2.1. Multi-Source Information and Characteristics Analysis

Multi-source information is the core component for constructing integrated space-air-ground monitoring. To utilize and analyze this information for intelligent geological hazard monitoring, the applications and data provided by space-based, airborne, and ground-based intelligent monitoring technologies were analyzed separately.

(1) Space-based: Satellite remote sensing information

Satellite remote sensing offers wide-area geographic information using space-based technologies that depict the overall macro-scale status of geological hazards. The global navigation satellite system (GNSS) is an important application that offers real-time monitoring, high-precision positioning, and ease of use, with millimeter-level monitoring accuracy ¹⁶. Critical hazard features were extracted by applying interpretation techniques using satellite remote sensing. Recently, more satellites have been deployed in near-Earth orbit, with high-resolution or ultra-high-resolution optical satellite sensors providing large volumes of Earth observation data ¹⁷. InSAR, with its all-time, all-weather observations, and cloud-penetrating capabilities, achieves millimeter-level monitoring ¹⁸.

Space-based technologies provide extensive coverage and periodic information, but lack detailed local information. The revisit cycles of satellites and time required for data processing limit the real-time nature of space-based information. Optical satellite data generally take hours to days from collection to distribution, whereas SAR data, owing to complex interferometric processing, are delayed by days to weeks. This low update frequency hinders real-time wide-area monitoring requirements and restricts the monitoring of localized geological hazards, thus requiring integration with airborne or ground-based data.

(2) Airborne: Aerial and UAV information

Airborne technologies are characterized by high flexibility and resolution, enabling the acquisition of detailed aerial and UAV information over localized regions, which supports timely and rapid responses and assessments. Airborne platforms equipped with diverse sensors address geological hazards using advanced information processing technologies ¹⁹. Airborne technologies include aerial photogrammetry, UAV-borne multispectral cameras, and airborne LiDAR. Monitoring with UAVs has described the geometric form of geological hazards and evaluated the morphological and topographic changes in recent years ²⁰. Aerial photogrammetry produces RGB optical images. UAV-borne multispectral cameras employ multispectral imaging to capture the detailed surface characteristics. A DEM was generated using airborne LiDAR point clouds. The generated digital orthophoto maps (DOM), digital surface models (DSM), and DEM were used to analyze the landslide morphology. Airborne technologies create three-dimensional (3D) models to support hazard analysis. Furthermore, UAVs are equipped with route planning and intelligent scheduling, leveraging edge computing to improve data collection efficiency, thereby achieving flexible deployment and swift responsiveness. However, constraints on battery capacity and flight altitude make it difficult to cover large areas, thereby limiting the operational and monitoring ranges. Under adverse weather conditions, UAV flights are affected to some extent, which limits their application.

(3) Ground-based: Surface monitoring information

Ground-based monitoring delivers high-precision, real-time measurements of surface deformation and key environmental triggers, such as hydrological and meteorological variables, which are essential for deciphering landslide mechanisms and supporting warning systems ²¹. Typical instruments include water-level gauges ²², ground-based InSAR (GB-InSAR) ²³, and rain gauges ²⁴, which enable continuous unmanned operation under all weather conditions. However, despite ground-based sensors being core high-precision real-time tools in landslide monitoring, particularly suitable for capturing local deformation and short-term dynamics, they are limited by their small coverage, high cost, and deployment difficulty. In practical applications, they are often combined with space- and air-based technologies (such as satellite InSAR and UAV surveys) to form a point-surface coordinated monitoring network, balancing accuracy and coverage to improve the reliability of landslide warnings.

(4) Geological survey materials

By studying geological survey materials, it is possible to identify the type and scale of past geological hazards and analyze the geological processes of evolution and change in a given area. Geological, topographic, and tectonic maps were used to identify potential hazard areas and to provide data for hazard monitoring and prediction.

2.2. Functional Requirements of the Monitoring System

The overall tasks related to geological hazards are summarized into three core functional aspects: intelligent hazard identification, intelligent spatiotemporal prediction, and intelligent real-time warning.

(1) Intelligent hazard identification

Intelligent hazard identification forms the foundation of geological hazard monitoring and analysis of multi-source information to identify impact areas.

(2) Intelligent spatiotemporal prediction

Intelligent spatiotemporal prediction aims to reveal the temporal and spatial evolution patterns of geological hazards and anticipate potential development stages in advance.

(3) Intelligent real-time warning

Intelligent real-time warning by combining outcomes from intelligent hazards identification and spatiotemporal prediction, based on predefined thresholds, promptly warns when hazards are imminent or have already occurred but have not yet caused severe consequences.

2.3. Intelligent Monitoring System Architecture

To satisfy the requirements of intelligent hazard identification, intelligent spatiotemporal prediction, and intelligent real-time warning, the overall architecture was designed to complement each other, as shown in Fig. 1, and included data acquisition, information fusion, intelligent algorithms, and monitoring platforms.

(1) Multi-source information fusion

By analyzing monitoring information from sensors of various types and spatiotemporal characteristics, extracting useful information, and removing redundancies and contradictions, space-based, airborne, and ground-based information is fused for wide-area local-scale detailed and space-air-ground integrated monitoring.

Geological hazard fields involve multi-source and multi-modal data. Through real-time acquisition by multiple sensors, raw data are preprocessed via filtering, normalization, spatiotemporal interpolation, and other techniques to form a multi-source heterogeneous database for hazard monitoring and data-level fusion. Correlating multi-dimensional information, extracting relevant features, and performing similarity matching enable feature-level fusion. The results of feature-level fusion were used to build intelligent hazard identification, intelligent spatiotemporal prediction, and intelligent real-time warning, ultimately achieving decision-level fusion for system functionality.

(2) Intelligent monitoring algorithms

Intelligent monitoring algorithms are at the core of monitoring systems. These are embedded in the system through normal operations by applying deep learning to automatically identify patterns, predict trends, and assess warnings from massive amounts of information to achieve intelligent monitoring.

(3) Database and information management

A database architecture was constructed to organize and manage data effectively, ensuring integrity, consistency, and rapid access, thereby providing strong and efficient data support for analysis, visualization, and intelligent warnings. Through information management, an online intelligent analysis of monitoring data can be performed to extract valuable information, enabling real-time access to target information, and supporting real-time monitoring and decision-making.

(4) Monitoring platform (server, client, and app)

A user interface was designed, and an information service center was established to achieve real-time monitoring of the core geological hazard parameters on both PC and mobile terminals. The server handles core business logic processing, data access control, user authentication and authorization, and task scheduling. The client enables data visualization, presents geological hazards data, and monitors the parameters on the PC. The app provides mobile access, including real-time viewing of data, instant warning notifications, handling, remote control, and service configuration.

3.1. Same Type Fusion Monitoring

Satellite remote sensing data from space-based platforms have been integrated with geological survey materials to achieve long-term, continuous, and intelligent monitoring. Surface condition information was obtained from RGB optical images, ground deformation was measured using InSAR, and fusion algorithms were applied to provide strong support for monitoring. Coordinated multi-source satellite sensing and airborne technologies offer rich and multi-perspective information.

The RGB optical images and the InSAR and DEM datasets were fused using different methods to ensure quality and consistency. Hazard-related features were extracted from different data sources, fused together, and then fed into classifiers and intelligent models for training. Optical features, such as texture, color, and spectral information, and geometric features, such as shape, form, boundary, and edge information, were extracted from the RGB optical images. Surface deformation features were extracted from the InSAR data, and topographic features were extracted from the DEM. These features were combined into a comprehensive feature vector to fully describe the hazardous areas. Incorporating these features into intelligent monitoring algorithms further produced training sample set for geological hazard identification, detailed regional maps, deformation composite maps, and multi-source image feature maps.

Two fusion methods were designed for different data sources. Specifically, decomposition and feature fusion rules ²⁵ were used to enhance the fusion of complementary information. A detailed workflow of the RGB optical images and SAR data is showed in Fig. 3.

The data were decomposed into a set of complementary features: the main structural features and detailed texture features. The source data and decomposed parts are defined as follows:

Next, the pixel saliency map and structure saliency map were extracted by fusing the main structural features. The fusion process is as follows:

Then, the detailed texture features from the optical and SAR images were fused, which were calculated as follows:

Finally, the color transformation method was used to transfer the spectral information, which was calculated as follows:

Inspired by U2Fusion ²⁶, a fusion method for RGB optical images and DEM data was designed to fuse complementary information from different domains. The detailed workflow is showed in Fig. 4. The fusion process is as follows:

3.2. Remote Sensing Collaborative Monitoring

Remote sensing collaborative monitoring focuses on globalized and localized characteristics using information sourced from space-based and airborne technologies. This process involves extracting features from aerial and UAV data for integration with space-based information or directly applying space-based and airborne technologies to monitor hazardous areas.

Surface DEM construction and differencing further revealed topographic changes. Airborne technologies such as LiDAR point clouds generate 3D models. In general, airborne methods transcend the constraints of space-based imagery and capture detailed information.

The RGB optical images obtained through aerial photogrammetry compensate for the limitations of satellite remote sensing imagery. Analyzing the technical strengths, data collection, and processing procedures of UAV systems and extracting the parameters of spatial, geometric, and optical features allows for the effective monitoring of geological hazards. Intelligent control technologies combined with edge computing can be adopted to accommodate diverse application scenarios. The fusion formulation can be expressed as follows:

Space-based and airborne techniques can be combined with on-site verification. The integrated application of multiple technologies and multi-source information further enhances the accuracy and comprehensiveness of hazard monitoring, offering a technical pathway for analyzing the formation patterns of hazards.

3.3. Space-Air-Ground Integrated Monitoring

The space-air-ground integrated monitoring unit satellite coverage, high-resolution aerial observations, and continuous ground-based monitoring. This study conducted deformation monitoring and prediction of geological hazards based on the integration of multi-source data. First, high-precision deformation fields derived from time-series InSAR analysis were combined with on-site GNSS displacement measurements to generate spatiotemporally continuous landslide displacement time series, serving as the baseline for prediction ²⁷. Concurrently, satellite imagery and UAV-derived DSMs have been incorporated to enhance the representation of surface details and landslide geomorphological structures, and to quantify surface changes ²⁸. These multi-source spatial datasets were further integrated with key environmental driving factors, including rainfall and groundwater table variations, to construct a multi-dimensional spatiotemporal feature matrix encompassing the deformation behavior, geomorphic structure, and environmental triggers. Based on this matrix, a deep neural network model was trained to achieve both short- and medium-to-long-term landslide displacement predictions, thereby improving the accuracy of the warning and the robustness of stability assessments.

A multi-source heterogeneous information fusion method was used to develop an intelligent fusion model based on the space-air-ground integrated system for the collaborative dynamic monitoring of geological hazards. Employing optical imagery, InSAR, UAV-based photogrammetry, and wireless sensor networks enable complete process monitoring.

4.1. Intelligent Hazard Identification

Intelligent hazard identification algorithms extract information from images using computer vision-based multi-source data interpretation, identify potential hazards via deep learning, and broadly classify them using semantic segmentation, change detection, and object detection.

(1) Semantic segmentation

Semantic segmentation algorithms still face challenges such as boundary blurring and inaccurate feature extraction, making edge delineation difficult. Transformer-based U-Net (TransUNet) captures local details and the global context in geological hazard image segmentation with strong adaptability to complex topography ²⁹. To address this challenge, improvements have been made to the TransUNet architecture by incorporating a vision encoder and a mask decoder, which are designed to enhance multi-scale feature extraction and facilitate the restoration of edge details, as shown in Fig. 6.

The vision encoder and mask decoder were designed to achieve semantic segmentation of geological hazard areas. The framework can be formulated as follows:

(2) Object detection

Object detection algorithms face limitations when dealing with complex backgrounds and multi-scale features, particularly with large variations in target orientation. YOLO has been widely applied in geological hazard object detection ³⁰. YOLO-based detection models were improved by optimizing the network structure and framework for object detection under complex conditions have been developed, as shown in Fig. 7. For detailed detection in specific scenarios, UAVs are used to enhance the object detection performance, which effectively addresses the real-world challenges of complex backgrounds and small-target interference, improving the detection accuracy and reliability for small targets.

Employing object detection algorithms for area identification can predict the labels of target objects and their existence in the candidate regions of the input images, narrowing the focus of the original images. The framework can be formulated as follows:

(3) Change detection

One challenge in change detection for geological hazards is that the generation of large damaged pixel blocks makes it difficult for models to distinguish between actual hazards and hazard-like features. ResUnet was used to capture the large-scale spatial and temporal dependencies of hazard regions ³¹. An improved hybrid method based on ResUnet treats hazards as objects with multiple attribute information, extracting and modeling features from local elements at various scales and global information. Multi-temporal remote sensing images acquired before and after the hazard were compared to detect changes, as shown in Fig. 8.

A model with two branches was designed to process the pre- and post-change images separately, and change detection was performed by comparing the resulting feature maps. The framework can be formulated as follows:

Intelligent hazard identification employs techniques such as semantic segmentation, change detection, and object detection, making full use of computer vision and deep learning to efficiently interpret multi-source remote sensing information. These algorithms not only improve identification accuracy and efficiency but also provide strong technical support for intelligent monitoring and real-time warning of geological hazards.

4.2. Intelligent Spatiotemporal Prediction

Predicting landslide displacement is a critical component of landslide hazard research. By analyzing temporal changes in landslide displacement, it is possible to predict future development trends. This landslide displacement prediction method was explored from two perspectives: time-series landslide displacement prediction and spatiotemporal coupled modeling landslide displacement prediction.

(1) Time-series landslide displacement prediction

Deep learning models based on time-series have been widely adopted because of their strong capacity to represent dynamic evolutionary processes. These approaches establish a predictive model by exploiting the correlations between historical displacements and exogenous driving factors. Typical model inputs include the cumulative displacement, precipitation, pore-water pressure, groundwater level, and surface temperature, all of which are incorporated as sequential covariates. Existing approaches that incorporate weighted moving average (WMA) decomposition with enhanced long short-term memory (LSTM) architectures have demonstrated significant improvements in prediction accuracy ³².

Building upon these decomposition–reconstruction frameworks, a multi-scale attention-enhanced temporal modeling paradigm has been widely adopted. In this paradigm, multi-scale convolutional operations and hierarchical attention mechanisms are used to capture both local and global dependencies within the displacement sequences. Furthermore, residual connections and dropout regularization were integrated to enhance the model stability and generalization capability, thereby achieving superior prediction accuracy and robustness in complex environmental scenarios. The overall workflow of the time-series prediction paradigm is shown in Fig. 9.

Within this time-series paradigm, let \(\mathbf{y}=\{y_t\}_{t=1}^{T}\) denote the cumulative displacement at a representative monitoring point. A generic decomposition operator \(\mathcal{D}_\phi\) such as a complete ensemble empirical mode decomposition with adaptive noise, variational mode decomposition, or singular spectrum analysis, maps \(\mathbf{y}\) into \(K\) components and a residual as follows:

For each component, multi-scale 1D convolutions with receptive fields \(\{s_\ell\}_{\ell=1}^{L}\) were applied as follows:

A hierarchical attention mechanism then fuses information across components and scales as follows:

The fused representation \(\widehat{\mathbf{h}}_t\) is then fed into a temporal prediction function \(f_{\theta}\) to obtain the future displacement as follows:

(2) Spatiotemporal coupled modeling for landslide displacement prediction

The core of spatiotemporal coupled modeling lies in the simultaneous characterization of spatial correlations between monitoring points and their time-evolving dynamic dependencies. By integrating sequential data, such as displacement observations and rainfall, into graph-structured network models, this approach synchronously captures the underlying propagation mechanisms and evolutionary trends of deformation within the spatiotemporal domain, thereby enabling the accurate prediction of landslide displacement. For multi-point landslide displacement prediction, methodologies based on graph deep learning and the T-GCN model, which combines a GRU with a GCN and an Attention-TGCN, have demonstrated high predictive accuracy and applicability in practical landslide case studies ^33,34.

Building upon these foundations, recent studies have introduced learnable dynamic adjacency matrices and multi-head attention mechanisms to adaptively update the spatial connections between monitoring points and effectively capture non-stationary spatial dependencies. The overall framework of the dynamic spatiotemporal graph model is shown in Fig. 10.

Concurrently, temporal convolutional networks (TCNs) are incorporated to extend the receptive field for long-term sequence modeling. This integration aims to achieve superior cross-scenario generalization capability while concurrently enabling precise modeling of both step-like and gradual deformation patterns. Let \(\mathbf{X}_t\in\mathbb{R}^{N\times F}\) denote the multivariate features of \(N\) monitoring sites at time \(t\) and \(\mathbf{A}_t\in\mathbb{R}^{N\times N}\) be a dynamic adjacency matrix. A generic fusion of physical, data-driven, and attention-based graphs can be written as follows:

After normalization, spatial propagation is performed via graph convolution as follows:

The temporal module models the spatial feature sequence \(\mathbf{Z}_{t-w+1:t}\) of length \(w\) to obtain the hidden state \(\mathbf{H}_t\), and then predicts the displacement vector \(\widehat{\mathbf{y}}_{t+\tau} \in \mathbb{R}^N\) for all monitoring sites at time \(t+\tau\) as follows:

4.3. Intelligent Real-Time Warning

Intelligent real-time warnings are the ultimate goal of intelligent geological hazard monitoring. Based on the evolution of geological environments, real-time monitoring data and historical data were analyzed to issue geological hazard warnings in advance. The warning criteria were designed by integrating probabilistic relationships from multi-source information, and the thresholds were dynamically adjusted. The warning criterion can be expressed as follows:

The warning index can be defined as follows:

When \(R\) exceeded the dynamic threshold \(T_{\mathit{d}}\), a geological hazard warning was triggered. Based on the regional guidance conclusions and pattern inference of intelligent hazard identification and intelligent spatiotemporal prediction, a multi-parameter geological hazard decision-making warning system was constructed, which was applied to real-time data monitoring and warning generation.

5.1. System Configuration

The system based on multi-source information fusion is divided into four layers, combining multi-platform collaborative to achieve hierarchical monitoring. The distributed system configuration is shown in Fig. 11.

The field layer consists of multi-source information acquisition modules, integrating space-, air-, and ground-based sensing to realize multi-scale and all-weather monitoring.

The data layer manages and stores data collected from the monitoring points. It includes functions such as the acquisition, storage, and processing of monitoring data. Data received from the device-end servers were preprocessed using a geographic information systems (GIS) platform and stored in the storage center database.

The business layer includes information management, data monitoring, and algorithm development, which processes, analyzes, and predicts data, identifies hazard regions, issues warnings, and maintains data exchange and coordination with the data layer.

The presentation layer provides real-time depictions of hazard areas, visualization of data from multiple sensors, displays of monitoring components, and warning management functions, including web and mobile applications.

5.2. Practical Applications

Two case studies of landslide identification using a deep learning method and an integrated intelligent monitoring system are presented below.

(1) Landslide identification based on segmentation network

Landslides were classified according to their occurrence time and characteristics into new landslides from Bijie City, Guizhou Province, China, and old landslides from Enshi Prefecture, Hubei Province, China. The landslides in Bijie City were identified using a model that improved U-Net ³⁵. A landslide identification method based on an integrated segmentation network and transfer learning was proposed to address the difficulties with indistinct optical features and improve the accuracy for small sample sizes. The accuracy of the proposed method reached 91.60%, the recall reached 65.49%, and the F1 score reached 62.77%. Experimental results demonstrated that the proposed method achieved improvements in segmentation tasks ³⁶.

(2) Intelligent monitoring and dynamic warning system for geological hazards

An intelligent monitoring system for geological hazards was developed for the Zigui–Badong region in the Three Gorges area, China to enable real-time monitoring and warnings. It integrates multi-source data and supports storage, management, analysis, and visualization through PC and mobile platforms. TCP protocols are used in portions of the area for warning simulations through semi-physical model-database interactions. Utilizing a GIS platform, the system facilitates efficient data processing and predictive model optimization. Deployed in Yichang, Hubei, and Qingyuan, Guangdong, it covers 136 monitoring sites with 733 sensors and provides reliable data and decision-making support for geological hazards ³⁷.

5.3. System Functional Results

This system ensures comprehensive spatiotemporal coverage by integrating space-, air-, and ground-based observations within a unified architecture that overcomes the limited scope of single-source systems. The representative results of typical application scenarios are presented.

The intelligent hazard identification functions are shown in Fig. 12. The spatial extent and morphology of the hazards were captured, indicating that the system provided spatial information.

The function of intelligent displacement prediction is shown in Fig. 13. The predicted time-series followed the observed displacement evolution for several monitoring points, demonstrating that time-series landslide displacement prediction and spatiotemporal coupled modeling for landslide displacement prediction are effective in characterizing the spatiotemporal dynamics of displacement.