Encyclopedia
Please note this is an old version of this entry, which may differ significantly from the current revision.
Landslide detection can generally be categorized into two approaches: traditional methods of landslide identification and automatic identification methods based on machine learning algorithms. Traditional methods of landslide detection often rely on field surveys conducted by experienced geologists, complemented by instrumental imaging techniques for analysis. The second category predominantly utilizes pre-existing datasets of landslides and facilitates automatic identification through the construction of algorithmic models.
- landslide detection
- machine learning
1. Introduction
Landslides are severe geological hazards that widely occur in mountainous environments with slopes and frequently lead to chain reactions such as mountain collapses and debris flows, which can pose serious risks to human life and property. Therefore, enhancing the detection and early warning systems for landslide-related geological catastrophes holds considerable implications in the context of China’s endeavors towards disaster mitigation and risk reduction [1,2].
Traditional landslide detection methods primarily rely on geologists, which often entails significant manpower and financial investments. However, the effectiveness of these methods may not always meet expectations. In light of the advancements in satellite imaging accuracy, researchers have increasingly proposed landslide detection approaches that leverage optical image data. Concurrently, machine learning has become increasingly popular in the field of landslide detection. Besides, the emergence of convolutional neural networks (CNNs) has led to the successful application of deep-learning-based object recognition algorithms in landslide detection, and they have gradually become mainstream. In contrast to machine learning approaches, deep learning techniques abandon the complicated artificially designed features, which adopt deeper convolutional neural networks to automatically acquire distinguishing characteristics. Furthermore, the data sample capacity of deep learning for landslide detection can be extensive, rendering it more appropriate for large-scale landslide identification and endowing it with a more-robust generalization capability. Although deep-learning-based algorithms have shown success in detecting landslides, they still face some challenges: These models necessitate substantial computational resources and numerous parameters, leading to diminished inference efficiency. As a result, employing them in power-limited contexts or embedded platforms with the objective of achieving real-time detection becomes challenging. Moreover, the scarcity of open-source repositories containing high-spatial-resolution images of landslides hampers the effective training and validation of these models.
2. Approaches of Landslide Detection
Landslide detection can generally be categorized into two approaches: traditional methods of landslide identification and automatic identification methods based on machine learning algorithms.
Traditional methods of landslide detection often rely on field surveys conducted by experienced geologists, complemented by instrumental imaging techniques for analysis. These methods involve on-site inspections, geological mapping, and the collection of ground truth data, for example using interferometry synthetic aperture radar (InSAR) technology to obtain multi-temporal data to observe whether the slope is deformed, which can be used as a basis to infer potential landslides [3]. While traditional methods have been widely practiced and have proven effective, they have the limitations of being time-consuming and resource-intensive.
The second category predominantly utilizes pre-existing datasets of landslides and facilitates automatic identification through the construction of algorithmic models. Generally, automatic landslide detection techniques can be categorized into machine learning approaches and deep learning approaches. Machine learning algorithms encompass methods such as Bayesian, logistic regression, support vector machines (SVMs), random forests, and artificial neural networks [4,5,6,7], which can utilize various features related to landslide occurrence, such as texture and terrain information for classification and prediction. For instance, Pourghasemi [8] applied random forest to evaluate the sensitivity of landslides, and Tien [9] utilized SVM and kernel logistic regression for landslide recognition. Artificial neural networks, including pulse-coupled neural networks (PCNNs) and spiking neural networks, have been shown to possess outstanding capabilities in image fusion and computer vision applications [10,11].
Owing to the swift advancements in hardware equipment and artificial intelligence, deep learning techniques have emerged as an additional potent data-driven approach for detection. Consequently, a multitude of sophisticated object-detection algorithms have surfaced, including two-stage object-detection algorithms represented by region-based convolutional neural networks (RCNNs), Fast R-CNN, and Faster R-CNN [12,13] and single-stage object detection networks represented by the SSD algorithm [14] and the YOLO algorithm series [15,16,17,18]. For instance, Ju [19] selected the YOLOv3 and Mask RCNN algorithms to achieve automatic recognition of loess landslides, with an optimal average precision of 18.9%. Ji [20] proposed an enhanced convolutional neural network for landslide detection of Bijie City, which demonstrated the effectiveness of a new landslide prediction based on the dataset. Hou [21] incorporated a coordinate attention mechanism [22] to enhance the YOLOX [23] object-detection model, effectively tackling the problem of the poor detection of complex mixed landslides. Tang [24] proposed SegFormer, a model based on the Transformer architecture, which is capable of automatically detecting landslides.
In recent years, various lightweight neural architectures such as MobileNetV1-3 [25,26,27], ShuffleNet [28], GhostNet [29], and FasterNet [30] have followed, aiming to achieve fewer parameters, a fast inference speed, and high performance. In brief, MobileNet incorporates depthwise separable convolution and an inverted residual structure to decrease the computational expense while simultaneously improving the detection performance. ShuffleNet substitutes the 1 × 1 convolution with group convolution and incorporates a shuffle operation to facilitate information flow among various groups. In GhostNet, to avoid redundant features maps, the spatial features are only captured by inexpensive operations for half of the features. FasterNet proposes a partial convolution to extract features with an efficient and parameter-friendly manner. Besides, residual connections [31] and dense connections [32] are widely used in these network designs to alleviate gradient degradation problems and aggregate features with diverse receptive fields.
This entry is adapted from the peer-reviewed paper 10.3390/app13158583
This entry is offline, you can click here to edit this entry!
