Keywords

1 Introduction

Deep Learning (DL), is a subset of Artificial Intelligence (AI), and Machine Learning (ML) approaches are entirely based on Artificial Neural Networks (ANN). ANNs are designed to mimic the human brain; hence deep learning is likewise a human brain mimic. DL algorithms are in high demand due to the exponential growth of data in all domains and the need for analysis and inferences from huge data. DL algorithms have proven to be highly powerful among the different popular techniques employed in ML. Their success is credited to the availability of a vast amount of data and the power of the graphical processor units. Deep learning models are used in several applications such as health care, autonomous vehicles, e-commerce, personal assistance, computer vision, etc. Several frameworks are available for implementing DL algorithms, each tailored to a certain context; TensorFlow, Pytorch, and Keras are the most well-known frameworks. Likewise, there are several types of DL architectures available. The multilayer perceptron (MLP), the convolutional neural network (CNN), the recurrent neural network (RNN), the self-organizing map (SOM), the generative adversarial network (GAN), and the auto-encoders (AE) are some of the most popular DL architectures. Each of these architectures has its own advantages and they work well for certain applications compared to others. In this section, we briefly discuss MLP, CNN, and RNN architectures in light of their application to landslide studies.

The MLP, or multilayer feed-forward neural network, is the first and most basic DL architecture. It has one input layer that receives the signal, one output layer that predicts the input, and numerous hidden layers that act as the network’s computational engine. Each layer has a different amount of neurons, and they are all fully linked. With a significant amount of training data, these networks are commonly utilized for regression and classification purposes. Backpropagation is used to train MLP architectures, which belong to the class of supervised learning techniques. Each perceptron in the network is connected to every other perceptron; therefore, the total number of parameters can quickly grow. Another drawback of MLP is that it does not take into account geographical information when making predictions. It uses flattened vectors as inputs, allowing it to make pixel-by-pixel predictions, which makes its usage limited in landslides studies that require spatio-temporal correlation.

CNN overcomes the inadequacies of MLP networks. CNNs are specifically designed for 2D data or structures such as images, videos, or sequences. CNN basically performs successive convolution operations to the input data in order to identify features and minimize the data sizes automatically. The CNN network has four basic layers in addition to the input and output layers: a convolution layer, a pooling layer, a flattening layer, and a fully connected layer. The main advantage of CNN is that it gathers spatial contextual information without human interaction by detecting relevant features through its convolutional layer. In the realm of computer vision and image interpretation, semantic segmentation of images using CNNs has produced promising results. Researchers are now using the same methodology to study landslides using satellite and aerial images such as susceptibility mapping, hazard mapping, landslide identification, etc.

While CNNs work well for extracting spatial information, RNNs are designed for sequential data, such as in time series, sound, or natural language processing. RNN networks have a feedback loop that allows them to make predictions using information from prior layers. However, two significant difficulties might arise while using RNN architectures: vanishing and exploding gradients. Two RNN variations have been created to address these limitations: the long short-term memory (LSTM) and, more recently, the gated recurrent unit (GRU). LSTM and GRU have nearly identical core ideas, have gated systems, and can handle longer sequences of tasks. The input, output, and forget gates are used in the LSTM design, whereas the reset and update gates are used in the GRU. In landslide studies, RNN and its variants are used to analyze long time-series data such as rainfall, moisture, displacement, etc.

In this paper, we have examined deep learning approaches applied in real-world case studies. We have reviewed the case studies based on (i) landslide mapping, (ii) landslide detection, (iii) landslide monitoring, and (iv) landslide prediction applications. In each of these reviews, the deep learning methodology used, the study area where it is implemented, additional benchmark algorithms implemented, model assessment metrics, the best model selected, and the limitations mentioned in the study by the authors are mentioned. Finally, the generic limitations in adapting deep learning to landslide studies are explored, as well as new prospects and breakthroughs that must be developed are discussed.

2 Paper Selection Criteria

Most of the papers retrieved by Google Scholar while searching for “deep learning and landslides” were about landslide susceptibility mapping and detection. Deep learning is commonly used for mapping applications in the landslide domain. This technology is directly derived from deeper architectures that are built and developed for applications like computer vision and augmented reality that works with RGB images. Other landslide studies, such as prediction and monitoring applications, use deep learning less than the landslide mapping studies. One reason for this is that deep learning outperforms most of the other ML-based algorithms when the amount of data is large. The availability of such large data in natural hazard areas like landslide monitoring and prediction is a limitation, and data with actual landslide incidents is indeed limited. To identify other landslide research that uses deep learning, we used Google Scholar’s advanced search option, excluding the words ‘susceptibility’, ‘mapping’, and ‘detection’ from the search. This paper considers the most relevant papers with the most citations from January 2017 to May 2022 for review.

3 Deep Learning in Landslide Susceptibility Mapping

The study Prakash et al. (2020) describes landslide mapping using CNN as a semantic segmentation problem. The study area is 1270 km2 in size and is located in Douglas County, Oregon, USA. The mapping for this investigation was done with a high-resolution Lidar DEM and a cloud-free optical image from Sentinel-2. Pixel-based, object-based, and DL methods are implemented for generating landslide susceptibility maps. This paper introduces CNN-based U-Net and ResNet architectures for mapping landslides. U-Net was introduced in 2015 by Ronneberger et al. (2015) for segmentation in biomedical images and has been modified to be used for mapping from satellite images (Peng et al. 2019; Schuegraf and Bittner 2019). U-net architecture is used for semantic segmentation of landslide-affected regions, and ResNet is used for feature identification in this paper. This paper applies U-Net architecture with a ResNet34 architecture for feature extraction backbone for landslide mapping and compares it to existing pixel-based and object-based machine learning approaches. The authors have demonstrated that the U-Net with ResNet34 strategy outperforms pixel-based and object-based machine learning algorithms on a regional scale. In the study area mentioned in Prakash et al. (2020), all three approaches identified landslides greater than 0.21 km2. However, all three methods were unsuccessful in detecting minor landslides. The pixel-based method performs best in detecting minor landslides; however, it has an extremely high probability of false detection. Authors also highlight that when landslides are close together, all approaches have trouble distinguishing individual landslides, and they prefer to predict them as one large landslide area. As a result, the predicted landslide profile does not resemble the actual landslide profile (Table 1).

Table 1 Deep learning case studies in landslide susceptibility mapping applications
Full size table

The researchers looked into the possibility of a deep convolution neural network for spatially predicting landslide susceptibility (Azarafza et al. 2021). The technique was tested using data from Iran’s Isfahan province (Azarafza et al. 2021). The landslide inventory dataset consisted of indices linked with 222 historical landslide occurrences, which were randomly separated into training (80%) and testing (20%) sets for the analysis. Based on field and remote sensing investigations, four key covariates were identified: geomorphologic, geologic, environmental, and human activity-related covariates. The deep convolution neural network model could accurately construct a susceptibility map for the research area. Compared to the benchmark models, the results reveal a considerable improvement in landslide susceptibility prediction accuracy. Despite the increased accuracy of the proposed deep convolution neural network, and predictive model for landslide susceptibility mapping, the authors highlight certain limitations to this work that should be taken into account in future research. (i) Fieldwork, historical landslide records, and remote-sensing data were used to create the primary database. Modeling was difficult due to the limited number of reference landslides in the recorded data (as is often the case); (ii) The quality of the input database was directly affected by the spatial resolutions of satellite imagery and DEM data quality, which influence the input data. (iii) During landslide susceptibility evaluations, the predictive model requires powerful processors to manage the inputs. The authors also addressed the question of whether landslides should be regarded as spatially continuous occurrences or spatial objects and another open research question.

To predict the susceptibility of future landslides in densely populated urban areas in Mt. Umyeon, Seoul, Korea, deep learning methodologies were used in Lee et al. (2020). Aerial photographs and a landslide inventory were used to create Deep Neural Network-DNN, kernel-based DNN, and convolutional neural network models. The average precision score and root mean square error for each of the three models were used to assess model performance. The average precision score curve revealed that the DNN, kernel-based DNN, and CNN models performed at 99.45%, 99.44%, and 99.41%, respectively. The accuracy of all three models was greater than 99%, indicating that each model is very good at predicting landslides. However, the kernel-based DNN and CNN models outperformed the DNN model slightly. This is evident that the models are based on the kernel approach. Lee et al. (2020) also mention one limitation in terms of choosing the appropriate hyperparameters. In order to understand the fundamental relationship between variables and landslides, it is necessary to apply a different methodology in advance when selecting variables, which requires experience.

The development and validation of a spatially explicit deep learning neural network model for predicting landslide susceptibility are described in Van Dao et al. (2020). Based on 217 landslide events from the Muong Lay district of Vietnam, a geospatial database was created, from which a set of nine landslide conditioning factors was derived with the help of the Relief-F feature selection method. Several performance metrics showed that the DL model performed well in terms of goodness-of-fit with the training dataset (AUC = 0.90; accuracy = 82%; RMSE = 0.36) and ability to predict future landslides (AUC = 0.89; accuracy = 82%; RMSE = 0.38). The model’s efficiency was compared to quadratic discriminant analysis, Fisher’s linear discriminant analysis, and a multilayer perceptron neural network. A Wilcoxon signed-rank test comparison revealed that the spatially explicit DL model outperformed the other models in terms of landslide prediction.

Landslide susceptibility assessments and comparison of its predictive performance to state-of-the-art machine learning models were implemented in Bui et al. (2020a, b). The efficiency of the Deep learning neural network model was estimated for the Kon Tum Province of Vietnam, which is characterized by the presence of landslide phenomena. 1657 landslide locations and nine landslide-related variables were used to generate the training and validation datasets for the landslide susceptibility assessment. The deep learning neural network model’s learning ability was evaluated and compared to a Multilayer Perceptron Neural Network, a Support Vector Machine, a C4.5-Decision Tree model, and a Random Forest model. To assess each model’s learning and predictive capacity, the classification accuracy, sensitivity, specificity, and area under the success and predictive rate curves (AUC) were calculated. According to the results, the proposed deep learning neural network model outperformed the four benchmark models. The study concludes that using a deep learning approach could be a viable alternative approach for landslide susceptibility mapping.

For a national-level landslide susceptibility mapping in Iran, two unique deep learning algorithms: the recurrent neural network (RNN) and the convolutional neural network (CNN), are used and compared in Ngo et al. (2021). A geospatial database was created with 4069 historical landslides and 11 conditioning factors. The data was split into two datasets: training and testing. RNN and CNN algorithms built landslide susceptibility maps for Iran using the training dataset. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) are used to quantify the landslide susceptibility maps using the testing dataset. The RNN algorithm (AUC = 0.88) outperformed the CNN method (AUC = 0.85) in both the training and testing phases. Authors assessed regions of susceptibility for each province and discovered that 6 and 14% of Iran’s land area are very highly and highly vulnerable to future landslides. Approximately 31% of Iran’s cities are prone to landslides. The findings of this research will aid in developing landslide risk mitigation strategies. Ngo et al. (2021) also highlight previous research, limitations, and future recommendations. The design of CNN and RNN algorithms, including the training technique, input window size, and layer depth, significantly impacts their performance. These algorithms yield consistent output sequences, yet they have unidentified or unacknowledged error sources; hence their results should be interpreted with caution. The authors also point out that, when compared to other models, RNN enhances landslide prediction accuracy, particularly in tropical locations.

Another study in Maoxian County, Sichuan, China, applied four deep learning algorithms; namely, (1) the convolutional neural network (CNN), (2) deep neural network (DNN), (3) long short-term memory (LSTM) networks, and (4) recurrent neural network (RNN), were used to assess the risk of landslides (Habumugisha et al. 2022). With historical records, field observations, and remote sensing techniques, a total of 1290 landslide records were created. According to this study, the DNN algorithm outperformed the LSTM, CNN, and RNN algorithms in detecting landslides in Maoxian County, with AUC values of 87.30%, 86.50%, 85.60%, and 82.90%, respectively.

Several other studies have also employed deep learning algorithms to map landslide susceptibility worldwide. However, most studies are primarily concerned with using deep learning approaches to achieve high accuracy in landslide mapping when compared to existing benchmark machine learning algorithms. There is little research focusing on making modifications and improvements to address the image segmentation problem for landslides, which is distinct from image segmentation in other disciplines such as biomedical images. On the other hand, landslide mapping is an unbalanced learning problem with more data belonging to the no landslide category and extremely limited data belonging to the landslide category. In particular, approaches to coping with unbalanced data and the segmentation process for landslides, must be specifically created. Few researchers have used deep learning to tackle landslide mapping as a semantic segmentation problem (Prakash et al. 2020). Conditional random fields have been successfully employed to post-process segmentation results in other disciplines (Christ et al. 2016). These methodologies specific to the context of landslide mapping should be examined more in the future.

4 Deep Learning in Landslide Detection

The study in Bui et al. (2020a, b) explores a system for detecting landslides from satellite images that combines deep learning and image transform algorithms. A convolution neural network is utilized in the deep learning section to classify satellite photos that contain landslides. This work presents a transformation technique, Hue–Bi-dimensional empirical mode decomposition, to calculate the landslide region and magnitude from landslide photographs categorized in order to reliably identify landslides under diverse illumination situations. After detecting the landslide’s position, the authors determine the landslide’s size changes over time. This paper presents the results of a simulation study using a limited set of satellite images, and it has not been tested or validated for identifying landslides in a real scenario using satellite images (Table 2).

Table 2 Deep learning case studies in landslide detection applications
Full size table

In the study in Kamiyama et al. (2018), the ability of CNN algorithms to detect landslides from differential interferograms is tested. Changes in the interference fringes in DInSAR differential interferograms could be caused by things other than ground motions too. Local impacts that occur on slopes are difficult to separate from global influences on a wide spatial scale. However, separating local impacts that appear on slopes from global effects on a large spatial scale is difficult. Moreover, experts would need a lot of time and effort to evaluate all of the differential interferograms formed from observational data collected at high frequency over large areas. The work in Kamiyama et al. (2018) examined the effectiveness of adopting a CNN model to detect interference fringes that may indicate landslides from differential interferograms with the goal of efficiently detecting interference fringes that may represent landslides. CNN models were found to be capable of detecting interference fringes with the possibility of landslides with high reproducibility, with recall values of around 90% in the validation instance. On the other hand, Landslide interference fringes were discovered in greater numbers than in the training data, indicating that the precision is low. Expert-like procedures were used to create fringes that represented landslide motions.

For landslide detection from satellite photos, a preprocessing method based on Bi-dimensional empirical mode decomposition is used along with deep learning by Bui et al. (2019). The results are better with this combination than with an individual CNN training model or solely identifying using the Bi-dimensional empirical mode decomposition process. This paper also presents the results of a simulation study using a limited set of satellite images. It has not been tested or validated for identifying landslides in a real scenario using satellite images.

Orland et al. (2020) suggested a novel framework for detecting and classifying natural disasters. The system relies primarily on a hybrid of the convolutional neural network (CNN) for feature extraction. Due to the lack of a dataset concerning several disasters, a new dataset was created to evaluate the framework’s capabilities. The model is made up of five separate CNNs that have been grouped together. Each CNN model makes use of a pre-trained AlexNet architecture on the ImageNet dataset that has been fine-tuned for the generated dataset. The proposed method employs CNN and SVM to identify and classify ten different types of natural disasters. The proposed framework was observed to outperform when compared with the state-of-the-art algorithms. According to the authors, this model can be used further by satellite and aerial real-time image processing systems to locate the geographical locations of places affected by these natural disasters (Table 3).

Table 3 Deep learning case studies in landslide monitoring applications
Full size table

5 Discussion

Deep learning techniques are now applied in almost all real-world applications, and the landslide domain is no exception. However, research into the use of DL algorithms for landslides is still in its early stages and has some unique challenges. The major factors that limit the use of DL for landslide research are summarized in this section.

Techniques are not well utilized for real-time monitoring and forecasting.

  1. (i)

    To acquire satisfactory results using DL algorithms, it is required to employ a large training dataset. Despite the fact that certain training datasets are available in the remote sensing community, there is, however, a major lack of data categorized as landslides, in contrast to the techniques such as computer vision, augmented reality etc., which uses large image datasets for creating deep learning models. Working with landslides, the issue of limited training samples is much more important. To get accurate predictions, it is essential to construct a large time series of data belonging to the landslide category.

  2. (ii)

    The majority of landslide studies rely on satellite data, and unlike RGB images, these images are complex and varied, the preprocessing procedure takes a long time and requires the assistance of a remote sensing expert. Although optical images are utilized, the presence of clouds on the photographs is a serious concern. Even though several strategies are used to overcome this limitation, the quality of the data produced is altered and as a result, the accuracy of prediction also decreases.

  3. (iii)

    Existing deeper architectures are mostly designed and developed for applications such as computer vision, and augmented reality that works with RGB images. Landslide studies, on the other hand, use (a) satellite and aerial images that are extremely large and may or may not be in RGB bands; (b) instrumental data such as geotechnical, geophysical and IoT data that are one dimensional or two dimensional time series unlike RGB data. When the size of images and sequence lengths rise, Moskolaï et al. (2020) state that using the ConvLSTM one among the deep learning architecture is not recommended. Therefore it is necessary to build deep learning models and architectures that are specific and suitable for natural hazard studies such as landslides.

  4. (iv)

    Existing DL methods are majorly utilized for mapping applications when it comes to landslide research. Application of DL in prediction, monitoring and early warning are limited and this is another area worth investigating because this component of landslide study indeed requires more precise results and forecasts.

6 Conclusion

This article reviewed recent papers from 2017 to till date that used Deep learning algorithms and architectures to study landslides. Deep learning finds its major application in susceptibility mapping of landslides. Promising results are obtained in all the studies that used deep learning approaches compared to the benchmark algorithms such as SVM, Naive Bayes, Decision trees, Neural networks, etc. Despite the positive results produced from deep learning approaches, limitations still exist, and these limitations are summarized in the discussion section. The major limitations are the lack of available training datasets, existing deep learning architectures and models are not fine-tuned to suit landslide data such as satellite images, the complexity of satellite images, the need of preprocessing, etc. Therefore, it is necessary to (i) design architectures and frameworks that are dedicated to natural disaster studies involving both satellite and ground data. (ii) Fine tune the existing deep learning architectures to be able to produce efficient outcomes for research of natural disaster studies such as landslides.