Rapid detection of landslides using remote sensing images benefits hazard assessment and mitigation. Many deep learning-based models have been proposed for this purpose, however, for small-scale landslide detection, excessive convolution and pooling process may cause potential texture information loss, which can lead to misjudgement of landslide target. In this paper, we present a novel UNet model for automatic detection of landslides, wherein the reversed image pyramid features (RIPFs) are adapted to compensate for the information loss caused by a succession of convolution and pooling. The proposed RIPF-Unet model is trained and validated using the open-source landslides dataset of the Bijie area, Guizhou Province, China, wherein the precision of the proposed model is observed to increase by 3.5% and 4.0%, compared to the conventional UNet and UNet + + model, respectively. The proposed RIPF-Unet model is further applied to the case of the Longtoushan region after the 2014 Ms.6.5 Ludian earthquake. Results show that the proposed model achieves a 96.63% accuracy for detecting landslides using remote sensing images. The RIPF-Unet model is also advanced in its compact parameter size, notably, it is 31% lighter compared to the UNet + + model.
Research Article
RIPF-Unet for regional landslides detection: a novel deep learning model boosted by reversed image pyramid features
https://doi.org/10.21203/rs.3.rs-1886017/v1
This work is licensed under a CC BY 4.0 License
Journal Publication
published 28 Aug, 2023
You are reading this latest preprint version
Landslide
detection
reversed image pyramid
UNet
remote sensing image
As a ubiquitous and typical geological hazard, landslides significantly endanger human lives, infrastructures, and other property security, and occur frequently worldwide (Cigna et al. 2018). A reliable inventory of landslides is therefore of critical importance for timely quantitative hazard assessment, disaster relief, and subsequent risk governance, especially in mountainous regions with complex environments and inconvenient transportation (Galli et al. 2008).
Traditionally, in-site manual investigation was the most commonly used scheme for landslide detection, which required experts and investigators to reach the landslide area and evaluate the landslides' magnitude, position, and hazard mechanisms (Ji et al. 2020). However, it is a time-consuming, laborious task, even sometimes risky (Li et al. 2022). With the boom of remote sensing technology, as well as improvements in the capacity of obtaining geospatial information on a regional scale, remote sensing technology has been extensively employed in various fields, such as for the investigation purpose of hydrology (Brinkhoff et al. 2022; Rajesh et al. 2021), geology (Barak et al. 2021), soil (Yu et al. 2022; Lourenco et al. 2022), environment (Petaja et al. 2022; Wang et al. 2021; Wiggins et al. 2021) and forestry (Eskandari et al. 2020; Altamirano et al. 2020). These remarkable works inspire the application of remote sensing technology in landslide detection. When high-resolution images are available, visual interpretation is a reliable method that creates appropriate interpretation markers to show the boundaries of landslides by considering the landslides’ colour, texture, shape, position, and other factors hidden in the remote sensing images. These markers are used by experts to further analyse and confirm the landslide area (Tanoli et al. 2017). The basic principle of using remote sensing images for landslide interpretation can be summed up as the spectral difference caused by the loss of vegetation and the exposure of fresh soil and rock (Li et al. 2014).
Although visual interpretation has been proved able to generate good results in landslides identification (Xu et al. 2015), it is usually inefficient and may suffer from interpreters' subjectivity (Fan et al. 2019). Thus, several computer visual mothods and machine learning based interpretation methods (Li et al. 2022; Chen et al. 2020; Merghadi et al. 2020; Nhu et al. 2020; Han et al. 2019; Tien Bui et al. 2019; Fan et al. 2019; Sun et al. 2017; Li et al. 2016; Behling et al. 2016; Li et al. 2014), including linear regression (LR), support vector machine (SVM), random forest (RF), k-nearest neighbour (KNN), and thresholding method, have been conducted to address these issues. These remarkable studies laid a solid foundation for landslide research. However, these works generally cost a lot of time in feature analysis, which consists of enumerating and selecting landslide conditioning factors (e.g., NDVI, NDWI, NDBI, TWI). Meanwhile, some of the crucial parameters in these methods (e.g., kernel function in SVM, max tree depth in RF, k value in KNN), require empirical adjusting, somewhat limiting the methods' flexibility. For instance, as we mentioned in our previous studies (Han et al. 2019; Li et al. 2014), grey-level threshold is crucial for panchromatic image binarization in landslide detection, but the determination of which is still debated.
Recently, convolutional neural networks (CNNs) have shown significant learning capacity in a variety of image-based tasks, including image classification (He et al. 2016; Simonyan and Zisserman 2014), object recognition (Ren et al. 2017), and semantic segmentation (Chen et al. 2018; Ronneberger et al. 2015). Many CNN algorithms have been proposed and implemented during the last two decades (Howard et al. 2017; Chollet 2017; He et al. 2016; Huang et al. 2016; Simonyan and Zisserman 2014), each based on a different rationale. In terms of landslides, the CNNs were used to extract the hidden features of the landslides in the remote sensing images, which helped to determine the existence of landslides at the location (Ding et al. 2016). Subsequently, CNNs have been widely used to retrieve images containing landslides from image datasets. Many studies (Yu et al. 2017; Li et al. 2014), used specific algorithms, such as automatic thresholding and region growth algorithms, to delineate the region of the landslides.
Up to date, various structures of CNN networks, with different convolutional layers, sizes, and input data channels, have been discussed to analyse the impact of network settings against accuracy (Ghorbanzadeh et al. 2019). In addition, multi-channel inputs of the combined spectral and topographic features were used to build CNN models (Su et al. 2021; Meena et al. 2021), which improved the accuracy of landslide detection. The convolution kernel of CNN can automatically obtain the effective features of images, which enables the understanding of semantic information of objects avoiding empirically determination of complex features (Lecun et al. 2015). Additionally, via network training, convolution kernel can further emphasise the distinction between landslide and background, which is essential for the purpose of landslide detection. Thus, CNN can be selected as an important method for image segmentation and inspection of landslides.
Among the whole CNN methods, the UNet model has been proven to be an efficient method for landslide detection (Meena et al. 2022; Yu et al. 2021; Prakash et al. 2020; Lei et al. 2019). With a succession of convolution, pooling, and deconvolution layers, this method tries to extract the required information from remote sensing images (Badrinarayanan et al. 2017; Shelhamer et al. 2017). However, the scale of landslides is different in remote sensing images. For tiny and slender landslide detection, the excessive convolution and pooling process may cause potential texture information loss (Li et al. 2019; Gu et al. 2019). For instance, a \(25\times 25\) pixel landslide target tends to be smaller than \(1\) (\((25÷{2}^{5})<1\)) after five pooling processes, which makes its texture information disappear on the feature map (Fig. 1). The lost texture information may lead to the following misjudgment of the landslide target. As a result, to successfully detect landslides, especially for a remote sensing image covering a large area, the UNet model needs to be revised so that it can detect landslide boundaries precisely.
In this paper, to address the issue of potential texture information loss caused by convolution and pooling in conventional CNN models, a reversed image pyramid feature (RIPF) boosted UNet algorithm for landslide detecting is presented. In this method, the original network structure of UNet models is utilised as the major network structure, while the reversed image pyramid features are employed to augment the feature in the network's decoding phase to increase the model's capacity to differentiate the area of landslides under a complicated backdrop. The RIPF-Unet model is supposed to better identify tiny and slender landslides from exposed surfaces. To verify the performance in RGB remote sensing image dataset, the proposed RIPF-Unet model is trained and tested using an open-source dataset, and the adaptability and effectiveness of the method in the actual area are further tested in the case of the Longtoushan region after the 2014 Ms.6.5 Ludian earthquake.
2.1 Convolutional neural network
The CNN-based networks receive images as input and automatically adjust network parameters during iterative training. The trained network, consisting of thousands of weights at the node of neutral network, will predict with the greatest degree of certainty. Convolution layers, activation layers and pooling layers are common components of CNN designs (Su et al. 2021). In general, the convolutional layer is usually followed by an activation operation, which is utilized to extract the nonlinear mapping of features \({F}_{i}\left(Y\right)\) according to Eq. (1), and improve the nonlinear representation capability of neural networks.
| \({F}_{i}\left(Y\right)=\sigma \left({W}_{i}\times Y+{B}_{i}\right)\) | (1) |
where \({F}_{i}\left(Y\right)\) in Eq. (1) represents the feature graph produced by image \(Y\) through the convolution layer and activation layer. \({W}_{i}\) and \({B}_{i}\) are independent representations of the convolution kernel and bias term that are trained through the network, respectively. The convolution layers are designed for image feature extraction. It is the matrix's dot product and sum operation. Convolution kernels of various sizes \((3\times 3, 5\times 5, 7\times 7)\) can be used in the convolution layers, and the typical convolution process is shown in Fig. 2a.
In Eq. (1), \(\sigma\) represents an activation function, which is commonly linked after the convolution layer to increase the neural network's nonlinear representation capacity. The typical activation functions are ReLU which could be represented by \(\left(\text{m}\text{a}\text{x}\right(0,x\left)\right)\) and Sigmoid \((1/(1+{\text{e}}^{-x}\left)\right)\), where \(x\) refers to an individual element in the image. The activation process using ReLU functions is shown in Fig. 2b. The pooling layer is used after several convolutions and activation layers to compress the feature graph for efficient computation. Max-pooling and average-pooling are two major pooling types that preserve the maximum and average of the pool window in the input feature map, respectively. The pool window is typically \(2\times 2\) in size, which corresponds to a downsampling process of \(2\times\). Figure 2c depicts the max-pooling process.
2.2 The conventianal UNet model
UNet is a Fully Convolutional Neural Network proposed to solve image-to-image mapping tasks (Fernández and Mehrkanoon 2021). Different with typical CNN architectures, UNet exploits a contracting and an expansive path during the convolution and pooling process. The former acts as an encoder of input information, and the latter is a decoder that performs classification on each pixel to reconstruct the segmented output. The encoder of UNet is changeable and aims to extract features from the input image (Liu and Wang 2022). In this study, the VGG-16 network was adopted as the encoder network for feature extraction. Meanwhile, a technique called skip connection creates a bridge between the encoder and the decoder, which was used to copy feature maps. Both contracting and expansive paths are more or less symmetrical in their operations, forming a U-shape process.
According to the structure of UNet in Fig. 3a, the contracting path (left side) contains five convolution blocks. Each convolution block consists of several \(3\times 3\) convolution operations (followed by a ReLU activation function) and a \(2\times 2\) max-pooling operation of stride \(2\) for downsampling. On the other hand, each convolution block in the expansive path (right side) consists of a \(2\times 2\) up-convolution operation for upsampling, a concatenation with the corresponding feature map from the contracting path, and two 3x3 convolutions, each followed by a ReLU activation function. To alleviate the checkerboard effect (Odena et al. 2016), the up-convolution step is replaced with a \(2\times 2\) upsampling step in this work. In the final stage, the extracted feature map convolves with a \(1\times 1\) kernel and is followed by a sigmoid activation function to carry out the pixel-wise classification part.
2.3 The proposed RIPF-Unet model
Through a succession of convolution and pooling layers, the UNet model extracts image feature maps and reconstructs segmented outputs. However, multiple stages of spatial pooling and convolutional layers may cause the potential loss of fine image structure information (Wang et al. 2017). Therefore, in the UNet model, the skip connection was employed to concatenate feature maps from the contracting path and expansive path for information transfer. In the same way, the UNet + + model adopts several short connections to minimise the information loss after base-level convolution.
The possible loss of image information in base-level convolution, especially high-frequency detailed information, is essential for landslide detection. Therefore, in this study, the reversed image pyramid features are employed to further expand the feature map in the expansive path to recover the information loss caused by convolution. Specifically, the initial image pyramid features are further stacked onto the output of the blocks in expansive path. The proposed network architecture is shown in Fig. 3b.
To stack image pyramid features into the feature maps in expansive path, a novel concatenation module (Fig. 4) has been proposed, which consists of three parts, i.e., the feature maps extracted by the contracting path, the original image pyramid feature maps, and the feature maps generated by the expansion path. Through average-pooling, the pyramid feature maps of the original image are scaled to the corresponding size that the concatenation operation requires. Notably, we extended the feature maps at each stage of the expansive path. After the output feature maps of the last convolution block in the expansive path, a convolution kernel of \(1\times 1\) is connected to control the number of channels, and the pixel-level binary classification result is obtained after a sigmoid activation function.
2.4 Loss functions and evaluation indicators
The proposed method uses a cross-entropy loss function to update the parameters of the network. The equation for calculating the cross-entropy loss is shown in Eq. (2):
| \(\text{l}\text{o}\text{s}\text{s}=-\frac{1}{n}\sum _{i=1}^{n} {y}_{i}\text{l}\text{g}\left({t}_{i}\right)+\left(1-{y}_{i}\right)\text{l}\text{g}\left(1-{t}_{i}\right)\) | (2) |
where \({t}_{\text{i}}\) represents the predicted distribution, \({y}_{\text{i}}\) represents the true distribution. The error between the predicted results and the ground truth is calculated using the loss function. The backpropagation updates the network parameters to minimize the value of the cross-entropy loss function.
To evaluate the model’s performance, many accuracy indexes can be referred to. In our study, we employ four indexes, including Precision, Recall, Accuracy, and F1 score, as shown in Eq. (3–6), to assess the network's recognition ability,
| \(\text{Precision}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}\) | (3) |
|---|---|
| \(\text{Recall}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{N}}\) | (4) |
| \(\text{Accuracy}=\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{F}\text{P}+\text{T}\text{N}+\text{F}\text{N}}\) | (5) |
| \(\text{F}1=2\times \frac{\text{ Precision }\times \text{ Recall }}{\text{ Precision }+\text{ Recall }}\) | (6) |
Results of true positives, true negatives, false positives, and false negatives are represented by \(\text{T}\text{P}\), \(\text{T}\text{N}\), \(\text{F}\text{P}\), and \(\text{F}\text{N}\) in the equation. To get a better segmentation result, the precision, recall and accuracy values should be as high as possible. However, when the segmentation target is smaller in the image, it will be a certain conflict between precision and recall. Thus, the comprehensive evaluation index F1 score is given to assess the model’s performance unbiasedly. Besides the abovementioned four indexes, to evaluate the model's predictive performance for an individual landslide in the dataset, the intersection over union (IOU) index in Eq. (7) is used.
| \(\text{I}\text{O}\text{U}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}+\text{F}\text{N}}\) | (7) |
3.1 Public data sets
To train the proposed RIPF-Unet model, an open-source dataset in the previous study (Ji et al. 2020) is used in this paper. The dataset contains both satellite optical images and landslides boundary shapefiles. The image data was obtained from TripleSat satellite imagery of the Bijie region, Guizhou Province, China (as shown in Fig. 5a), from May to August 2018. The satellite images were all pre-processed with radiometric calibration, atmospheric correction, orthographic correction, and image fusion. As shown in Fig. 5b, a total of 770 samples of landslides are included and manually delineated in this dataset.
The included landslides samples in this dataset vary in size, and the dataset was created using the anchor box of the present landslides example within a 40m×40m extension as a backdrop. The RGB remote sensing images have a ground resolution of 0.8m, and the shape vector of each landslide was produced using ArcGIS software. Some examples of the landslides and their boundaries are shown in Fig. 6.
3.2 Model parameter setting
In our study, 80% of the image data in the dataset is used for training, while the remaining 20% is for testing. To verify the effectiveness of the proposed model, the conventional UNet and UNet + + networks are also well trained for comparison. All networks have the same training setup. In detail, these models are trained for 20 epochs using the Adam optimizer with a batch size of 4. The original learning rate starts at 10− 4 and decreases exponentially to 10% after 5 epochs. The size of the input image is adjusted to \(256\times 256\) pixels, and the image pixel values are normalized to 
During the training process, the data is augmented using random rotation and cropping. The network's output is the probability that any pixel in the image belongs to landslides. The model is trained using a Windows Server equipped with a GeForce TITAN V 12G GPU and an Intel Xeon(R) Gold 6133 CPU. All code is implemented in the Tensorflow deep learning environment.
3.3 Analysis of training results
Table 1 lists the accuracy statistics of the different landslide detection methods on the training and validation datasets, which are schematically shown in Fig. 6. It demonstrates that all three models are able to detect landslides. The overall accuracy for the validation set of RIPF-Unet is 0.9725, the precision is 0.8991, the recall is 0.9490, and the F1-score is 0.9234. The statistical results of the validation dataset show that compared to the conventional UNet and UNet++, the proposed RIPF-Unet model obtains a better detection result. The precision index improves by 3.5% and 4.0%, respectively, while the F1 index improved by 2.3% and 2.7%.
|
Model |
Training data |
Validation data |
||||||
|---|---|---|---|---|---|---|---|---|
|
Precision |
Recall |
Accuracy |
F1 |
Precision |
Recall |
Accuracy |
F1 |
|
|
UNet |
0.8884 |
0.9621 |
0.9725 |
0.9238 |
0.8682 |
0.9390 |
0.9644 |
0.9023 |
|
UNet++ |
0.8863 |
0.9574 |
0.9713 |
0.9205 |
0.8644 |
0.9369 |
0.9633 |
0.8992 |
|
The proposed RIPF-Unet |
0.9199 |
0.9763 |
0.9811 |
0.9472 |
0.8991 |
0.9490 |
0.9725 |
0.9234 |
The above details support that the proposed RIPF-Unet model has a better accuracy index for the training dataset, compared to the conventional UNet and UNet + + models. This improvement can be explained by the fact that the RIPF-Unet model appropriately complements the decoding network’s information, increasing the abundance of features for decoding network extraction.
The proposed RIPF-Unet model has shown satisfactory accuracy using the training and validating dataset. In this section, to test the model’s performance for delineating landslides, the Longtoushan region after the 2014 Ms.6.5 Ludian earthquake is selected as a case study.
As shown in Fig. 9, the Longtoushan region (27°05′30″-27°07′43″N, 103°16′37″-103°19′27″E) locates in Ludian County, Zhaotong City, Yunnan Province. The area was impacted by the Ms.6.5 earthquake on August 3, 2014 (Chen et al. 2015). We manually delineated the boundary of landslides in the study area with the empirical interpretation of remote sensing images, and the result is as shown in Fig. 9d.
Figure 10 presents the final detection result using the proposed RIPF-Unet model. The components circled in red lines highlight the landslides, and the rest part refers to the background. To enhance the detection result, we highlight the TP, FN, and FP regions using different colours. Figure 10 demonstrates that most landslides can be effectively detected, with a few incorrect detections of the cloud-covered region. In terms of landslide boundaries, most landslide borders can be delineated, although some trivial errors exist compared to the manual delineated boundaries.
Figure 11 depicts the comparison detection results of sub-areas using different models. It is shown that all three models are able to detect the majority area of landslides. Compared to the other two models, RIPF-Unet successfully detects small-medium size landslides as shown in Fig. 11a and Fig. 11b, owing to the concatenation of reversed image pyramid features. In addition, the detected landslides’ boundary of RIPF-Unet is relatively consistent with the manually delineated boundary as shown in Fig. 11c and Fig. 11d. Further, a clear distinction between landslides and roads (Fig. 11e) is also observed for the prediction result by the proposed RIPF-Unet, which helps improve the detection accuracy.
For the detection results of landslides based on the three deep learning models, a statistical quantification analysis of the detection results is carried out compared to the manually visual interpretation results. Table 2 summarizes the quantitative assessment of the different landslide detection models. As shown in Table 2, the proposed RIPF-Unet model generally outperforms other models, respectively achieving the highest precision, recall, and accuracy of 0.6764, 0.8109, and 0.9663, and yielding F1 of 0.7376. This improved result is attributed to the correct detection of tiny and slender landslides, as well as more precise boundary detection, compared with other models.
Compared to UNet, our proposed RIPF-Unet has better assessment indexes with no substantial increase in parameter memory. In contrast with these baseline models, RIPF-Unet's recall metrics improved by 0.4% and 1.9%, respectively. While the F1 metrics improved by 1.6% and 1.6%, demonstrating the RIPF-Unet’s advantages in the landslide detection task.
| Model | Parameters Size | Assessment Index | |||
|---|---|---|---|---|---|
| Precision | Recall | Accuracy | F1 | ||
| UNet | 95.0MB | 0.6599 | 0.8071 | 0.9644 | 0.7261 |
| UNet++ | 138.0MB | 0.6674 | 0.7951 | 0.9649 | 0.7257 |
| RIPF-Unet | 95.1MB | 0.6764 | 0.8109 | 0.9663 | 0.7376 |
5.1 Impact of initial learning rate
Although the proposed RIPF-Unet model has a relatively better detection result, and the dynamic learning rate setting has a certain universality for model training, it still needs further discussion on the initial learning rate. Because the initial learning rate often determines whether the model can converge quickly in a short time (Shi et al. 2021), and is an essential network parameter set in the study. In this work, we have performed a sensitivity analysis using different initial learning rates, i.e., \(5{\text{E}}^{-5}\), \(1{\text{E}}^{-5}\), \(1{\text{E}}^{-4}\), and \(5{\text{E}}^{-4}\), against the model’s performance. However, in the actual experiment, the model cannot converge while the learning rate is set to \(5{\text{E}}^{-4}\), hence the influence of the learning rate of \(5{\text{E}}^{-4}\) is not taken into account here.
Using the RIPF-Unet model as an example, we trained the model using the rest three initial learning rates and calculated the loss in the validation dataset. As seen in the graph (Fig. 12), the loss under the three learning rates tends to be steady eventually. In general, a lower loss indicates a better performance of the model on the validation dataset. When the learning rate is set to \(1{\text{E}}^{-4}\), it can be seen that the overall loss of the model is superior. As a result, in the model training process, the initial learning rate was set to \(1{\text{E}}^{-4}\), and the model with the lowest loss on the validation dataset is chosen as the actual prediction model.
5.2 Impact of image pyramid’s embedding position
In the study, an important design of our method is to incorporate the image pyramid features into the UNet model. However, various embedding positions will result in different extraction results of the model, which potentially have an impact on the model's performance. In this experiment, we try to concatenate the reversed image pyramid features in three different embedding patterns in the model, i.e., embedding the image pyramid in the expansive path (RIPF-Unet-R), the contracting path (RIPF-Unet-L), and directly embedding the image pyramid in both the contracting path and the expansion path (RIPF-Unet-LR).
Figure 13 represent the loss curves of different models in the validation dataset, it can be found that the loss of RIPF-Unet-R, RIPF-Unet-L, and RIPF-Unet-LR that are embedded with the image pyramid is better than that of conventional UNet and UNet++, indicating that the method of embedding the image pyramid features is effective. Compared to these methods, it can be concluded that RIPF-Unet-R has a relatively lower loss. Therefore, incorporating the reversed image pyramid into the expansive path is a better choice.
5.3 Limitations and Future Work
Our proposed RIPF-Unet has achieved satisfactory results on the open-source dataset. Meanwhile, we also carried out landslide detection in the Longtoushan region as a case study. From the detected results of the case study (as shown in Fig. 10), it can be seen that the landslides have been effectively detected. However, due to the differences in image spectrum, resolution, and shadow between regions, the detection precision of the case study is not as outstanding as that of the open-source dataset. To improve the generalisation ability of the model, it is crucial to building a large-scale landslide dataset that contains more landslide samples from different regions.
In addition, considering that RGB images are the most commonly used and easily accessible remote sensing data when landslide events occur, we only explore the feasibility of the model to identify landslides on RGB images. In future work, additional consideration of some geological or geographical factors is more conducive to landslide detection tasks.
To address the issue of imprecise boundary detection of landslides owing to the information loss caused by convolution and pooling, this study proposed a novel UNet model (RIPF-Unet) boosted by reversed image pyramid features (RIPFs). Specifically, we designed a novel concatenation module, which incorporates the high-level and low-level pyramid features of the original image into the expansive path to compensate for the information loss.
The proposed RIPF-Unet model is trained and validated using the open-source landslide dataset of the Bijie area, Guizhou Province, China, wherein the precision of the proposed model is observed to increase by 3.5% and 4.0%, and the F1 index improved by 2.3% and 2.7%, compared to the conventional UNet and UNet + + models, respectively. This improvement can be explained by the fact that the RIPF-Unet model appropriately complements the decoding network’s information, increasing the abundance of features for decoding network extraction.
In addition, the proposed RIPF-Unet model is applied to the case of the Longtoushan region after the 2014 Ms.6.5 Ludian earthquake. Results show that the proposed model achieves a 96.63% accuracy for detecting the landslides using the Gaofen-1 images. Notably, the RIPF-Unet model is advanced in its compact parameter size and is 31% lighter compared to the Unet + + model, which indicates that the RIPF-Unet model is more suitable for mobile and portable platforms.
However, the proposed FIPF-Unet is potentially limited for landslide detection in a larger area, due to the differences in image spectrum, resolution, and shadow between regions. To improve the generalisation ability of the model, in future work, it is crucial to building a large-scale landslide dataset.
Competing interests
The authors declare that they have no competing interests.
Acknowledgements
This study was financially supported by the National Key R&D Program of China (Grant No. 2018YFD1100401); the National Natural Science Foundation of China (Grant No. 52078493); the Natural Science Foundation for Excellent Young Scholars of Hunan (Grant No. 2021JJ20057); the Innovation Provincial Program of Hunan Province (Grant No. 2020RC3002). These financial supports are gratefully acknowledged.
- Altamirano A, Miranda A, Aplin P, Carrasco J, Catalan G, Cayuela L, Fuentes-Castillo T, Hernandez A, Martínez-Harms MJ, and Peluso F (2020) Natural forests loss and tree plantations: large-scale tree cover loss differentiation in a threatened biodiversity hotspot. Environmental Research Letters 15:124055. https://doi.org/10.1088/1748-9326/abca64.
- Badrinarayanan V, Kendall A, and Cipolla R (2017) SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis & Machine Intelligence 39:2481-2495. https://doi.org/10.1109/TPAMI.2016.2644615.
- Barak S, Imamalipour A, Abedi M, Bahroudi A, and Khalifani FM (2021) Comprehensive modeling of mineral potential mapping by integration of multiset geosciences data. Geochemistry 81:125824. https://doi.org/10.1016/j.chemer.2021.125824.
- Behling R, Roessner S, Golovko D, and Kleinschmit B (2016) Derivation of long-term spatiotemporal landslide activity-A multi-sensor time series approach. Remote Sensing of Environment 186:88-104. https://doi.org/10.1016/j.rse.2016.07.017.
- Brinkhoff J, Backhouse G, Saunders ME, Bower DS, and Hunter JT (2022) Remote sensing to characterize inundation and vegetation dynamics of upland lagoons. Ecosphere 13:e3906. https://doi.org/10.1002/ecs2.3906.
- Chen L-C, Zhu Y, Papandreou G, Schroff F, and Adam H (2018) Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In: 15th European Conference on Computer Vision (ECCV), Munich, GERMANY, pp 833-851.
- Chen S, Miao Z, Wu L, and He Y (2020) Application of an incomplete landslide inventory and one class classifier to earthquake-induced landslide susceptibility mapping. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 13:1649-1660. https://doi.org/10.1109/JSTARS.2020.2985088.
- Chen X, Zhou Q, and Liu C (2015) Distribution pattern of coseismic landslides triggered by the 2014 Ludian, Yunnan, China Mw6.1 earthquake: special controlling conditions of local topography. Landslides 12:1159-1168. https://doi.org/10.1007/s10346-015-0641-y.
- Chollet F (2017) Xception: Deep Learning with Depthwise Separable Convolutions. In: 30th IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, pp 1800-1807.
- Cigna F, Tapete D, and Lee K (2018) Geological hazards in the UNESCO World Heritage sites of the UK: From the global to the local scale perspective. Earth-Science Reviews 176:166-194. https://doi.org/10.1016/j.earscirev.2017.09.016.
- Ding A, Zhang Q, Zhou X, and Dai B (2016) Automatic Recognition of Landslide Based on CNN and Texture Change Detection. In: 31st Youth Academic Annual Conference of Chinese-Association-of-Automation (YAC), Wuhan, PEOPLES R CHINA, pp 444-448.
- Eskandari R, Mahdianpari M, Mohammadimanesh F, Salehi B, Brisco B, and Homayouni S (2020) Meta-analysis of Unmanned Aerial Vehicle (UAV) Imagery for Agro-environmental Monitoring Using Machine Learning and Statistical Models. Remote Sensing 12:3511. https://doi.org/10.3390/rs12213511.
- Fan X, Scaringi G, Korup O, West AJ, van Westen CJ, Tanyas H, Hovius N, et al. (2019) Earthquake-Induced Chains of Geologic Hazards: Patterns, Mechanisms, and Impacts. Reviews of Geophysics 57:421-503. https://doi.org/10.1029/2018rg000626.
- Fernández JG, and Mehrkanoon S (2021) Broad-UNet: Multi-scale feature learning for nowcasting tasks. Neural Networks 144:419-427. https://doi.org/10.1016/j.neunet.2021.08.036.
- Galli M, Ardizzone F, Cardinali M, Guzzetti F, and Reichenbach P (2008) Comparing landslide inventory maps. Geomorphology 94:268-289. https://doi.org/10.1016/j.geomorph.2006.09.023.
- Ghorbanzadeh O, Blaschke T, Gholamnia K, Meena SR, Tiede D, and Aryal J (2019) Evaluation of Different Machine Learning Methods and Deep-Learning Convolutional Neural Networks for Landslide Detection. Remote Sensing 11:196. https://doi.org/10.3390/rs11020196.
- Gu Z, Cheng J, Fu H, Zhou K, Hao H, Zhao Y, Zhang T, Gao S, and Liu J (2019) Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 38:2281-2292. https://doi.org/10.1109/TMI.2019.2903562.
- Han Z, Li Y, Du Y, Wang W, and Chen G (2019) Noncontact detection of earthquake-induced landslides by an enhanced image binarization method incorporating with Monte-Carlo simulation. Geomatics Natural Hazards & Risk 10:219-241. https://doi.org/10.1080/19475705.2018.1520745.
- He K, Zhang X, Ren S, and Sun J (2016) Deep Residual Learning for Image Recognition. In: IEEE conference on computer vision and pattern recognition, pp 770-778.
- Howard AG, Zhu M, Chen B, Kalenichenko D, Wang W, Weyand T, Andreetto M, and Adam H (2017) MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint. https://doi.org/10.48550/arXiv.1704.04861.
- Huang G, Liu Z, Laurens V, and Weinberger KQ (2016) Densely Connected Convolutional Networks. In: IEEE conference on computer vision and pattern recognition, pp 4700-4708.
- Ji S, Yu D, Shen C, Li W, and Xu Q (2020) Landslide detection from an open satellite imagery and digital elevation model dataset using attention boosted convolutional neural networks. Landslides 17:1337-1352. https://doi.org/10.1007/s10346-020-01353-2.
- Lecun Y, Bengio Y, and Hinton G (2015) Deep learning. Nature 521:436-444. https://doi.org/10.1038/nature14539.
- Lei T, Zhang Y, Lv Z, Li S, Liu S, and Nandi AK (2019) Landslide inventory mapping from bitemporal images using deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters 16:982-986. https://doi.org/10.1109/LGRS.2018.2889307.
- Li H, He Y, Xu Q, Deng J, Li W, and Wei Y (2022) Detection and segmentation of loess landslides via satellite images: a two-phase framework. Landslides 19:673-686. https://doi.org/10.1007/s10346-021-01789-0.
- Li H, Huang J, and Ji S (2019) Bearing fault diagnosis with a feature fusion method based on an ensemble convolutional neural network and deep neural network. Sensors 19:2034. https://doi.org/10.3390/s19092034.
- Li Y, Chen G, Han Z, Zheng L, and Zhang F (2014) A hybrid automatic thresholding approach using panchromatic imagery for rapid mapping of landslides. GIScience & Remote Sensing 51:710-730. https://doi.org/10.1080/15481603.2014.972867.
- Li Z, Shi W, Lu P, Yan L, Wang Q, and Miao Z (2016) Landslide mapping from aerial photographs using change detection-based Markov random field. Remote Sensing of Environment 187:76-90. https://doi.org/10.1016/j.rse.2016.10.008.
- Liu F, and Wang L (2022) UNet-based model for crack detection integrating visual explanations. Construction and Building Materials 322:126265. https://doi.org/10.1016/j.conbuildmat.2021.126265.
- Lourenco M, Fitchett JM, and Woodborne S (2022) Angolan highlands peatlands: Extent, age and growth dynamics. Science of the Total Environment 810:152315. https://doi.org/10.1016/j.scitotenv.2021.152315.
- Meena SR, Ghorbanzadeh O, van Westen CJ, Nachappa TG, Blaschke T, Singh RP, and Sarkar R (2021) Rapid mapping of landslides in the Western Ghats (India) triggered by 2018 extreme monsoon rainfall using a deep learning approach. Landslides 18:1937-1950. https://doi.org/10.1007/s10346-020-01602-4.
- Meena SR, Soares LP, Grohmann CH, van Westen C, Bhuyan K, Singh RP, Floris M, and Catani F (2022) Landslide detection in the Himalayas using machine learning algorithms and U-Net. Landslides 19:1209-1229. https://doi.org/10.1007/s10346-022-01861-3.
- Merghadi A, Yunus AP, Dou J, Whiteley J, ThaiPham B, Bui DT, Avtar R, and Abderrahmane B (2020) Machine learning methods for landslide susceptibility studies: A comparative overview of algorithm performance. Earth-Science Reviews 207:103225. https://doi.org/10.1016/j.earscirev.2020.103225.
- Nhu V-H, Mohammadi A, Shahabi H, Ahmad BB, Al-Ansari N, Shirzadi A, Clague JJ, Jaafari A, Chen W, and Nguyen H (2020) Landslide susceptibility mapping using machine learning algorithms and remote sensing data in a tropical environment. International journal of environmental research and public health 17:4933. https://doi.org/10.3390/ijerph17144933.
- Odena A, Dumoulin V, and Olah C (2016) Deconvolution and Checkerboard Artifacts. Distill 1:e3. https://doi.org/10.23915/distill.00003.
- Petaja T, Tabakova K, Manninen A, Ezhova E, O'Connor E, Moisseev D, Sinclair VA, et al. (2022) Influence of biogenic emissions from boreal forests on aerosol-cloud interactions. Nature Geoscience 15:42-47. https://doi.org/10.1038/s41561-021-00876-0.
- Prakash N, Manconi A, and Loew S (2020) Mapping landslides on EO data: Performance of deep learning models vs. traditional machine learning models. Remote Sensing 12:346. https://doi.org/10.3390/rs12030346.
- Rajesh J, Pande CB, Kadam SA, Gorantiwar SD, and Shinde MG (2021) Exploration of groundwater potential zones using analytical hierarchical process (AHP) approach in the Godavari river basin of Maharashtra in India. Applied Water Science 11:1-11. https://doi.org/10.1007/s13201-021-01518-x.
- Ren S, He K, Girshick R, and Sun J (2017) Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. Ieee Transactions on Pattern Analysis and Machine Intelligence 39:1137-1149. https://doi.org/10.1109/tpami.2016.2577031.
- Ronneberger O, Fischer P, and Brox T (2015) U-Net: Convolutional Networks for Biomedical Image Segmentation. In: International Conference on Medical image computing and computer-assisted intervention, Springer, CHAM, pp 234-241.
- Shelhamer E, Long J, and Darrell T (2017) Fully Convolutional Networks for Semantic Segmentation. Ieee Transactions on Pattern Analysis and Machine Intelligence 39:640-651. https://doi.org/10.1109/tpami.2016.2572683.
- Shi P, Xue P, Liu A, and Han D (2021) A Novel Rotating Machinery Fault Diagnosis Method Based on Adaptive Deep Belief Network Structure and Dynamic Learning Rate Under Variable Working Conditions. IEEE Access:44569-44579. https://doi.org/10.1109/ACCESS.2021.3066594.
- Simonyan K, and Zisserman A (2014) Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint. https://doi.org/10.48550/arXiv.1409.1556.
- Su Z, Chow JK, Tan PS, Wu J, Ho YK, and Wang Y-H (2021) Deep convolutional neural network-based pixel-wise landslide inventory mapping. Landslides 18:1421-1443. https://doi.org/10.1007/s10346-020-01557-6.
- Sun W, Tian Y, Mu X, Zhai J, Gao P, and Zhao G (2017) Loess Landslide Inventory Map Based on GF-1 Satellite Imagery. Remote Sensing 9:314. https://doi.org/10.3390/rs9040314.
- Tanoli JI, Chen N, Regmi AD, and Jun L (2017) Spatial distribution analysis and susceptibility mapping of landslides triggered before and after Mw7.8 Gorkha earthquake along Upper Bhote Koshi, Nepal. Arabian Journal of Geosciences 10:1-24. https://doi.org/10.1007/s12517-017-3026-9.
- Tien Bui D, Shahabi H, Omidvar E, Shirzadi A, Geertsema M, Clague JJ, Khosravi K, Pradhan B, Pham BT, and Chapi K (2019) Shallow landslide prediction using a novel hybrid functional machine learning algorithm. Remote Sensing 11:931. https://doi.org/10.3390/rs11080931.
- Wang L, Chen C, Xie F, Hu Z, Zhang Z, Chen H, He X, and Chu Y (2021) Estimation of the value of regional ecosystem services of an archipelago using satellite remote sensing technology: A case study of Zhoushan Archipelago, China. International Journal of Applied Earth Observation and Geoinformation 105:102616. https://doi.org/10.1016/j.jag.2021.102616.
- Wang T, Borji A, Zhang L, Zhang P, and Lu H (2017) A stagewise refinement model for detecting salient objects in images. In: Proceedings of the IEEE international conference on computer vision, pp 4019-4028.
- Wiggins EB, Anderson BE, Brown MD, Campuzano-Jost P, Chen G, Crawford J, Crosbie EC, et al. (2021) Reconciling Assumptions in Bottom-Up and Top-Down Approaches for Estimating Aerosol Emission Rates From Wildland Fires Using Observations From FIREX-AQ. Journal of Geophysical Research: Atmospheres 126:e2021JD035692. https://doi.org/10.1029/2021jd035692.
- Xu C, Xu X, Shyu JBH, Gao M, Tan X, Ran Y, and Zheng W (2015) Landslides triggered by the 20 April 2013 Lushan, China, Mw 6.6 earthquake from field investigations and preliminary analyses. Landslides 12:365-385. https://doi.org/10.1007/s10346-014-0546-1.
- Yu B, Chen F, Xu C, Wang L, and Wang N (2021) Matrix SegNet: a practical deep learning framework for landslide mapping from images of different areas with different spatial resolutions. Remote Sensing 13:3158. https://doi.org/10.3390/rs13163158.
- Yu H, Ma Y, Wang L, Zhai Y, and Wang X (2017) A Landslide Intelligent Detection Method based on CNN and RSG_R. In: IEEE International Conference on Mechatronics and Automation (ICMA), Takamatsu, JAPAN, pp 40-44.
- Yu W, Cui J, Gao Y, Zhu M, Shao L, Shen Y, Zhang X, Guo C, and Zhang H (2022) Evolution of Desertification Types on the North Shore of Qinghai Lake. Cmc-Computers Materials & Continua 71:3635-3646. https://doi.org/10.32604/cmc.2022.023195.