2.1.  Remote Sensing for Landslide Analysis

A landslide is a substantial displacement of earth, rock, or debris down an inclined surface. It is a form of “mass wasting,” which is caused by the downward movement of soil and rock due to the direct influence of gravity. Typically, landslides have multiple contributing factors. They can occur due to unstable slopes or reduced strength in the underlying materials. Various factors can initiate landslides in slopes already predisposed to movement, including rainfall, snowmelt, changes in water levels, stream erosion, fluctuations in groundwater levels, earthquakes, volcanic activity, human-induced disturbances, or a combination of these elements. Seismic activity and other external forces can also trigger landslides underwater, referred to as submarine landslides. These submarine landslides have the potential to generate tsunamis, which can cause significant damage to coastal regions.

In the context of geological phenomena, landslides are frequently initiated by external factors such as precipitation or seismic activity. Furthermore, the occurrence of landslides is influenced by various proxies, including soil moisture content, topographical characteristics, and the geological composition of the soil. To mitigate the catastrophic impact of landslides, timely detection and estimation of land displacement are crucial. Remote sensing data plays a pivotal role in this regard, enabling the analysis of landslide detection and displacement. However, the challenge lies in the requirement for extensive data processing of satellite imagery and in situ observations, which can be computationally demanding.

To address this challenge, automated analysis approaches that incorporate computer vision and machine learning techniques have been proposed. However, these approaches often lack scalability and do not support cloud-optimized development. In this work, we present Landslideflow, a Common Workflow Language (CWL) based workflow designed to generate landslide displacement maps as output, given multi-modal remote sensing datasets including satellite images (MSI), digital elevation models (DEMs), derived slope, and aspect maps as inputs. The proposed Landslideflow is based on ZOO-Project and ZOO-Services.

2.2.  ZOO-Project and ZOO-Services

The ZOO-Project is an Open-Source Processing Engine and an OSGeo Project that adheres to the OGC Standards, commencing with WPS 1.0.0 and then 2.0.0. It serves as a Reference Implementation of the OGC API — Processes — Part 1: Core. The ZOO-Project is built upon the ZOO-Kernel, which is capable of managing the execution of local processes implemented in various programming languages, including C/C++, Java, C\#, Python, PHP, Perl, Ruby, Fortran, R, and JavaScript (limozjs or Node.js). Furthermore, it supports existing processing applications such as OrfeoToolBox (OTB) and SAGA-GIS, automatically transforming them into standard OGC Web Services or API..

The primary focus of the ZOO-Project is the utilization of the MapServer for on-demand publication of Open Geospatial Consortium (OGC) Web Services (WMS, WFS, and WCS). To ensure the successful publication of a resultant dataset through MapServer, modification of solely the process’s metadata is necessary. This eliminates the requirement for alterations to the process code. The ZOO-Project inherently supports flexibility and facilitates seamless integration with a diverse range of other open-source libraries, including GDAL, CGAL, GRASS GIS, OrfeoToolbox, SAGA GIS, and many more. It is designed to utilize existing geospatial algorithms through the standardized Web Processing Service (WPS) and offers numerous notable examples to assist in the construction of customized solutions.

The ZOO-Project has been used in various projects including the GeoSUD and Phidias-HPC project. These projects have facilitated the implementation of remote task execution on High-Performance Computing (HPC) systems, utilising the Secure Shell (SSH) protocol for connectivity purposes, as well as the SLURM workload manager for scheduling and monitoring tasks. Initially, applications from the OrfeoToolBox (OTB) were employed, which were later supplemented with workflows encapsulated in Singularity containers. This implementation was based on the WPS 2.0.0 Standard and leveraged the inner outputs to provide the URLs to the OGC Web Services (WMS+WFS or WMS+WCS) associated with the result.

Within the context of OGC Testbed19, the ZOO-Project was employed to facilitate the deployment of singularity containers encompassing workflows or individual applications, subsequently executing them on a remote high-performance computing (HPC) infrastructure. This framework leverages the Open Geospatial Consortium (OGC) API — Processes — Part 2: Deploy, Replace, and Undeploy (DRU) draft specification for deployment purposes, whereas Part 1: Core is utilised for process execution invocation. Drawing upon the OGC Best Practice for Earth Observation Application Package, the ZOO-Project was integrated into the Earth Observation European Platform for Cloud and Analytics (EOEPCA) project to support the Common Workflow Language (CWL) conformance class outlined in the OGC API — Processes — Part 2: DRU draft. Furthermore, the ZOO-Project-DRU Helm chart simplifies the deployment of the ZOO-Project, encompassing both DRU and CWL support.

During the Testbed19, the ZOO-Project prototyped an implementation for deploying OpenEO User-Defined Processes utilising the OGC API-Processes Part 2: DRU draft specification, enabling their execution. Furthermore, this activity showcased a prototype platform that facilitates the coexistence of both eoAPI and the ZOO-Project, with the latter providing a unified OpenAPI for accessing both services. The combination of Processing and STAC API capabilities through a unified API presents novel opportunities. The implementation of DRU capabilities necessitates the fulfilment of specific security requirements. It is impossible to restrict the use of specific endpoints internally if the ZOO-Project passes request headers to the secured services. In this context, it is essential to acknowledge that the ZOO-Project’s OpenAPI enables user authentication via Basic Authentication or OpenID Connect. In the Open Geospatial Consortium (OGC) CDRP 2024, the functionality developed during the Testbed-19 is extended as Landslideflow to support a scientific use case. This use case involves deploying Landslideflow for remote process execution of a landslide displacement approach as a Common Workflow Language (CWL) workflow.

2.3.  Methodology

This section describes the methodology for the proposed Landslideflow platform.

2.3.1.  Brief Understanding of Service Architecture

 

Figure 2 — Illustration of an architecture comprising of OGC web services for a sample landslide detection and monitoring client

Traditionally, a Landslide detection and monitoring client would necessitate interaction with Open Geospatial Consortium (OGC) Web Processing Service (WPS) services and OGC APIs (refer to the figure above). This interaction entails web services interfacing with the Features and imagery databases in the form of Geodatacubes based on predetermined bounding boxes, thereby adhering to the initial defined work for OGC GeoDataCube. The standards for GeoDataCube are still in development.. These web services facilitate data (both raw and derived) access as Request-Response, enabling standardized and interoperable remote processing for a specified geographical region.

2.3.2.  Inference-as-a-service based on OGC API Processes

 

Figure 3 — Illustration of the inference-as-a-service based on OGC API - Processes

The retrieved data is passed to the inference-as-a-service (See above figure) for generating the displacement maps based on the inference generated from the optical flow model. The output is then visualised in a Swagger-Ui client based on the OpenAPI. Some of the processes used in this processing are as shown below.

 

Figure 4 — Illustration of the processes used in the Landslideflow platform

2.3.3.  LandslideFlow: ML Inferencing as a CWL Workflow

 

Figure 5 — Execution mechanisms for Earth observation algorithms

However, this model-data interaction can be abstracted in a containerised platform for deploying the end-to-end analysis using CWL workflow. The OGC EOAP facilitates this development by enabling algorithm execution as an OGC EO Application Package

2.3.4.  Algorithm: Optical Flow based Landslide Displacement Analysis

 

Figure 6 — Illustration of high-level framework for generating landslide displacement maps.

We tested the proposed Landslideflow for two use-cases based on different types of data used for processing.

1. Based on using Sentinel 1 — SAR data comprising of VV and VH polarisation for earthquake induced landslide of Hokkaido, Japan [2] (Source: https://zenodo.org/records/7248056)

2. Based on Sentinel 2 data for rainfall triggered landslides in Taiwan. We used sentinel-2 data made available on AWS by Element84. (Source : https://registry.opendata.aws/sentinel-2-l2a-cogs)

2.3.4.1.  Case-study 1: Hokkaido landslide

2.3.4.1.1.  Study Area

On 6th September 2018 Hokkiado was impacted by an earthquake of magnitude 6.7 causing more than 6000 landslides. The study area spans over Hokkaido, Japan having the temporal range from 7th July, 2018 to to 4th Nov, 2018 and spatial extent covering latitude range from 42°31’31.658” to 43°0’52.7” and longitude from 141°42’35.899” to 142°14’23.546”.

2.3.4.1.2.  Dataset Description

The dataset comprises Sentinel-1 acquired imagery. These images are Radiometrically and Terrain Corrected (RTC) images derived from the “Ground Range Detected”. The RTC images are stacked for VH and VV polarization channels and time. Additionally, the dataset incorporates topography-based data, specifically Digital Elevation Model (DEM), and derived products utilizing DEM such as slope, aspect, and curvature. Furthermore, a label layer is integrated into the dataset.

The dataset is structured with dimensions of time, y-axis, and x-axis. The resultant dataset is a data cube comprising various data variables. In addition to polarimetric information, the dataset also contains interferometric data, including phase and coherence components. However, these components have not been utilised for analysis. We used the data provided by Vanessa et. al. (source: https://zenodo.org/records/7248056)

 

Figure 7 — Study area and components of the generated data cube. (Source: https://arxiv.org/pdf/2211.02869)

2.3.4.2.  Case-study 2: Taiwan Landslide

2.3.4.2.1.  Study Area

 

Figure 8 — Taoyuan City-Fuxing district large-scale collapse location and aerial photo. (Courtesy: Agency of Rural Development and Soil and Water Conservation, ARDSWC, MOA, Taiwan)

2.3.4.2.2.  Dataset Description

Although we did not use the entire provided data, the dataset comprised of data as described in below table along with their sources.

Table 1 — Available datasets

Dataset Name	Type of Data	Source	If used for Modeling
Surface extensometer Readings	Observational	ARDSWC	x
Electronic water pressure gauge	Observational	ARDSWC	x
Geologic Data	Observational	ARDSWC	x
UAV images	Images	ARDSWC	✓
Rain Gauge measurements	Observational	ARDSWC	x
GNSS measurements	Observational	ARDSWC	x
Satellite images	Images	Copernicus Sentinel Data access	✓
Digital Elevation Model	Images	ARDSWC	✓
Slope Map	Images	Derived	✓
Aspect Map	Images	Derived	✓

2.3.4.3.  Implementation

 

Figure 9 — Illustration of the Landslideflow platform for the Taiwan Landslide case study

2.3.5.  ML inference-as-a-CWL Workflow

 

Figure 10 — Deploy inference

 

Figure 11 — Run inference

 

Figure 12 — Access results

2.4.  Broad Impact

This proposed prototype is targeted to provide multiple advantages and create broad impact as follows:

2.5.  Future Directions

2.6.  Conclusion

The work provides a standardised implementation for detection of landslides using imagery as source for two example cases, which will in turn help to improve the detection and decrease the detection time. The decreased detection will further impact the overall response time and support early relief and planning the response. The overall process is standardised and adapts CWL, making it available to users, authorities and government task forces, without involving much technical details.

2.7.  Acknowledgements

We formally acknowledge the Agency of Rural Development and Soil and Water Conservation, ARDSWC, MOA, Taiwan for providing us with the Taiwan Landslide data. Additionally, we recognize and express our gratitude to the scientific community for making the code and data related to this research publicly accessible.