrease the uncertainty of a single data-driven model and identify the most flood-prone areas. To improve the interpretability of the data model, a geographical detector was incorporated to demonstrate the explanatory capacity of 12 variables and the process of waterlogging. Building Density (BD) exhibits the highest explanatory power in relation to explain waterlogging susceptibility (Q value = 0.202), followed by Distance to Road, Frequency of Heavy Rainstorms (FHR), DEM, etc. The interaction between BD and FHR results in a nonlinear increase in the explanatory power of waterlogging susceptibility. The presence of waterlogging susceptibility risk in the research area can be attributed to the interactions of multiple factors.

Landslide susceptibility assessment using deep learning considering unbalanced samples distribution

Landslide susceptibility assessment (LSA) is fundamental for managing landslide geological disasters. This study presents a deep learning approach (DNN-MSFM) designed to enhance LSA modeling, particularly addressing limitations caused by the unbalanced distribution of data samples in applied datasets. DNN-MSFM approach combines a deep neural network (DNN) and a mean squared false misclassification loss function (MSFM) to handle unbalanced samples from the algorithmic perspective. The model's performance was evaluated on an unbalanced dataset containing mapping units' records of 293 landslide samples and 653 non-landslide samples from the Baota District, China. Its effectiveness was assessed through statistical metrics and compared against DNN and Support Vector Machine (SVM) basic models. The results demonstrated a significant performance enhancement from the DNN-MSFM (OverallAccuracy = 0.889 and area under the receiver operating characteristic curve (AUC) = 0.84), indicating its effectiveness in learning the underlying landslide susceptibility features and demonstrating its ability to provide improved predictions even in areas with unbalanced landslide samples. Moreover, the study emphasizes the importance of considering balanced loss functions in training DNN under various imbalance degrees and contributes to expanding the applicability of DNN in LSA modeling. Also, this study builds a foundation for further enhancements of deep learning methods for geological disaster assessments.

Landslide susceptibility assessment based on frequency ratio and semi-supervised heterogeneous ensemble learning model

Epistemic uncertainty in data-driven landslide susceptibility assessment often tends to be increased by the limited accuracy of an individual model, as well as uncertainties associated with the selection of non-landslide samples. To address these issues, this paper centers on the landslide disaster in Ji'an City, China, and proposes a heterogeneous ensemble learning method incorporating frequency ratio (FR) and semi-supervised sample expansion. Based on the superimposed results of 12 environmental factor frequency ratios (FFR), non-landslide samples were selected and input into light gradient boosting machine (LightGBM), random forest (RF), and convolutional neural network (CNN) models for prediction along with historical landslide samples. The predicted probability values are integrated by four heterogeneous ensemble strategies to expand samples from high-confidence results. The model's performance is evaluated using the area under the receiver operating characteristic curve (AUC), partition frequency ratio (PFR), and other verification methods. The results demonstrate that the negative sample based on FFR sampling is more accurate than the random sampling method, and the FR-SSELR model based on frequency ratio sampling and semi-supervised ensemble strategy exhibits the highest performance (AUC = 0.971, ACC = 0.941). A more reasonable landslide susceptibility map was drawn based on this model, with the lowest percentage of landslides in the low and very low susceptibility zones (sum of PFR = 0.194), as well as the highest percentage of landslides in the high and very high susceptibility zones (sum of PFR = 6.800). Furthermore, the FR-SSELR model improved economic benefits by 3.82-14.2%, offering valuable guidance for decision-making regarding landslide management and the sustainability of Ji'an City.

Debris flow susceptibility assessment based on information value and machine learning coupling method: from the perspective of sustainable development

Accurately assessing the susceptibility of debris flow disasters is of great significance for reducing the cost of disaster prevention and mitigation, as well as disaster losses. Machine learning (ML) models have been widely used in the susceptibility assessment of debris flow disasters. However, these models often have randomness in the selection of non-disaster data, which can lead to redundant information and poor applicability and accuracy of susceptibility evaluation results. To address this issue, this paper focuses on debris flow disasters in Yongji County, Jilin Province, China; optimizes the sampling method of non-disaster datasets in machine learning susceptibility assessment; and proposes a susceptibility prediction model that couples information value (IV) with artificial neural network (ANN) and logistic regression (LR) models. A debris flow disaster susceptibility distribution map with higher accuracy was drawn based on this model. The model's performance is evaluated using the area under the receiver operating characteristic curve (AUC), information gain ratio (IGR), and typical disaster point verification methods. The results show that the rainfall and topography were found to be decisive factors in the occurrence of debris flow disasters, and the IV-ANN model established in this study had the highest accuracy (AUC = 0.968). Compared to traditional machine learning models, the coupling model produced an increase in economic benefit of about 25% while reducing the average disaster prevention and control investment cost by about 8%. Based on model's susceptibility map, this paper proposes practical disaster prevention and control suggestions that promote sustainable development in the region, such as establishing monitoring systems and information platforms to aid disaster management.

Landslide Susceptibility Prediction Based on High-Trust Non-Landslide Point Selection

Landslide susceptibility prediction has the disadvantages of being challenging to apply to expanding landslide samples and the low accuracy of a subjective random selection of non-landslide samples. Taking Fu’an City, Fujian Province, as an example, a model based on a semi-supervised framework using particle swarm optimization to optimize extreme learning machines (SS-PSO-ELM) is proposed. Based on the landslide samples, a semi-supervised learning framework is constructed through Density Peak Clustering (DPC), Frequency Ratio (FR), and Random Forest (RF) models to expand and divide the landslide sample data. The landslide susceptibility was predicted using high-trust sample data as the input variables of the data-driven model. The results show that the area under the curve (AUC) valued at the SS-PSO-ELM model for landslide susceptibility prediction is 0.893 and the root means square error (RMSE) is 0.370, which is better than ELM and PSO-ELM models without the semi-supervised framework. It shows that the SS-PSO-ELM model is more effective in landslide susceptibility. Thus, it provides a new research idea for predicting landslide susceptibility.

Convolutional neural network (CNN) with metaheuristic optimization algorithms for landslide susceptibility mapping in Icheon, South Korea

Landslides are a geological hazard that can pose a serious threat to human health and the environment of highlands or mountain slopes. Landslide susceptibility mapping is an essential tool for predicting and mitigating landslides. This study aimed to investigate the application of deep learning algorithms based on convolutional neural networks (CNNs) with metaheuristic optimization algorithms, namely the grey wolf optimizer (GWO) and imperialist competitive algorithm (ICA), to landslide susceptibility mapping. The study area was Icheon City, South Korea, for which an accurate landslide inventory dataset was available. The landslide inventory map was prepared and randomly divided into datasets of 70% for training and 30% for validation. Additionally, 18 landslide-related factors, including geo-environmental and topo-hydrological factors, were considered as predictive variables. The models were compared using area under the curve (AUC) values in receiver operating characteristic (ROC) curve analysis. The validation results showed that optimized models based on CNN-GWO (AUC = 0.876, RMSE = 0.08) and CNN-ICA (AUC = 0.852, RMSE = 0.09) outperformed the standalone CNN model (AUC = 0.847, RMSE = 0.12). Nevertheless, the CNN model outperformed previous research that used a machine learning algorithm alone. Thus, the deep learning algorithm with optimization algorithms proposed in this study can generate more suitable models for landslide susceptibility mapping in the study area due to its improved accuracy.

Analysis of Changes in Landslide Susceptibility according to Land Use over 38 Years in Lixian County, China

Landslides occur frequently in Lixian County, China, and land use has changed significantly in recent decades. We obtained land use data for the years 1980, 2000, and 2018, as well as three landslide susceptibility maps from a Random Forest model. Agricultural land, low coverage grassland, water area, and urban, rural and other construction land were prone to landslides. Landslide susceptibility was low in areas of woodland, moderate and high coverage grassland, bare rock land, desert and tundra. Areas with high landslide susceptibility were mainly located in the catchment of the study region, and a 2.61% decrease in high landslide susceptibility areas over the 38-year period was primarily driven by changes in agricultural and rural land. By contrast, a 1.42% increase in low landslide susceptibility areas over the 38-year period was driven by changes in moderate and high coverage woodland and moderate coverage grassland. There is a need for effective management measures to be implemented because areas with high landslide susceptibility are still present. We also found that human aggregations, or the absence of these, vary in their effects on the areas of Lixian County most susceptible to landslides.

Landslide Detection from Open Satellite Imagery Using Distant Domain Transfer Learning

Using convolutional neural network (CNN) methods and satellite images for landslide identification and classification is a very efficient and popular task in geological hazard investigations. However, traditional CNNs have two disadvantages: (1) insufficient training images from the study area and (2) uneven distribution of the training set and validation set. In this paper, we introduced distant domain transfer learning (DDTL) methods for landslide detection and classification. We first introduce scene classification satellite imagery into the landslide detection task. In addition, in order to more effectively extract information from satellite images, we innovatively add an attention mechanism to DDTL (AM-DDTL). In this paper, the Longgang study area, a district in Shenzhen City, Guangdong Province, has only 177 samples as the landslide target domain. We examine the effect of DDTL by comparing three methods: the convolutional CNN, pretrained model and DDTL. We compare different attention mechanisms based on the DDTL. The experimental results show that the DDTL method has better detection performance than the normal CNN, and the AM-DDTL models achieve 94% classification accuracy, which is 7% higher than the conventional DDTL method. The requirements for the detection and classification of potential landslides at different disaster zones can be met by applying the AM-DDTL algorithm, which outperforms traditional CNN methods.

Measuring landslide vulnerability status of Chukha, Bhutan using deep learning algorithms

Landslides are major natural hazards that have a wide impact on human life, property, and natural environment. This study is intended to provide an improved framework for the assessment of landslide vulnerability mapping (LVM) in Chukha Dzongkhags (district) of Bhutan. Both physical (22 nos.) and social (9 nos.) conditioning factors were considered to model vulnerability using deep learning neural network (DLNN), artificial neural network (ANN) and convolution neural network (CNN) approaches. Selection of the factors was conceded by the collinearity test and information gain ratio. Using Google Earth images, official data, and field inquiry a total of 350 (present and historical) landslides were recorded and training and validation sets were prepared following the 70:30 ratio. Nine LVMs were produced i.e. a landslide susceptibility (LS), one social vulnerability (SV) and a relative vulnerability (RLV) map for each model. The performance of the models was evaluated by area under curve (AUC) of receiver operating characteristics (ROC), relative landslide density index (R-index) and different statistical measures. The combined vulnerability map of social and physical factors using CNN (CNN-RLV) had the highest goodness-of-fit and excellent performance (AUC = 0.921, 0.928) followed by DLNN and ANN models. This approach of combined physical and social factors create an appropriate and more accurate LVM that may-support landslide prediction and management.

Predicting and Interpreting Students’ Grades in Distance Higher Education through a Semi-Regression Method

Multi-view learning is a machine learning app0roach aiming to exploit the knowledge retrieved from data, represented by multiple feature subsets known as views. Co-training is considered the most representative form of multi-view learning, a very effective semi-supervised classification algorithm for building highly accurate and robust predictive models. Even though it has been implemented in various scientific fields, it has not adequately used in educational data mining and learning analytics, since the hypothesis about the existence of two feature views cannot be easily implemented. Some notable studies have emerged recently dealing with semi-supervised classification tasks, such as student performance or student dropout prediction, while semi-supervised regression is uncharted territory. Therefore, the present study attempts to implement a semi-regression algorithm for predicting the grades of undergraduate students in the final exams of a one-year online course, which exploits three independent and naturally formed feature views, since they are derived from different sources. Moreover, we examine a well-established framework for interpreting the acquired results regarding their contribution to the final outcome per student/instance. To this purpose, a plethora of experiments is conducted based on data offered by the Hellenic Open University and representative machine learning algorithms. The experimental results demonstrate that the early prognosis of students at risk of failure can be accurately achieved compared to supervised models, even for a small amount of initially collected data from the first two semesters. The robustness of the applying semi-supervised regression scheme along with supervised learners and the investigation of features’ reasoning could highly benefit the educational domain.