Detection of Sinkholes and Landslides in a Semi-Arid Environment Using Deep-Learning Methods, UAV images, and Topographical Derivatives

doi:10.21203/rs.3.rs-2847897/v1

Download PDF

Research Article

Detection of Sinkholes and Landslides in a Semi-Arid Environment Using Deep-Learning Methods, UAV images, and Topographical Derivatives

https://doi.org/10.21203/rs.3.rs-2847897/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Sinkholes and landslides occur when parts of a soil collapse mainly in more gentle or steeper slopes respectively, both often triggered by intensive rainfall. These processes often cause problems in the hilly regions in the “Golestan province” of Iran, and their detection is the essential aim for this research. The production of soil landforms maps is typically based on visual interpretation of aerial and satellite images eventually supported by field surveys. Recent advances in the acquisition of images from “unmanned aerial vehicles (UAV)” and of “deep learning (DL)” methods inherited from computer vision have made it feasible to propose semi-automated soil landforms detection methodologies for large areas at an unprecedented spatial resolution. In this study, we evaluate the potential of two cutting-edge DL segmentation models, the vanilla “U-Net model” and the “Attention Deep Supervision Multi-Scale U-Net” model, applied to “UAV”-derived products, to map landslides and sinkholes in a semi-arid environment, the “Golestan Province” (north-east Iran) Results show that our framework can successfully map landslides in a challenging environment (with an F1-score of 69%), and that topographical derivates from “UAV-derived DSM” decrease the capacity of mapping sinkholes of the models calibrated with optical data.

Unmanned aerial vehicle

Sinkholes

Landslide detection

Deep learning

Landslides and sinkholes are very common rainfall induced processes in Iran. Landslides cause widespread damage to social and economic infrastructure, as well as to natural features (Sarro et al., 2018). They have indirect and direct influence on major infrastructure, including human habitations, critically impacting land use change and rural to urban migration (Miura, 2019; Thakur et al., 2023). Landslide analysis commonly aims both to detect the failures, and to present the areas affected to generate inventories. These inventories contain information, including the spatial patterns of the landslides/sinkholes, the volume of the failed slope, and the type/data of failure (Mezaal et al., 2018; Wijaya et al., 2023).

Sinkholes Sinkholes develop and expand in a wide range of pedo/eco-climatic landscapes, throughout the world (Poesen, 2018, Bernatek-Jakiel and Poesen, 2018). They have effective impacts on landscape hydrology and have a main role in landscape evolution (Jones and Crane, 1984) including degradation. Meanwhile, this kind of erosion, as a sediment source fingerprint, is related to drainage networks (Bernatek-Jakiel, 2015) and it leads to channel extension (Higgins et al., 1990). It also intensifies annual gully head migration five times more than gully head retreats alone (Crouch, 1983; Daley et al., 2023).

To perform landslide and sinkhole detection, it is important to examine different accessible inventory of preceding hazards, their areal extent, and their map to identify other landslide and sinkhole-prone areas (Agrawal et al., 2017; Chen et al., 2018; Chen et al., 2021). The outputs are extremely reliant on the accuracy of the input data (Ghorbanzadeh et al., 2018).

The advent of advanced remote sensing technology has helped pave an efficient way to map surface changes, particularly changes in the landscapes accruing to extreme climatological events, anthropogenic factors, and ground shaking. Remote sensing is very beneficial as it allows for large-scale mapping of slope failures (Fernández et al., 2016). Remote sensing provides mostly updated data with relatively low resolution. However, “unmanned aerial vehicles (UAVs)” have become a very suitable tool for preparing low-cost, up-to-date, and precise field datasets (Ghorbanzadeh et al., 2019; Zhang et al., 2019). The use of “UAVs” in natural events and urgent situations is developing (Watson et al., 2019; Windrim et al., 2019), and compared to conventional tools, it has the potential to prepare data with better resolution (Brovkina et al., 2018). Recently “UAV” was applied in multi-hazard susceptibility mapping in the same hilly loess region of the current study by Kariminejad et al. (2022) where they reported sinkholes were the most perilous natural hazard, followed by landslides, and headcuts. UAVs have been used on several papers to collect various data for landslide inventories, and piping/headcut monitoring (Lin et al., 2010; Yang et al., 2015; Fernández et al., 2016), but analysis/classification of UAV remote sensing imagery to detect sinkholes and landslides are a less studied existing technique in the present research.

The approaches applied for image analysis/classification are mainly based of pixel and object which is extended for natural hazard detection applying various machine- and deep-learning algorithms (Hölbling et al., 2012; Dou et al., 2015; Mezaal et al., 2018). Recently, deep learning models have been applied for segmentation and classification of “remotely sensed imagery” (Du et al., 2019; Qayyum et al., 2019; Ghadi et al., 2022). In literature, DL-based automated techniques for sinkhole mapping are currently uncommon. Lee et al. (2016) employed light CNNs on thermal far-infrared (FIR) imagery to map sinkholes in the context of urban areas. Hoai et al. (2019) presented a sinkhole tracking approach that takes use of recent developments in CNN transfer learning on FIR imagery, however, it is still inconclusive if such methods would work in very complex semi-arid regions and using conventional (and usually more available) RGB optical imagery. As for landslides, there are few studies that employ VHR UAV photography in conjunction with DL algorithms for automated mapping. For example, Ghorbanzadeh et al. (2019) conducted an intriguing work in which they trained different CNN architectures to map landslides in a densely forested area of the Himalayas. Karantanellis et al. (2021) compared multiple ML models with “object-based image analysis” to determine the boundaries of two rotating slides in Greece. However, this research was performed in the context of only two landslide bodies.

The literature indicates that there are few studies that examine the use of UAV with DL to automatically map landslides and sinkholes, while none of them is tested in a semi-arid environment. The use of Deep learning framework is an innovative method for landslide and sinkhole detection. deep learning framework rather than UAV. Also, similar applications in semi-arid areas are scarce due to extraction difficulties from a little spectrum difference. By applying the “spectral information from the UAV-derived imagery”, in conjunction with the slope dataset, we have illustrated the efficiency of DL-based automated approaches and the benefits and shortcomings of applying topographic dataset, especially slope dataset, for sinkhole and landslide detection. We contrasted the output maps prepared by applying automated approaches with manually extracted sinkhole and landslide inventory datasets created from a variety of sources. The detected sinkholes and landslides were then validated applying “conventional computer vision validation techniques”.

Landslides are often hazardous, especially those that happen along a road (Sarro et al., 2018; May et al., 2023) and sinkholes that are critically important as a geo-hazard in loess regions are in positive relation with roads and drainage networks (Hosseinalizadeh et al., 2018). The point is that landslide- and sinkhole-susceptible zones in a terrain where emergency responses are not easily approachable, require attention and that's where landslide- and sinkhole-detecting maps are very important for stakeholders to make decisions for first response time. Moreover, there have been a few investigations into various types of landslide susceptibility mapping, and early warning strategies are exceedingly being used to circumvent damage to infrastructures (Meena et al., 2019).

The primary motivation for conducting the study area was the quantitative and qualitative miracle nature of sinkholes and associated problems, such as the i limited traffic of agricultural machinery, cattle being trapped inside, and the extraordinary hydrological response of loess landscapes. Sometimes the depth of these sinkholes and following gully heads reaches more than ten meters and this causes the death of cattles. In addition, due to numerical and dimensional increasing of sinkholes followed by gully head retreat and development and sever land degradation, residential immigration toward Kalaleh city and nearby towns increased. Because of the high severity of erosion including sinkholes, landslides, and gullies in a short period of time caused by climatic and anthropogenic factors, croplands are divided into several parts and exposed to infertile, highly erodible, and degraded lands. Due to torrential rainfall in nearby areas the number of sinkholes and gully-heads increased in two years (2018–2019) (Kariminejad et al., 2021) that shows the high importance of landscape changes and evolution. In the study area for 2020 and 2021 susceptible regions to hazards (sinkhole, landslide, gully-head) and compound events increases. Apart from a variety of on-site or site-specific effects of landslides, sinkholes, and their derivatives (diverse types of gullies) on landscape evolution/land degradation, their off-site effects are of foremost importance that is beyond the scope of research and this region overlooks the main road of “Kalaleh-Marveh Tappeh” (road destruction due to erosion). Most percentage of interested areas are rangeland that in earlier study in loess regions (Hosseinalizadeh et al., 2018) have been proved as susceptible land use to sinkhole erosion. This land use experiences overgrazing and intense animal trampling for the whole of the year. Landslides have also endangered the residents of the study area and their properties. Since these geohazards are rainfall-induced, they increase in size and severity after erosive rainfall (Kariminejad et al., 2022). Therefore, these issues, apart from identifying the factors influencing the formation of these land degradation processes, led us to restate the problems and their impacts to finally find suitable solutions for their management.

Moreover, since this kind of soil erosion is the main origin of some major soil erosion including gully initiation and extension, applying new technology namely, “UAV and deep learning” is highly important and recommended. Apart from that, excessive grazing as an important man-made factor should be managed in the loess driven soils since these landscapes are fertile and fragile deposits and hence, very susceptible to natural hazards including sinkholes and landslides.

Due to erosive rainfall in the same region, the number and volume of sinkholes and landslides increased that shows the importance of soil erosion and highly erodible conditions of loess landscapes (Kariminejad et al., 2022). Based on field surveys, sinkholes precede and are formed earlier than drainage networks. Based on the geomorphological approach, in order to express the spatial interaction between sinkholes and landslides, it can be stated that in stable environmental conditions, sinkholes are present in a scattered pattern with a low number and density. With the influence of environmental factors, the spatial pattern of sinkholes was clustered, and their density increased. Although, a factor such as slope alone cannot lead to the initiation and expansion of landforms (such as gully heads, landslides, and sinkholes erosion) in fact, the formation of these landforms is influenced by very complex and unpredicted internal and external processes of the earth, which, if not managed in time, can increase the propensity to natural hazards and disasters.

As stated, most areas of the studied region are rangeland and some old sinkholes are encompassed by woody rangeland (covered with trees namely pomegranate, Ailanthus, Paliurus spina, Lycium shawii, Reed, and Fig). Therefore, sinkholes have strange dual behavior including 1) land degradation or landscape change/evolution, and 2) natural flowerpots (natural holes with suitable conditions for self-growing trees). The first and second behavior happens in young and old sinkholes in turn. Hence, in this case biodiversity of sinkholes is highly important and can be considered in soil and water conservation practices. So, their identification, monitoring, managing, and bioengineering are critically important. In earlier research of the Iranian Loess Plateau (ILP) (Kariminejad et al., 2022), it was mentioned that sinkholes and gully heads are highly concentrated in rangeland, and they are managed and controlled in cropland by landowners and agricultural practices.

It is in the “Golestan Province” (north-east Iran) between (55° 39′ 30˚ E to 55° 41′ 30˚ E) longitudes, and (37″ 36′ 30˚ N to 37″ 38′ 30˚ N) latitudes, with 515 hectares of mostly semi-arid climatic. The region is in the “loess-driven soils” with the average annual rainfall of 261 mm (IMA, 2022). The minimum and maximum altitudes are 211 and 549 meters above sea level (Fig. 1). The “silt loam” is the dominant texture of the soil surface and the whole area covered by loess. The geographic map of the region and some picture examples of “sinkholes and landslides” are shown in Fig. 1 and Fig. 2, respectively. Moreover, this study area is one of the places that is very important for ranchers and farmers. For instance, many shepherds who go to graze with their sheep, when they return, they see that some of the sheep are missing. it is because of the impact of soil piping erosion on grazing practices. We have taken a drone video (Video 1) that show a sheep has fallen into a sinkhole, and then we took hold of it and lift it out.

3.1 Dataset creation

The “UAV-based photogrammetric” capture yielded accurate landslide and sinkhole location identification. In the investigation, landslides and sinkholes were found in areas with optical and morphological erosion and deposition evidence. A stationary “UAV” (“Sensefly eBee x”) with the camera type “Sensefly Aeria X” and a focal length of 18.50 mm was used. Acquisitions were scheduled to ensure a “Ground Sampling Distance (GSD)” of 5.0 cm on average. The orthomosaic's resulting pixel size is 5.0 cm/pixel. An experienced UAV operator planned the flight using the “senseFly eMotion” flight control program. A “GPS RTK” adjustment has also been included. The UAV trajectories were constructed with a 70% lateral overlap and an 85% fore-and-aft overlap. The flight's mean altitude was 220 meters above the ground. The Pix4Dmapper photogrammetry program was used to process aerial pictures captured during the study.

For each composite image (see Table 1) and the corresponding mask, we extracted a large tile from the whole research region to construct the corresponding dataset. The tile's dimensions were 512x512 pixels, and 3, 4, or 7 bands, depending on the band arrangement. In the case of the landslide’s dataset, the number of pixels in the studied region reveals that there are 2064288 pixels inside landslides and 487868622 pixels in the background class, for a ratio of 236. The tile and mask were separated into 512 x 512-pixel patches without overlapping, resulting, in this case, in 6508 patches. As for the sinkhole’s dataset, instead, we find 2096241 pixels belonging to the sinkholes class and 1864415983 pixels to the background class, for a ratio of 899. In both cases, we only saved images that had at least one pixel from the researched target (132 for landslides and 203 for sinkholes). This sampling approach was chosen to reduce the disparity between classes while also supplying a constant quantity of target pixels. We used 80% of all patches for training and 20% for testing (Agrawal et al., 2017). A further validation dataset is created using 20% of the training dataset. The patches were randomly shuffled before being divided into three datasets including, test, validation, and training. In both cases, “patches” were sampled without overlap, to avoid biases in the calibration process. We finally standardized the highest and lowest pixel values between 0 and 1 for each patch.

Table 1

Landslide and sinkhole evaluation with three different band combinations, namely, RGB, DHSC, and the combination of optical and topographical derivates (ALL).
Target	Dataset	Band1	Band2	Band3	Band4	Band5	Band6	Band7
Landslides	RGBS	Red	Green	Blue	Slope	x	x	x
	RGB	Red	Green	Blue	x	x	x	x
Sinkholes	DSHC	DSM	Slope	Hillshade	P Curve	x	x	x
	ALL	Red	Green	Blue	DSM	Slope	Hillshade	Profile Curve

Four different band combinations are investigated. For the landslide mapping task, the RGBS combination was used, since it is evident from the literature that Earth Observation (EO) imagery with the aid of slope, in this case, is the most suitable band combination (Ghorbanzadeh et al., 2019). As for the sinkhole detection, three different band combinations were evaluated, namely, RGB (optical), DHSC (topographical derivatives), and the combination of optical and topographical derivatives (ALL) (Table 1).

3.2 Models architecture and training strategy

3.2.1 U-Net

Numerous semantic segmentation applications have used “U-Net”, with excellent results (Abderrahim et al., 2020; Chandra et al., 2023). U-Net was first used for biomedical image segmentation (Ronneberger et al., 2015) and has been used successfully in several landslide detection tasks, proving to be a good solution when dealing with Earth Observation image segmentation tasks (Bhuyan et al., 2022; Meena et al., 2022; Nava et al., 2022). A contracting path (encoder) records low-level representation, whereas a decoding path record high-level representations. The “encoding path” is like a conventional “CNN” structure and consists of succeeding “convolution blocks”. Every convolutional block has two convolutional layers with a 3 × 3 kernel size and a 2 x 2 max-pooling layer. Each layer of the convolutional network is activated by the “rectified linear unit (ReLU) activation function” (Agarap, 2018). To perform non-linear downsampling, a 2 x 2 max-pooling layer is placed at the end of the “convolutional block” in the encoder path. In contrast, 2 x 2 upsampling layers are added inside the decoder path and are followed by a 3x3 convolutional layer which is the upsampling layer (see Fig. 3). This combination is referred to as learnable upconvolution (Kundu et al., 2020).

3.2.2 “Attention deep supervision multi-scale (ADSMS) U-Net”

The traditional “U-Net” structure has significant drawbacks, notably when working with imbalanced data and tiny targets like we do in present study. We use the “ADSMS U-Net” architecture established by Abraham and Khan (2019) for skin lesion segmentation to overcome the constraint (Fig. 4). The model may gather class information that is more readily available at many scales thanks to multi-scale inputs. This holds for both background features and targets. Furthermore, because of cascading convolutions, obtaining inaccurate predictions for tiny objects with high form variability is simple. Soft attention gates (AGs), which assist in identifying pertinent “spatial information” from low level feature maps (where critical information is present but the original image is significantly distorted), are therefore utilized to minimize the problem. Lastly, deep supervision conducts a potent "regularization” when training data is few, and networks are shallow.

3.2.3 Model training

When training “U-Net”, “binary cross-entropy” was applied as a loss function. As an exception, the final output of the “ADSMS U-Net” model is directed by the conventional Tversky Loss. At the same time, each high-dimensional feature representation is governed by “Focal Tversky Loss” to prevent loss over-suppression. The intermediate layers in this “deep supervision technique” must be semantically discriminative at various scales, according to Lee et al. (2015). Additionally, it helps to make sure that the attention module has the flexibility to alter how it responds to a range of visual foreground stimuli. Abraham and Mefraz Khan, (2018) is who put out the ADSMS U-Net architecture along with it. For both models, we employed a stochastic gradient descent method (Adam), which is effective in issues involving ambiguous data and sparse gradients. Adam is according to an adaptive estimate of “first order and second-order moments” (Kingma and Lei Ba, 2014).

In this section, we present and discuss the results of the “ADSMS U-Net” architecture while comparing it with a control model, that is the conventional “U-Net”, to identify landslides and sinkholes. The results of applying the control “U-Net” and “ADSMS U-Net” models to identify landslides and sinkholes on “UAV imagery” are shown. The findings of segmentation are shown for both the landslides (Table 2) and sinkholes (Table 3). To achieve the best results, hyperparameters are adjusted for both models.

Table 2

Landslide mapping: evaluation metrics for the U-Net and ADSMS U-Net models on the unseen test set on the RGBS dataset.
Model	Dataset	Batch Size	Learning Rate	Filters	Precision	Recall	F1-Score	IOU Score
U-Net	RGBS	8	0.0005	8	0.7225	0.6000	0.6444	0.4778
ADSMS U-Net	RGBS	4	0.0001	8	0.6488	0.7631	0.6896	0.5281

Table 3

Sinkhole mapping: evaluation metrics for the U-Net and ADSMS U-Net models on the unseen test set for the RGB, DSHC, and ALL datasets.
Model	Dataset	Batch Size	Learning Rate	Filters	Precision	Recall	F1-Score	IOU Score
U-Net	RGB	8	0.0005	4	0.4646	0.4613	0.4624	0.3018
	DSHC	8	0.0001	8	0.1422	0.5137	0.2143	0.1204
	ALL	4	0.0005	4	0.5398	0.3707	0.4394	0.2829
ADSMS U-Net	RGB	4	0.0001	16	0.4173	0.6403	0.5045	0.3380
	DSHC	4	0.0001	32	0.3937	0.3410	0.3643	0.2246
	ALL	4	0.0001	32	0.5059	0.5382	0.5215	0.3527

4.1 Landslides

While running the two models, the two had quite opposite tendencies in terms of their prediction capabilities. The performance of the U-Net model depicts that the model had fewer false positives but at the same time, missed a lot of “landslide pixels” which is explained by the lower Recall score. At the same time, the “ADSMS U-Net” model performed better at not missing out on landslide pixels but attributed a lot to false predictions (explained by the lower Precision score). Overall, the trade-off between Precision and Recall, i.e., the F1 score seems very well balanced. The best F1-score of 0.68 was obtained with the ‘ADSMS U-Net” model with a significantly higher recall (Table 2).

Figure 5 shows the power of the two models, particularly that of the “ADSMS U-Net” model where the predictions are much closer to the “ground truth” than the ones of “U-Net”. This model identifies most of the landslides present in the test patches. In certain cases, landslide footprints were detected even more accurately than the manual mapping used as ground truth, while detecting the footprints where the landslides were not mapped (Fig. 5). It is feasible to observe that the U-Net produces misleading positive landslide predictions when none exist. Using the ADSMS U-Net eliminates this tendency for misclassification and allows for a more accurate delineation of the target landslide polygon(s).

4.2 Sinkholes

For all band combinations, the performances of “ADSMS U-Net” are superior to the ones of the “conventional U-Net” architecture. The best overall scores are yielded by the “ADSMS U-Net” when trained on the ALL dataset. We can notice, as in the previous case, the higher recall of the “ADSMS U-Net”, with respect to the conventional architecture. As for the effectiveness of the different composites evaluated in this paper, we can notice contradictory behavior of the models. The conventional “U-Net” achieves the best performance when using just the optical RGB combination, while “ADSMS U-Net” can exploit the information of both topographical derivatives and optical data, achieving the best results with the ALL composite. Moreover, it is noticeable that the DSHC dataset alone achieves the worst results for both models (Table 3).

By a visual evaluation of the results in Fig. 6, we can notice a tendency of the models to underpredict the results when predicting using just topographical derivatives. On the other hand, when using the RGB and ALL datasets, we see that almost all objects are predicted correctly. However, we can notice some imprecise predictions of the sinkhole boundaries’ delineation. Overall, training with the RGB data tends to outperform the data curated simply of the “UAV-derived topographic” data. However, as seen in the true color images, quite often trees are present in the sinkholes and therefore contribute to errors, while using the DSMs, in detecting the sinkholes. In fact, the measurements of the depth of the sinkholes are hampered by vegetation and the resulting DSM often does not represent the actual sinkhole structure.

Now, the automated technique is the only viable alternative for mapping broad areas with enough “spatial and temporal” accuracy (in a short time span) for scientific and operational reasons. However, effective, repeatable, and accurate processes for automating landslide detection on VHR “UAV” pictures are currently absent. Automated “pipelines” that strive to address these issues can therefore overcome these challenges, considerably reducing the requirement for human interventions, and thus being viable for reliable real-time monitoring and mapping of “natural hazards”. Coupling the two DL models with UAV imagery made the automated mapping beneficial.

The RGBs combination is confirmed to be a valid solution to automatically map landslides with DL models. In this case, reliable predictions are yielded by both “U-Net” and “ADSMS U-Net”. However, it is possible to notice that the U-Net yields false positive classification of landslides where no real ones are present. This misclassification tendency is not found when using the “ADSMS U-Net” and this promotes a good predictive capability of the latter model. This behavior is due to the multiscale inputs as well as the Soft Attention Gates as described above. As for sinkholes, it is evident how some sinkholes are more easily detectable with RGB data while others are more clearly mapped using topographical derivatives. This difference can be due to the presence of trees in most of the collapsed pipes, which are captured by both the optical data and DSM. When using the ALL dataset, however, the optical and topographic information combined allowed “ADSMS U-Net” to reach the highest mapping performances.

If we analyze the “U-Net” performances on both DSHC and ALL datasets it is clear how such a model is not able to exploit successfully the information carried by the topographical layers. On the other hand, the ability of the “ADSMS U-Net” to get higher accuracies can be attributed to the fact that the model uses a soft attention mechanism that enables retention of the low-level information during “convolution”, and at the same time, “multi-scale” inputs allow the model to gather that class information that is more readily available at many scales. We feel that this framework can be used to its full capacity, particularly in circumstances when the data resolution is quite high as in such cases.

DL-based automated approaches for sinkhole mapping are still scarce in the literature (Hoai et al., 2019). However, few studies in which VHR UAV imagery is used in combination with DL algorithms to map landslides are present. Ghorbanzadeh et al. (2019) trained several CNN architectures to map landslides in a highly vegetated area in the Himalayas (F1-score = 0.85). However, in our study area, the semi-arid environment makes it even more challenging as it is difficult to distinguish between the abundant bare soil and the targets (landslides and sinkholes). Also, Karantanellis et al. (2021) evaluated several ML models combined “with object-based image analysis” to get the boundaries of two rotational slides in Greece (F1-score = 0.85). However, in this case, just two landslides are included in the test set. This might not be plausible when upscaling predictions to much larger areas.

This study investigated rapid sinkhole and landslide detection approaches in a semi-arid environment, and we assess, analyze, and discuss the suggested architecture's prediction performance compared to the conventional “U-Net”. Moreover, we investigate the use of “UAV-derived DSM” to detect sinkholes, from the standpoint of two cutting-edge models. We purposefully picked a difficult location for mapping to develop an adaptive approach that is reliable under harsh detecting settings. On the other hand, the technique is easily scalable to less difficult conditions, such as densely vegetated regions. This study demonstrates one of the first “CNN”-based mappings of sinkholes and landslides in a difficult semi-arid region. Furthermore, to the best of the experts' knowledge, it is the first study where the “ADSMS U-Net” architecture is employed in conjunction with UAV data. Our results revealed the combination of DL and UAV-derived data as a reliable approach and instrument for mapping landslides also in such environments. The findings show that topographical derivatives from UAV-derived DSM are not a solid solution to map sinkholes. Further research will investigate the use of such an approach to map landslides and sinkholes over time, to assess the temporal evolution of such natural hazards.

Acknowledgments

This study was supported by “Iran National Science Foundation (INSF)” (Grant No. 99011991). We wholeheartedly thanked the International Association of Geomorphology (IAG) for bringing us as young geomorphologists together to do valuable scientific research and projects. We are grateful to Lorenzo Nava and his scientific team for their constructive remarks on methods that significantly improved the quality of the manuscript.

Compliance with Ethical Standards

The authors declare that they have no conflict of interest.

Agarap AF (2018) Deep learning using rectified linear units (RELU). arXiv preprint arXiv:1803.08375
Abraham N, Khan NM (2019) A novel focal Tversky loss function with improved attention u-net for lesion segmentation. IEEE 16th international symposium on biomedical imaging (ISBI 2019) (pp. 683–687). IEEE
Agrawal K, Baweja Y, Dwivedi D, Saha R, Prasad P, Agrawal S, Kapoor S, Chaturvedi P, Mali N, Kala VU, Dutt V (2017) December. A comparison of class imbalance techniques for real-world landslide predictions. International Conference on Machine Learning and Data Science (MLDS) (pp. 1–8). IEEE
Qayyum A, Malik A, Saad M, N. and, Mazher M (2019) Designing deep CNN models based on sparse coding for aerial imagery: a deep-features reduction approach. Eur J Remote Sens 52(1):221–239
Bernatek A (2015) The influence of piping on mid-mountain relief: A case study from the polish bieszczady Mts. (Eastern Carpathians). Carpathian J Earth Environ Sci 10(1):107–120
Bernatek-Jakiel A, Poesen J (2018) Subsurface erosion by soil piping: significance and research needs. Earth Sci Rev 185:1107–1128
Brovkina O, Cienciala E, Surový P, Janata P (2018) Unmanned aerial vehicles (UAV) for assessment of qualitative classification of Norway spruce in temperate forest stands. GeoSpatial Inform Sci 21:12–20
Chandra N, Sawant S, Vaidya H (2023) An Efficient U-Net Model for Improved Landslide Detection from Satellite Images. PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science, pp.1–16
Chen W, Shahabi H, Zhang S, Khosravi K, Shirzadi A, Chapi K, Pham BT, Zhang T, Zhang L, Chai H et al (2018) Landslide susceptibility modeling based on GIS and novel bagging-based kernel logistic regression. Appl Sci 8:2540
Chen Y, Chen W, Janizadeh S, Bhunia GS, Bera A, Pham QB, Linh NTT, Balogun AL, Wang X (2021) Deep learning and boosting framework for piping erosion susceptibility modeling: spatial evaluation of agricultural areas in the semi-arid region.Geocarto International, pp.1–27
Crouch RJ (1983) The role of tunnel erosion in gully head progression. J Soil Conserv New South Wales 39(2):148–155
Daley JS, Spencer JR, Brooks AP, Stout JC, Thwaites R (2023) Direct rain splash and downwearing of internal surfaces as an important erosion process in alluvial gully development. CATENA 221:106760
Dou J, Chang K-T, Chen S, Yunus A, Liu J-K, Xia H, Zhu Z (2015) Automatic case-based reasoning approach for landslide detection: Integration of object-oriented image analysis and a genetic algorithm. Remote Sens 7:4318
Du Z, Yang J, Ou C, Zhang T (2019) Smallholder crop area mapped with a semantic segmentation deep learning method. Remote Sens 11:888
Fernández T, Pérez JL, Cardenal J, Gómez JM, Colomo C, Delgado J (2016) Analysis of landslide evolution affecting olive groves using UAV and photogrammetric techniques. Remote Sens 8:837
Ghadi YY, Rafique AA, Al Shloul T, Alsuhibany SA, Jalal A, Park J (2022) Robust object categorization and Scene classification over remote sensing images via features fusion and fully convolutional network. Remote Sensing, 14(7), p.1550
Ghorbanzadeh O, Rostamzadeh H, Blaschke T, Gholaminia K, Aryal J (2018) A new GIS-based data mining technique using an adaptive neuro-fuzzy inference system (ANFIS) and k-fold cross-validation approach for land subsidence susceptibility mapping. Nat Hazards 94:497–517
Ghorbanzadeh O, Meena SR, Blaschke T, Aryal J (2019) UAV-based landslide detection using deep-learning convolutional neural networks. Remote Sensing, 11(17), 2046
Higgins CG, Coates DR, and others (1990) Groundwater geomorphology: The role of subsurface water in Earth-surface processes and landforms, vol 252. Geological Society of America
Hoai NV, Dung NM, Ro S (2019) July. Sinkhole detection by deep learning and data association. Eleventh International Conference on Ubiquitous and Future Networks (ICUFN) (pp. 211–213). IEEE
Hölbling D, Füreder P, Antolini F, Cigna F, Casagli N, Lang S (2012) A semi-automated object-based approach for landslide detection validated by persistent scatterer interferometry measures and landslide inventories. Remote Sens 4:1310–1336
Hosseinalizadeh M, Kariminejad N, Alinejad M (2018) An application of different summary statistics for modelling piping collapses and gully headcuts to evaluate their geomorphological interactions in Golestan Province, Iran, vol 171. Catena, pp 613–621
Jones JAA, Crane FG (1984) Pipe flow and pipe erosion in the Maesnant experimental catchment. International Geographical Union Commission on Field Experiments in Geomorphology, pp 55–72
Karantanellis E, Marinos V, Vassilakis E, Hölbling D (2021) Evaluation of machine learning algorithms for object-based mapping of landslide zones using UAV data. Geosciences 11(8):305
Kariminejad N, Hosseinalizadeh M, Pourghasemi HR, Tiefenbacher JP (2021) Change detection in piping, gully head forms, and mechanisms. CATENA 206:105550
Kariminejad N, Pourghasemi HR, Hosseinalizadeh M (2022) Analytical techniques for mapping multi – hazard with geo – environmental modeling approaches and UAV images.Scientific Reports,1–17
Kingma DP, Ba J (2014) Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
Kundu S, Mostafa H, Sridhar SN, Sundaresan S (2020) Attention-based Image Upsampling. arXiv preprint arXiv:2012.09904
Lee CY, Xie S, Gallagher P, Zhang Z, Tu Z (2015) Deeply-supervised nets. Artificial intelligence and statistics. PMLR, pp 562–570
Lin J, Tao H, Wang Y, Huang Z (2010) Practical application of unmanned aerial vehicles for mountain hazards survey. In Proceedings of the International Conference on Geoinformatics, Beijing, China
Meena SR, Mishra BK, Tavakkoli Piralilou S (2019) A hybrid spatial multi-criteria evaluation method for mapping landslide susceptible areas in Kullu valley. Himalayas Geosci 9:156
Meena SR, Soares LP, Grohmann CH, van Westen C, Bhuyan K, Singh RP, Catani F (2022) Landslide detection in the Himalayas using machine learning algorithms and U-Net. Landslides 19(5):1209–1229
Mey J, Guntu RK, Plakias A, Silva de Almeida I, Schwanghart W (2023) More than one landslide per road kilometer–surveying and modelling mass movements along the Rishikesh-Joshimath (NH-7) highway, Uttarakhand, India. Natural Hazards and Earth System Sciences Discussions, pp.1–25
Mezaal M, Pradhan B, Rizeei H (2018) Improving landslide detection from airborne laser scanning data using optimized Dempster–Shafer. Remote Sens 10:1029
Miura H (2019) Fusion analysis of optical satellite images and digital elevation model for quantifying volume in debris flow disaster. Remote Sensing, 11(9), p.1096
Abderrahim NYQ, Abderrahim S, Rida A (2020) “Road Segmentation using U-Net architecture,” in 2020 IEEE International conference of Moroccan Geomatics (Morgeo), 1–4
Nava L, Bhuyan K, Meena SR, Monserrat O, Catani F (2022) Rapid Mapping of Landslides on SAR Data by Attention U-Net. Remote Sens 14(6):1449
Poesen J (2018) Soil erosion in the Anthropocene: Research needs.Earth Surface Processes and Landforms, 43(1)
Ronneberger O, Fischer P, Brox T (2015) U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234–241). Springer, Cham
Sarro R, Riquelme A, García-Davalillo JC, Mateos RM, Tomás R, Pastor JL, Cano M, Herrera G (2018) Rockfall simulation based on UAV photogrammetry data obtained during an emergency declaration: Application at a cultural heritage site. Remote Sensing, 10, 1923
Thakur M, Kumar N, Dhiman RK, Malik JN (2023) Geological and geotechnical investigations of the Sataun landslide along the Active Sirmauri Tal Fault, Sataun, Northwestern Himalaya, India. Landslides, pp.1–19
Yang Z-h, Lan H-x, Gao X, Li L-p, Meng Y-s (2015) Wu, Y.-m. Urgent landslide susceptibility assessment in the 2013 Lushan earthquake-impacted area, Sichuan province, china. Nat Hazards 75:2467–2487
Watson CS, Kargel JS, Tiruwa B (2019) Uav-derived Himalayan topography: Hazard assessments and comparison with global dem products.Drones3,18
Wang L, Li XA, Zheng ZY, Zheng H, Ren YB, Chen WJ, Lei HN (2022) Analysis of the landslide mechanism a under tunnel erosion environment in the south-eastern Loess Plateau in China. Catena, 212, p.106039
Wijaya IP, Straka W, Mergili M, Ottner F, Wriessnig K, Arndt R, Andreatta P, Arifianti Y, Zangerl C (2023) Geological characterization and failure analysis of a catastrophic landslide in volcaniclastic soils: the Banjarnegara–Jemblung Landslide (Indonesia).Quarterly Journal of Engineering Geology and Hydrogeology, 56(1)
Windrim L, Bryson M, McLean, Randle M, J (2019) Stone, C. Automated mapping of woody debris over harvested forest plantations using UAVs, high-resolution imagery, and machine learning. Remote Sens 11:733
Zhang Y, Yue P, Zhang G, Guan T, Lv M, Zhong D (2019) Augmented reality mapping of rock mass discontinuities and rockfall susceptibility based on unmanned aerial vehicle photogrammetry. Remote Sens 11:1311

Vidio1.mp4

Download PDF

Version 1

posted

You are reading this latest preprint version

Detection of Sinkholes and Landslides in a Semi-Arid Environment Using Deep-Learning Methods, UAV images, and Topographical Derivatives

Status:

Version 1

Abstract

Figures

1. Introduction

2. Rationale

3. Study area and Materials

3. Methodology

3.1 Dataset creation

3.2 Models architecture and training strategy

3.2.1 U-Net

3.2.2 “Attention deep supervision multi-scale (ADSMS) U-Net”

3.2.3 Model training

4. Results

4.1 Landslides

4.2 Sinkholes

5. Discussion

6. Conclusion

Declarations

References

Supplementary Files

Status:

Version 1