CGHD: Dual-Temporal Dataset of Composite Geological Hazards via Multi-Source Optical Remote Sensing Images

Highlights

What are the main findings?

• We constructed the Composite Geological Hazards Dataset (CGHD), a large-scale, multi-scale and multi-resolution dual-temporal dataset integrating both landslides and debris flows from diverse optical satellite sources.
• Experimental results demonstrate that the proposed use of dual-temporal and multi-source optical remote sensing data in CGHD significantly improves detection accuracy and enhances generalization across diverse geographic environments.

What are the implications of the main findings?

• CGHD establishes a solid data foundation for landslide and debris flows hazard research, enabling models to effectively learn temporal dynamics and adapt to varying spatial resolutions and sensor characteristics in complex terrains.
• This resource is pivotal for advancing intelligent disaster monitoring and prevention, facilitating the development of reliable automated systems for rapid landslides and debris flows mapping and emergency response.

Abstract

Geological hazards are characterized by their sudden occurrence, high destructiveness, and wide spatial impact. In particular, landslides and debris flows triggered by earthquakes and intense rainfall often lead to severe casualties and substantial property losses. Therefore, the rapid delineation of affected areas is crucial for disaster assessment and post-disaster reconstruction. To this end, several geohazard datasets have been developed from remote sensing imagery, focusing on specific regions, disaster types, and data sources, providing valuable support for geohazard detection and risk assessment. Our study addresses the diversity of real-world geological disasters in terms of their types, causes, and spatial distribution and constructs the Composite Geological Hazards Dataset (CGHD), a dual-temporal geohazard dataset that enhances generalisation and practical applicability. CGHD incorporates pre- and post-disaster remote sensing images of 14 landslide and debris flow events that occurred worldwide between 2017 and 2024, collected using four remote sensing platforms and encompassing multiple spatial scales and land-cover categories. The affected areas varied significantly in size and shape, with land-cover types including roads, buildings, vegetation, farmland, and water bodies. This resulted in 3963 pairs of pre- and post-disaster images, each with a size of 1024 × 1024 pixels. We validated the reliability of the CGHD through experiments with nine change-detection models and further evaluated its generalisation capability using an unseen dataset. The experimental results demonstrate that CGHD achieves high recognition accuracy and strong generalisation across diverse geographic environments, providing comprehensive data support for intelligent geohazard recognition and disaster assessment.

1. Introduction

Geological hazards such as debris flows and landslides may be triggered by earthquakes and intense rainfall [1]. These secondary geohazards typically emerge suddenly, cause considerable destruction, and have spatial complexity, potentially posing severe threats to human life and property, critical infrastructure, and the ecological environment [2,3]. The accurate localisation of geohazard-affected areas is a critical foundation for disaster situation analysis, risk assessment, and post-disaster reconstruction [4]. Researchers relied on field investigations to identify geohazard-affected areas, which offered high accuracy, and were the primary method given the limited technical means in the early days [5]. However, geological hazards occur predominantly in mountainous regions, and field surveys are often time-consuming, logistically demanding, and inherently dangerous [6].

With the continuous advancement of remote sensing imaging and digital image processing technologies, remote sensing imagery has been widely applied to identify geohazard-affected areas owing to its high resolution and wide coverage. Recently, synthetic aperture radar (SAR) and optical images have been increasingly used for geohazard detection. Cigna et al. [7] employed SAR images and applied the persistent scatterer interferometry technique, together with a time-series post-processing approach, to detect ground deformation patterns associated with the 2005 geohazard event in the Naro area of Italy. However, the interpretation of SAR imagery is often challenging in interpreting cities or areas with complex features, and its revisit cycle limits its effectiveness for rapid disaster assessment. By contrast, optical imagery is more straightforward for visual interpretation and is particularly suitable for large-scale disaster detection. Sato and Harp [8] identified 257 large-scale landslides triggered by the Wenchuan earthquake on 12 May 2008 in Beichuan County, using FORMOSAT-2 images, SPOT-5 images, and digital elevation model (DEM) data. Similarly, Martha et al. [9] mapped 1196 landslides induced by the 2011 Sikkim earthquake in India based on 123 high-resolution optical remote sensing images. Numerous similar datasets have been constructed in earlier studies and successfully applied to their respective research domains. However, the geographic locations, geological conditions, and disaster types of these datasets have limited their application and expansion. The construction of high-quality and representative datasets has become a key approach to improving the accuracy and robustness of disaster scene analysis. Consequently, researchers have devoted substantial efforts to dataset construction and refinement, particularly focusing on debris flows and landslide hazards, for which numerous effective datasets have already been established.

In debris-flow research, Pham and Kim [10] extracted images from videos of debris-flow events and obtained 5950 valid debris-flow-containing samples that were then used to construct the Debrisflow21 dataset. Yuan et al. [11] combined Gaofen-1 imagery with DEM data from Nujiang, Yunnan, and designed a dual-channel convolutional neural network to evaluate debris-flow susceptibility. Ma et al. [12] created a Google remote sensing image dataset of debris flows in the Xinjiang Uygur Autonomous Region, China. This dataset comprises 4585 JPG images, each with a resolution of 512 × 512 pixels. Although these studies have made valuable contributions to debris-flow detection and susceptibility assessments, the field lacks an open and comprehensive remote sensing image dataset.

In landslide research, Ji et al. [13] developed a public landslide detection dataset using TripleSat satellite imagery, which included 770 annotated landslide instances in Bijie City, Guizhou Province, China. The Landslide4Sense dataset was constructed from Sentinel-2 multispectral imagery covering four landslide events and served as a valuable source of spectral information [14]. The High-Resolution Global landslide Detector Database (HR-GLDD) was compiled from PlanetScope imagery of landslide instances across 10 different regions worldwide [15]. Although these datasets vary in geographic coverage, temporal depth, and spectral characteristics, most rely on data from a single sensor, thereby limiting the generalisation and robustness of models when applied to multi-source optical remote sensing data and multi-resolution scenarios. The Chinese Academy of Sciences (CAS) landslide dataset integrated satellite and unmanned aerial vehicles (UAV) imagery from nine regions and comprised 20,865 remote sensing images [16]. The Globally Distributed Coseismic Landslide Dataset (GDCLD) incorporates multisource remote sensing imagery, including PlanetScope, Gaofen-6, Tianditu, and UAV data, and covers nine disaster events across diverse geographic and geological contexts [17]. The Large-scale Multi-source High-resolution Landslide Dataset (LMHLD) further increases data diversity by collecting images from five different satellite sensors across seven global study areas [18]. Most of the aforementioned datasets achieved a relatively high accuracy within their respective tasks, thereby demonstrating their effectiveness for specific disaster scenarios. However, their reliance on single-temporal, post-disaster imagery limits their applicability to other regions and hinders accurate localisation of affected areas in complex real-world scenarios. In contrast, dual-temporal remote sensing imagery not only reveals the evolutionary differences in surface features before and after disasters but also delineates affected regions more clearly. Moreover, it effectively reduces the interference caused by inherent geomorphic features such as exposed rocks that resemble disaster characteristics in single-temporal data. A dual-temporal Global Very-high-resolution Landslide Mapping (GVLM) dataset was constructed using high-resolution pre- and post-disaster images from 17 disaster events worldwide via Google Earth, providing a valuable reference for change-detection research across diverse disaster scenarios [19]. The experimental results indicate that dual-temporal data enable a more accurate extraction of affected areas. Therefore, the development of multi-source dual-temporal optical remote sensing disaster datasets is essential for accurately locating affected areas and analysing complex disaster scenarios.

Landslides and debris flows exhibit highly similar visual characteristics in optical remote sensing imagery, such as a significant reduction in vegetation cover, exposure of surface soil or rock, and typical elongated or tongue-shaped patterns. In addition, these two types of hazards often show strong cascading behavior within disaster chains [20]. However, most existing datasets focus on a single type of hazard (e.g., landslides or debris flows) and a fixed region, which limits their applicability to joint analysis and scenario modelling of landslides and debris flows. Focusing on the concurrency of geological disasters and the complex characteristics of geohazard-affected areas, including diverse surface features, irregular spatial distributions, and two disaster types, we constructed the Composite Geological Hazards Dataset (CGHD). CGHD integrates pre- and post-event imagery from 14 landslide and debris-flow events triggered by earthquakes or intense rainfall and covers multiple remote sensing platforms, including Google Earth, Jilin-1 Satellites, the European Space Agency (Sentinel-2), and Maxar (WorldView series). The spatial resolution varies from 0.3 m to 10 m, which ensures high spatial detail and broad variation in affected area scales. After data collection, standardized preprocessing and cropping were performed, resulting in a total of 3963 dual-temporal remote sensing image pairs with a spatial size of 1024 × 1024 pixels. To establish a systematic benchmark, nine mainstream change detection models were evaluated on the CGHD. The experimental results demonstrate that CGHD is highly challenging and possesses strong discriminative capability, enabling effective differentiation of model performance. Furthermore, generalization experiments conducted on an unseen dataset further validate the cross-regional adaptability and robustness of the dataset. Compared with existing public datasets, CGHD exhibits significant advantages in terms of land-cover complexity, scale diversity, spatial coverage, and multi-resolution characteristics. It effectively captures the pre- and post-event changes of landslides and debris flows under diverse environmental conditions, thereby providing a solid data foundation for automated geological hazard identification and change detection in complex disaster scenarios.

The remainder of this paper is structured as follows: Section 2 describes the data collection and preparation process of CGHD, focusing on data acquisition channels, dual-temporal image registration, and label generation. Section 3 details the dataset validation methods using typical change-detection algorithms, including model architectures, loss functions, and training parameter settings. Section 4 presents the experimental results, evaluates the reliability and accuracy of CGHD, and examines the generalisability of the models trained on this dataset for cross-regional disaster identification.

2. Composite Geological Hazards Dataset (CGHD)

The construction of CGHD in our study involved two main stages: the collection of remote sensing imagery and the creation and verification of labels. In the first stage, we collected pre- and post-disaster remote sensing images of landslide and debris-flow events triggered by earthquakes or intense rainfall over the past eight years from four data platforms. In the second stage, we generated pixel-level label maps based on historical disaster records and expert interpretations and conducted multiple rounds of cross-checking to ensure consistency and accuracy. Figure 1 illustrates the complete workflow, from raw data acquisition to final label production.

2.1. Study Areas

Based on a comprehensive survey and analysis, we selected 14 representative geological hazard events triggered by earthquakes or intense rainfall, covering six continents worldwide, including Asia, Europe, Africa, North America, South America, and Oceania, as shown in Figure 2. These disasters affect diverse landforms, including plateaus, mountains, hills, plains, and basins. Our work focuses on several representative regions, including the Qinghai-Tibet Plateau, Lianhua Mountains, Massif de la Hotte, Longmenshan Fault Zone, East African Rift Valley, Himalayas, and the Caucasus Mountains. The lithological conditions vary, ranging from dolomite, gabbro, and limestone to granite and shale. The climatic conditions are diverse and include tropical rainforests, tropical monsoons, subtropical humid, temperate continental, and alpine climates. Overall, the dataset exhibits high diversity in spatial distribution, geomorphic features, lithology, and climatic conditions, providing extensive data for geological hazard analyses. Detailed information on the geological hazard events in our dataset is provided in

Table 1. Overview of geological hazard events contained in the CGHD.

Locations	Geographic Coordinates	Occurrence Date	Triggers	Disaster Type
Jiuzhaigou	33.28°N, 103.76°E	2017.08.08	Earthquake	Landslides
Sierra Leone	08.43°N, 13.23°W	2017.08.14	Rainfall	Landslides, debris flows
Guangdong	23.27°N, 115.57°E	2018.08.27–31	Rainfall	Landslides, debris flows
Haiti	18.36°N, 74.00°W	2021.08.14	Earthquake	Landslides
Lushan	30.37°N, 102.94°E	2022.06.01	Earthquake	Landslides
Luding	29.59°N, 102.08°E	2022.09.05	Earthquake	Landslides
DR Congo	02.01°S, 28.90°E	2023.05.02–05	Rainfall	Landslides, debris flows
Italy	44.18°N, 11.85°E	2023.05.16–17	Rainfall	Landslides, debris flows
Bhutan	27.57°N, 91.07°E	2023.07.20	Rainfall	Landslides, debris flows
Georgia	42.70°N, 43.64°E	2023.08.03	Rainfall	Landslides, debris flows
Brazil	29.25°S, 51.89°W	2024.04.23–05.06	Rainfall	Landslides, debris flows
Papua New Guinea	05.37°S, 143.36°E	2024.05.24	Rainfall	Landslides
Ya’an	29.57°N, 103.75°E	2024.07.20	Rainfall	Landslides, debris flows
Nepal	27.62°N, 85.38°E	2024.09.26–28	Rainfall	Landslides, debris flows

.

• Jiuzhaigou
  On 8 August 2017, an Mw 6.5 earthquake struck Jiuzhaigou County, Sichuan Province, China, with an epicentre at 103.76°E and 33.28°N. The region exhibits an average elevation above 4 km, with predominant slope gradients of over 30° [21]. The earthquake triggered at least 4834 landslides, mainly small- to medium-scale rockfalls and debris avalanches, primarily occurring in regions underlain by Carboniferous limestone and dolomite. The total landslide-affected area was 9.64 km², of which 189 individual landslides exceeded 0.01 km² in area [22].
• Sierra Leone
  On 14 August 2017, the capital city of Freetown, Sierra Leone, experienced large-scale flooding and debris flows following three consecutive days of heavy rainfall, with a total precipitation of 1040 mm. The disaster resulted in over 500 deaths, approximately 600 missing persons, and the destruction of hundreds of houses [23]. This mixed disaster primarily involved landslides, debris flows, and floods, resulting in widespread and severe damage. The local geology is dominated by weathered gabbro with low resistance to weathering, whereas the terrain is characterised by steep slopes and strong relief, which reduce slope stability and exacerbate the risk of slope failure [24].
• Guangdong
  From 27 to 31 August 2018, the eastern region of Guangdong Province experienced continuous heavy rainfall that triggered widespread geological hazards, including landslides and debris flows. The affected areas are situated within the Lianhua Mountain Range, characterised predominantly by mountainous, hilly, and plain topography with pronounced elevation variations. During this heavy rainfall event, 1844 rainfall-induced landslides were recorded in Jiexi County, with a total affected area of approximately 3.39 km². The largest single landslide measured 0.0223 km², while the smallest measured 0.000417 km² [25]. In Luhe County, 2241 landslide events were identified, with affected areas ranging from 0.000126 to 0.019761 km², reflecting their widespread distribution and notable variations in scale [26].
• Haiti
  On 14 August 2021, a magnitude 7.2 earthquake struck the Nippes region of Haiti with its epicentre at 18.36°N and 74.00°W. The earthquake, combined with Hurricane Grace two days later, triggered numerous geological hazards. The calamity claimed over 2500 lives, injured more than 10,000 people, and caused at least 8444 landslides with a total affected area of 45.6 km² [27]. The landslides were mainly distributed near Pico Macaya National Park in Massif de la Hotte, and approximately 89.4% of them occurred at elevations above 1 km. Limestone outcrops constitute the primary landslide development zones [28,29].
• Lushan
  On 1 June 2022, an Mw 5.8 earthquake struck Lushan County, Sichuan Province, China. The epicentre was located at 30.37°N, 102.94°E, with a focal depth of approximately 12.0 km. The earthquake occurred along the southern segment of the Longmenshan Fault Zone and caused at least 2352 landslides across an affected area of approximately 1470 km². The total landslide area was 5.51 km², with an average individual landslide size of approximately 0.0023 km². The landslides were predominantly concentrated in areas with slopes ranging from 40° to 50° and elevations between 1.3 and 2.5 km [30]. The geology of the region is primarily composed of intrusive rocks, including granite, gabbro, and amphibolite [31].
• Luding
  On 5 September 2022, an Mw 6.6 earthquake struck Luding County, Sichuan Province, China. The epicentre was located at 29.59°N, 102.08°E, with a focal depth of approximately 16.0 km [32]. The epicentral area lies within the Hengduan Mountains in the southeastern Qinghai-Tibet Plateau and features typical mountainous gorge topography [33]. The earthquake triggered 9142 landslides, covering an aggregate area of 49.51 km² [34]. Landslides were mainly concentrated at elevations between 1 and 2.3 km and slopes ranging from 20° to 50°, with areas underlain by granite formations and dense forest cover being particularly susceptible to slope failure [35].
• DR Congo
  From 2 to 5 May 2023, the Great Lakes region of Africa experienced intense rainfall, which triggered severe floods, landslides, and debris flows in the eastern Democratic Republic of Congo and western Rwanda [36]. The affected areas are located within the East African Rift Valley. This region is characterised by a predominantly mountainous terrain and frequent tectonic activity, which makes it highly susceptible to landslides and erosion [37,38]. In the Kalehe region of the DR Congo, 452 people were killed, 6206 went missing, and 200 were injured, with extensive destruction of infrastructure, including homes, schools, and medical facilities [39]. In Rwanda, 136 people died, 112 were injured, and 2713 houses were destroyed [40].
• Italy
  From 1 to 17 May 2023, the Emilia-Romagna region of Italy experienced two successive episodes of heavy rainfall that triggered floods and 80,997 landslides. The disaster resulted in economic losses exceeding €9 billion and 17 fatalities [41]. The northeastern part consists of low plains where river levels are higher than the surrounding terrain, which makes it highly prone to flooding, while the southwestern part comprises the Apennine Hills with slopes ranging from 0° to 70°, where the complex topography significantly increases the likelihood of landslides [42]. The disaster affected a total landslide area of 83.33 km², with the largest single landslide covering 0.5 km² and an average landslide area of approximately 0.001039 km² [43].
• Bhutan
  On 20 July 2023, heavy rainfall struck Wangkar in Bhutan, triggering flash floods and mudslides that partially destroyed a small hydropower station. The disaster resulted in at least seven fatalities and 16 missing people [44]. Situated in the Himalayan region, Bhutan is one of the world’s most geologically fragile and meteorologically active zones, making it highly susceptible to diverse natural hazards [45]. The disaster occurred within an area characterised by exceptionally steep topography and pronounced erosion. In this area, two minor tributaries converge in the main river channel, forming a typical geological setting in which small-scale landslides accumulate and evolve into channel-type debris flows [46].
• Georgia
  On 3 August 2023, a catastrophic debris flow occurred in the Shovi region of Georgia owing to the combined effects of heavy rainfall and glacial melt. This disaster resulted in 32 fatalities and the deposition of approximately 1 million m³ of sediment. Situated within the Caucasus Mountains, the geological structure of the region primarily comprises Jurassic–Lower Cretaceous shales, clay shales, sandstones, and limestone. In some areas, prolonged erosion has carved depths exceeding 1 km, with slopes generally ranging between 40° and 60° [47]. The high-altitude zones fall within the glacial belt, whereas the lower elevations exhibit alpine-subalpine landscapes. This region has historically experienced recurrent landslides and debris-flow disasters [48].
• Brazil
  Between 23 April and 6 May 2024, the state of Rio Grande do Sul in Brazil experienced extremely heavy rainfall, with cumulative precipitation reaching 850 mm. This triggered 15,057 geological hazards, predominantly translational landslides and debris flows. The affected area spanned approximately 92 km², with individual landslides averaging 0.0066 km². The smallest recorded landslide covered 0.000075 km², whereas the largest exceeded 0.1 km². The disaster resulted in 183 fatalities, 27 missing persons, and the displacement of approximately 500,000 people. Landslides predominantly occur in mountainous terrain at elevations between 200 and 400 m, with slopes ranging from 30° to 35° facing north, northeast, or east. The geological substrate primarily consists of basalt and andesite formations [49].
• Papua New Guinea
  On 24 May 2024, a village in Enga Province, Papua New Guinea, was affected by a large-scale landslide. Rainfall in the region during the first five months of 2024 reached its highest level within a decade, with landslides triggered by a combination of tectonic activity, heavy rainfall, and steep topography. The landslide deposit covers approximately 0.072 km², with a total volume of approximately 500,000 m³. Papua New Guinea lies along the active boundary between the Australian and Pacific tectonic plates, with an average elevation exceeding 2 km. The landslide source area features a slope gradient of approximately 60°, a vertical drop of nearly 100 m, and geological formations dominated by low-strength, heavily weathered quartz sandstone and limestone. Since the beginning of the 21st century, nearly 50 landslide events have occurred in this region, approximately 90% of which were triggered by rainfall [50].
• Ya’an
  On 20 July 2024, sudden flash floods and debris flows triggered by heavy rainfall in Hanyuan County, Ya’an City, Sichuan Province, disrupted local communications and infrastructure and caused 41 fatalities or missing persons [51]. Hanyuan County lies on the eastern margin of the northern section of the Hengduan Mountains and is characterised by a landscape dominated by low-to-medium mountains and river valleys [52]. The topography exhibits extreme elevation variation within the watershed, with the highest peak reaching 2.8 km above sea level and the lowest elevation at 1.6 m. Approximately 79% of the watershed’s slopes range between 30° and 50° [53]. Furthermore, the region exhibits fractured bedrock and active neotectonic movement. Geologically, the exposed strata include Permian basalts and limestones, with intense rock weathering alongside widespread Quaternary loose deposits, which further exacerbate disaster risks [54].
• Nepal
  From 26 to 28 September 2024, central Nepal experienced prolonged heavy rainfall that triggered extensive landslides and debris flows. The disaster caused 250 fatalities, 18 missing persons, and economic losses exceeding US $341 million. The affected area lies within a mountainous basin at elevations ranging from 1.2 to 2.7 km, situated within the Lesser Himalayas. The region exhibits complex geological conditions, characterised predominantly by weathered rock masses and weak strata such as phyllite and slate, with a steep topography featuring slopes predominantly between 35° and 45° [55,56].

Among the 14 events, 10 were triggered by rainfall and 4 by earthquakes. This imbalance reflects the natural occurrence patterns of these hazards. In the study regions, rainfall-induced shallow landslides and debris flows occur more frequently than coseismic landslides. Although efforts were made to collect additional earthquake-triggered events, the observed imbalance is consistent with the inherent distribution of hazard occurrences in nature.

2.2. Data Collection

In our study, we integrated pre- and post-disaster images from the above-mentioned 14 study areas, obtained from Google Earth, Jilin-1 Satellites, European Space Agency, and Maxar. Multiple rounds of internal cross-validation were conducted to rigorously control data consistency, usability, and annotation accuracy, which ensures the reliability and suitability of the dataset for high-resolution geohazard analysis. Detailed image information, including locations, data sources, spatial resolutions, acquisition dates, and research areas, is provided in

Table 2. Detailed information on the remote sensing data sources used in each case study.

Locations	Data Sources	Resolution	Pre-Image Capture Time	Post-Image Capture Time	Research Area [km2]
Jiuzhaigou	Google Earth	0.59 m	2015.12.07	2017.08.14	140.27
Sierra Leone	MAXAR	0.33 m	2017.03.03	–	36.20
		0.72 m	–	2017.08.15
Guangdong	Google Earth	0.59 m	2017.12.07, 2018.03.10	2018.09.15	47.79
Haiti	Google Earth	0.59 m	2020.01.10, 2021.02.27, 2021.05.18	2021.08.23, 2022.12.08, 2021.08.28	98.42
Lushan	Jilin-1 Satellites	0.75 m	2021.08.03, 2021.11.10	2022.07.06	476.35
		0.5 m	–	2022.06.17
Luding	Google Earth	0.59 m	2019.12.25, 2021.02.03	2022.09.10	120.58
DR Congo	MAXAR	0.3 m	2023.04.10	2023.05.12	36.17
Italy	MAXAR	0.3 m	2021.05.20, 2023.04.06	2023.05.23	918.59
Bhutan	Google Earth	0.59 m	2020.11.09	2023.10.25	4.59
Georgia	MAXAR	0.3 m	2017.06.27	2023.08.08	44.30
Brazil	MAXAR	0.3 m	2024.03.02	2024.05.07	482.09
Papua New Guinea	MAXAR	0.3 m	2023.06.27	2024.05.27	216.33
Ya’an	European Space Agency	10 m	2024.06.12	2024.08.01	12,056.04
Nepal	MAXAR	0.3 m	2024.01.13, 2024.04.25	2024.10.06	415.07
Sum	–	–	–	–	15,092.79

. The collected image data comprised 260 pairs of dual-temporal images, with individual image dimensions ranging from 2112 × 1406 to 37,691 × 37,010 pixels, an average pixel size of 12,858 × 11,724. The spatial resolutions varying between 0.3 m and 10 m. The image acquisition spanned multiple pre- and post-disaster time points and covers a total area of approximately 15,092.79 km², with the affected area accounting for about 73.51 km². The collected images exhibited both rich spatiotemporal diversity and extensive geographic coverage, which allowed fine-scale feature extraction at local scales while supporting comprehensive analyses over large-scale regions, thus fulfilling the demands of scientific research work.

As shown in Table 2, in some cases (e.g., Jiuzhaigou), the temporal interval between pre- and post-event images reaches 1–2 years. Such relatively long intervals may introduce uncertainties in change detection. This limitation is primarily due to the difficulty of acquiring high-quality, cloud-free optical imagery, as natural disasters are often accompanied by extreme weather conditions (e.g., heavy rainfall and persistent cloud cover). In addition, platforms such as Google Earth integrate imagery from multiple satellite sources and provide selectively processed observations, which may further constrain the temporal proximity between available images and the actual event time. Despite these constraints, we have made every effort to select imagery as close as possible to the event occurrence. From another perspective, the presence of such temporal gaps reflects the practical challenges commonly encountered in real-world applications, and highlights the necessity for developing change detection models that are robust to temporal inconsistencies.

2.3. Datase Preprocessing

In this study, we first performed a dual-temporal registration of pre- and post-disaster remote sensing images using ENVI 5.6 software and the OpenCV 4.10.0 library to ensure spatial consistency. For pre- and post-event images with different spatial resolutions within the same hazard event, the images are resampled to a unified resolution during the co-registration process to achieve pixel-level spatial alignment. Subsequently, the affected areas were annotated with polygons using the LabelMe 5.6 tool and Adobe Photoshop 25.0.0. The labelling process relied on a comparison of morphological changes between co-registered pre- and post-disaster images, with a focus on identifying variations in surface texture, color, and geomorphological characteristics associated with landslides and debris flows before and after the events. The labeling process follows established practices in landslide and debris flow mapping [57,58]. Figure 3 presents representative examples of polygon annotations and illustrates that the use of dual-temporal imagery effectively distinguishes inherent landforms from actual hazard zones, thereby improving the accuracy of geological hazard identification and interpretation, particularly enabling a more precise annotation of bare land (Figure 3a,f,i) or cultivated fields (Figure 3h,k) that are difficult to differentiate in single-temporal images.

After completing the polygon annotations, all results were converted into semantic-level label maps. Specifically, polygons were drawn onto binary masks using NumPy and OpenCV in Python and saved as PNG images using the Pillow library, where a pixel value of 255 indicated the foreground (disaster-affected areas) and that of 0 indicated the background (non-affected areas). In particular, our project team had an expert team dedicated to ensuring the scientific and accurate nature of the labels. The expert team comprised five domain experts: two professors with more than 12 years of experience in disaster monitoring and assessment, two senior researchers with more than 10 years of research experience, and one technical expert with extensive experience in dataset construction and evaluation.

We conducted a year-long collaborative effort to interpret and semantically annotate pre- and post-disaster remote sensing images with careful attention to content, quality, and research relevance. Pixel-level reviews and successive rounds of refinement removed obvious sources of interference and produced the high-precision CGHD, which contains 3963 non-overlapping subsamples of 1024 × 1024 pixels. Among these subsamples, 1800 pairs contained disaster-stricken areas of varying sizes. The smallest affected area contains only 5 pixels at a spatial resolution of 10 m, while the largest affected area contains 1,039,461 pixels at a spatial resolution of 0.3 m. In total, approximately 1.31 × 10⁸ pixels are annotated as affected regions.

CGHD enhances the robustness and generalisation of geological hazard recognition tasks by incorporating a set of representative negative samples, including clouds, roads, rivers, buildings, cultivated land, and shadows, which can easily be confused with geological hazards in remote sensing analyses. Figure 4 presents the respective image sample of each event in CGHD and shows that the dataset effectively mitigates interference from clouds (Figure 4d,f,j,n), various types of roads (Figure 4a,c,e,h,j,m,n), rivers (Figure 4e), buildings (Figure 4a,b,e,i,k,l,n), cultivated land (Figure 4h,i), and shadows (Figure 4d,f), thereby improving recognition accuracy in complex scenarios.

2.4. Dataset Sample Richness

We quantitatively assessed the information richness and complexity of CGHD by employing the Shannon entropy [59] as a metric and compared it with the representative GVLM dataset, which is widely regarded as a high-quality dataset in the field of landslide identification [60]. Shannon entropy captures the uncertainty in pixel intensity or colour distribution, effectively reflecting image complexity and diversity. In our study, the Shannon entropy was calculated jointly across the RGB channels to fully capture the information distribution of the colour images. The computation was defined as follows:

$$H(R,G,B)=-\sum _{r,g,b}p(r,g,b){log}_{2}p(r,g,b)$$

(1)

Here, p(r,g,b) denotes the joint probability distribution of the pixels with values (r,g,b). Higher entropy indicates that the image contains more information and exhibits greater texture complexity, whereas lower entropy corresponds to simpler textures and less information. Therefore, a wider entropy range indicates that the dataset covers a broader spectrum of scene complexities and reflects higher diversity and representativeness. As shown in Figure 5, CGHD outperformed the GVLM dataset in both the Shannon entropy range (vertical) and the density of the sample distribution (horizontal) dimensions. Specifically, CGHD exhibited a wider range of Shannon entropy values, implying that it included both low- and high-complexity samples with diverse textures. Moreover, most samples in CGHD were concentrated in regions with higher entropy, indicating that a larger portion of the data originated from areas with complex surface features, which better reflects the characteristics of real-world geological disaster scenarios. Generally, CGHD exhibits a more balanced distribution in terms of information richness and complexity. This feature corresponds more accurately to the authentic characteristics of image information in real-world scenarios, consequently delivering enhanced comprehensiveness and representativeness in the data support for model training.

3. Dataset Validation Methods

To verify the accuracy and scientific validity of the CGHD annotations, we employed change-detection methods. Change detection is a technique that identifies changes in land features by comparing remote sensing images of the same area at different times. The accuracy and labelling quality of the dataset are objectively verified through the overlap analysis between the detection results of advanced methods and the annotated ground truth. When the overlap degree between the detection results and the labelled area is relatively high, it indicates that the annotations effectively reflect the actual surface changes caused by geological hazards, thereby confirming their accuracy and scientific reliability. Therefore, we selected nine representative dual-temporal remote sensing image data change-detection algorithms to evaluate the precision and reliability of the CGHD annotations.

We further assessed the cross-scenario generalisation capability of CGHD by employing a high-quality publicly available GVLM dataset as an independent test dataset. Samples corresponding to the same geohazard events with spatiotemporal overlap between the CGHD and GVLM were removed to ensure complete independence between training and testing. The performance of the models trained on CGHD was then evaluated using GVLM to validate the generalisation capacity of CGHD.

3.1. Change-Detection Algorithms

We selected nine representative change-detection algorithms for remote sensing images that encompass different structural paradigms, including classical convolution-based models, attention-enhanced convolutional models, and transformer architectures. Specifically, FC-Siam-conc represents a typical convolutional neural network approach; DTCDSCN, SNUNet-CD, DMINet, LRBNet, and SEIFNet incorporate attention mechanisms to enhance convolutional feature extraction; BIT and ChangeFormer are purely transformer architectures; and MDA-CD integrates the advantages of both convolution and transformers, forming a hybrid structure.

• FC-Siam-conc [61]. FC-Siam-conc introduced a fully convolutional Siamese architecture for change-detection tasks. During the decoding phase, the skip connection mechanism from UNet is incorporated, directly concatenating features from the corresponding layers of the two encoding streams to fuse spatial details with abstract representations from dual-temporal images. Compared with earlier methods, FC-Siam-conc exhibits superior performance in change detection, which has led to its widespread adoption and subsequent refinement. Consequently, it has become a widely recognised baseline model in the field of change detection.
• DTCDSCN [62]. DTCDSCN aims to address the problems of incomplete change extraction and blurred boundaries in building change detection from remote sensing imagery caused by insufficient feature discriminability. Based on a Siamese convolutional network architecture, the DTCDSCN adopts dual-task joint optimisation, where building change detection serves as the main task and dual-temporal semantic segmentation serves as the auxiliary task, a dual attention module (DAM), and an improved change detection loss (CDL). These designs enable precise change-region extraction and object segmentation in high-resolution building change-detection tasks.
• SNUNet-CD [63]. SNUNet-CD addresses the problems of spatial localisation loss in deep networks, boundary pixel uncertainty, and the omission of small objects in very high-resolution (VHR) remote sensing change detection. This method employs a densely connected Siamese network that constructs dense skip connections within encoder–decoder modules and integrates an ensemble channel attention module (ECAM) to solve these problems. SNUNet-CD effectively improves metrics such as the F1 score, balances accuracy and computational efficiency, and enhances small-object detection accuracy in multiclass remote sensing change-detection tasks.
• BIT [64]. The BIT was proposed to overcome the limitations of convolution-based methods in long-range spatiotemporal dependencies and the inefficiency of nonlocal self-attention mechanisms in high-resolution optical change detection. This method converts dual-temporal image features into a compact set of semantic tokens using a semantic tokenizer, models the token-level spatiotemporal context using a transformer encoder, and refines pixel-level representations using a transformer decoder. The experimental results demonstrate that BIT significantly outperforms convolutional baselines on multiclass high-resolution remote sensing datasets, achieving comparable or superior accuracy with a three-fold lower computational cost and parameter count.
• ChangeFormer [65]. ChangeFormer was designed to address the inability of convolution-based remote sensing change-detection methods to effectively capture long-range spatiotemporal dependencies. It builds on a Siamese architecture and incorporates a hierarchical transformer encoder, four multiscale feature-difference modules, and a lightweight MLP decoder. The experimental results show that it captures finer details of changes across diverse land-cover types in remote sensing imagery.
• DMINet [66]. DMINet aims to mitigate the challenges of class imbalance between the foreground and background, limited training samples, and seasonal variations in remote sensing change detection. The network integrates a cross-temporal joint-attention (JointAtt) module that integrates self-attention (SelfAtt) and cross-attention (CrossAtt), a dual-branch difference acquisition structure that combines pixel-level subtraction and channel-level concatenation, and utilises a multilevel difference aggregation mechanism based on the Hadamard product. DMINet achieved superior performance in limited-sample scenarios and maintained a low computational overhead.
• LRBNet [67]. The LRBNet was proposed to reduce the parameter count and computational complexity of post-disaster building change detection. Based on the Siamese UNet++ architecture, a lightweight residual block (LRB) integrates a lightweight compression module (LCM) and efficient channel attention (ECA) and introduces a multilevel damage feature aggregation attention module (MDAAM). In dual-temporal high-resolution remote sensing imagery before and after disasters, LRBNet achieves high detection accuracy while substantially reducing the computational cost.
• SEIFNet [68]. SEIFNet addresses the issues of false changes and scale variations in remote sensing change detection. It employs a Siamese ResNet18 backbone and integrates a spatiotemporal difference enhancement module (ST-DEM), an adaptive context fusion module (ACFM), and refinement modules. SEIFNet effectively reduces false changes and accurately identifies scale-varying targets in change-detection tasks involving diverse land-cover types in remote sensing imagery.
• MDA-CD [69]. MDA-CD aims to address the challenge of fine-grained feature extraction in building damage assessments under multihazard scenarios. It adopts an encoder-bridge-decoder architecture and integrates global feature aggregation (GFA), dual-temporal image transformer compression (BITC), and subtle feature attention (SFA) modules. MDA-CD improves the accuracy of building damage level classification, particularly for the identification of slightly damaged structures.

The aforementioned methods demonstrate outstanding performance in the field of remote sensing image change detection, with diverse network structures and feature extraction mechanisms that are capable of effectively capturing complex multiscale, local, and global change patterns in geohazard detection tasks. Hence, these methods are suitable for systematic validation and performance evaluation of novel datasets.

3.2. Loss Functions and Accuracy Evaluation

We selected a combination of Binary Cross-Entropy (BCE) loss and Dice loss to evaluate the accuracy of the dataset in addressing the class imbalance problem in disaster-affected area localisation. The BCE loss measures the difference between the predicted probabilities and true labels to ensure overall classification accuracy, whereas the Dice loss calculates the overlap between the predicted regions and ground truth to mitigate class imbalance and enhance the focus of the model on positive samples, such as disaster-affected areas. This combined loss maintains the global classification stability to improve the sensitivity of the model for small target regions and accurately reflects the authenticity and reliability of the dataset annotations. The composite loss formulation based on the BCE and Dice losses is as follows:

$$\begin{array}{cc}\hfill & L(y,\widehat{y})=\lambda L_{\mathrm{BCE}}(y,\widehat{y})+(1-\lambda )L_{\mathrm{Dice}}(y,\widehat{y})\hfill \end{array}$$

(2)

$$\begin{array}{cc}\hfill & L_{\mathrm{BCE}}(y,\widehat{y})=-\left(ylog\left(\widehat{y}\right)+(1-y)log(1-\widehat{y})\right)\hfill \end{array}$$

(3)

$$\begin{array}{cc}\hfill & L_{\mathrm{Dice}}(y,\widehat{y})=1-{\displaystyle \frac{2y\widehat{y}+1}{y+\widehat{y}+1}}\hfill \end{array}$$

(4)

here, L denotes the overall loss function, y represents the true label of the sample, $\widehat{y}$ signifies the predicted probability for the sample, and $\lambda$ is a parameter used to balance the weights of $L_{\mathrm{BCE}}$ and $L_{\mathrm{Dice}}$. Following the optimal parameter configuration validated by Huang et al. [68] across multiple public datasets, this study set $\lambda$ = 0.75 to effectively combine the advantages of both loss functions. In this study, we employed four representative evaluation metrics, namely precision, recall, $F_1$ score, and mean intersection over union (mIoU), to evaluate the performance of the dataset. The respective formulae are as follows:

$$\begin{array}{cc}\hfill & \mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}}\hfill \end{array}$$

(5)

$$\begin{array}{cc}\hfill & \mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}\hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill & F_1={\displaystyle \frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}\hfill \end{array}$$

(7)

$$\begin{array}{cc}\hfill & \mathrm{mIoU}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{T{P}_i}{T{P}_i+F{P}_i+F{N}_i}}\hfill \end{array}$$

(8)

where $TP$ denotes true positives, $FP$ denotes false positives, $TN$ denotes true negatives, $FN$ denotes false negatives, i indicates the class index, and N represents the total number of classes used for the evaluation.

3.3. Equipment and Parameters

The experiments were conducted on a workstation equipped with an Intel^® Xeon^® Gold 6230R CPU @ 2.10 GHz and 500 GB of system memory. The GPU acceleration was provided by an NVIDIA GeForce RTX 3090 with 24 GB of VRAM. The operating system was Ubuntu 20.04.6 LTS. The deep-learning framework employed was PyTorch 1.11.0, running on Python 3.7.0, CUDA 11.3, and cuDNN 8.2.0. For model optimisation, the AdamW optimiser [70] was used with an initial learning rate of 0.0001, momentum of 0.9, weight decay of 0.01, and batch size of 8. Subsequently, the models were trained for 100 epochs.

4. Dataset Validation Results

The CGHD was evaluated through two groups of comparative experiments designed to assess its accuracy and generalizability. The first set of comparative experiments was conducted to verify the accuracy of the dataset. In this experiment, the dataset was split into 52,203 pairs of 256 × 256 pixel images. These pairs were then split into training, validation, and test sets at a ratio of 8:1:1, resulting in 41,762 training pairs, 5220 validation pairs, and 5221 test pairs. Nine state-of-the-art change-detection algorithms were employed to evaluate CGHD, and the experimental results are listed in

Table 3. Comparison results on the test set of the CGHD (best scores are highlighted in bold).

Method	Precision [%]	Recall [%]	${\mathit{F}}_1$ [%]	mIoU [%]
FC-Siam-conc	72.633	65.316	68.781	75.195
DTCDSCN	83.445	80.676	82.037	84.166
SNUNet	84.248	82.307	83.266	85.097
BIT	81.645	70.748	75.807	79.747
ChangeFormer	83.143	78.062	80.522	83.050
DMINet	86.559	78.847	82.355	84.421
LRBNet	85.277	80.176	82.648	84.297
SEIFNet	83.511	80.456	81.955	84.105
MDA-CD	86.632	80.360	83.393	84.818

.

As shown in Table 3, the mIoU values ranged from 75.195% to 85.097%, and the $F_1$ scores on CGHD ranged from 68.781% to 83.393%, and for most models, precision is higher than recall. This indicates that the models tend to focus on regions with significant feature changes while avoiding false positives. This phenomenon indirectly reflects the high quality and fine granularity of the CGHD annotations: due to the presence of numerous small-scale, finely labeled change regions, models aiming for high precision inevitably suffer from relatively higher omission rates. Taking BIT in Table 3 as an example, its recall is relatively low, while its visual results are clean and well-structured, confirming that BIT prioritizes precision and produces almost no false positives, despite missing some subtle changes. Specifically, MDA-CD obtained the best results with a precision of 86.632%, recall of 80.360%, $F_1$ score of 83.393%, and mIoU of 84.818%, whereas FC-Siam-conc achieved the lowest performance with an $F_1$ score of 68.781% and an mIoU of 75.195%. Figure 6 provides a visualisation of the model performance for CGHD. The results indicate that in most scenarios (Figure 6a,c,e), the models produced relatively accurate predictions, whereas in complex cases with cloud cover or shadow interference (Figure 6b,d,f–h), the prediction consistency decreased. Overall, these findings confirm that CGHD demonstrates high reliability and annotation quality across diverse terrains and interference conditions, establishing it as a robust benchmark for evaluating change-detection algorithms.

The second group of comparative experiments was conducted to evaluate the generalisation capability of CGHD. The publicly available GVLM dataset was used as an independent test set to validate the models trained in the first experiment. Potential interference from overlapping data was minimised by excluding the Jiuzhaigou earthquake-induced landslide event, which shared the same source and acquisition time in both the GVLM and CGHD datasets. The remaining GVLM scenes were tiled into 6978 non-overlapping 256 × 256 image pairs and directly processed by nine models previously trained on CGHD. This evaluation protocol effectively separated the influence of model selection from the intrinsic generalisation capability of CGHD. The experimental results of the assessment of the unseen dataset are presented in

Table 4. Comparison of experimental results on the GVLM test set (highest scores in bold).

Method	Precision [%]	Recall [%]	${\mathit{F}}_1$ [%]	mIoU [%]
FC-Siam-conc	59.542	80.754	68.544	73.408
DTCDSCN	69.557	76.584	72.901	76.639
SNUNet	68.326	70.520	69.406	74.355
BIT	78.855	73.856	76.274	79.181
ChangeFormer	58.545	64.645	61.444	69.291
DMINet	72.819	65.405	68.913	74.188
LRBNet	72.300	83.375	77.444	79.682
SEIFNet	72.697	68.785	70.687	75.299
MDA-CD	81.187	74.989	77.975	80.365

.

As shown in Table 4, the models trained on CGHD achieved satisfactory detection performance on the unseen dataset. Specifically, the mIoU values ranged from 69.291% to 80.365%, and the $F_1$ scores ranged from 61.444% to 77.975%. Among the tested models, MDA-CD exhibited the strongest performance on the unseen dataset, achieving a precision of 81.187%, a recall of 74.989%, and $F_1$ and mIoU scores of 77.975% and 80.365%, respectively. Compared to the results obtained on the CGHD test set, the performance degradation was minimal, indicating that the models trained on CGHD maintained high recognition accuracy on unseen data. These results verify the effectiveness of CGHD in enhancing model generalisation and evaluating cross-regional adaptability.

As illustrated in Figure 7, most models accurately identified disaster-affected areas in most scenes (Figure 7a,b,d–h). However, when the affected and unaffected regions shared similar colour and texture characteristics (Figure 7c), the detection performance decreased.

These results indicate that models trained on CGHD exhibited strong generalisation to unseen datasets, further confirming the utility of CGHD for evaluating model adaptability. As shown in Table 4, most models achieve higher recall than precision, indicating a tendency to favor detecting all potential hazard areas at the expense of increased false positives, which aligns with practical requirements where missing any suspicious hazard is undesirable. For example, the LRBNet model achieves relatively high recall on unseen datasets, but its visual results (as shown in Figure 7) contain more false alarms, further illustrating this trade-off. In addition, as illustrated in Figure 7e, all nine models successfully detect small-scale landslide regions that are not annotated in the ground truth, which indirectly suggests the high completeness and reliability of the dataset annotations.

A comparative analysis of Table 3 and Table 4 and Figure 6 and Figure 7 reveals that while CNN-based methods (e.g., SNUNet) excel in-domain, they suffer from fragmented detections on unseen data. Conversely, transformer-based methods (e.g., BIT) and hybrid models (e.g., MDA-CD) provide better generalization and cleaner outputs. These findings provide valuable references for the research of universal change detection models, especially in addressing the complex scenarios of mixed landslide and debris flow disasters, helping researchers to make informed method choices. Meanwhile, the large-scale, multi-scale, multi-resolution geological hazard dataset constructed in this paper provides strong support for improving the model’s generalization ability across different disaster types and change scales, thus advancing further development in the field of change detection.

5. Future Research Directions

Considering the differences in disaster characteristics across various surface types, future work will focus on collecting additional samples that exhibit subtle pre- and post-disaster changes in remote sensing imagery (e.g., landslides in the Loess Plateau region). Moreover, we aim to further enhance the capability of the model to identify weak-change scenarios and improve the diversity and applicability of the dataset. Future studies should consider the long duration and complex evolutionary stages of rainfall-induced landslides and construct multitemporal pre-, during-, and post-disaster image data to explore change-detection methods suitable for different temporal scales.

It should be noted that the dataset constructed in this paper is based solely on optical sensors, which makes it difficult to capture surface deformation processes and movement rates. To further enhance the application range of the model, we will continuously monitor disaster events and expand the dataset, especially by collecting more samples under different surface types and disaster characteristics, in order to improve the model’s ability to identify subtle changes and complex disaster scenarios.

In addition, to enable the precise identification and assessment of geological hazard-affected areas, a specialised disaster-zone detection model is under development. The research and application of this model aim to establish a practical framework for integrating artificial intelligence into geological hazard detection, thereby promoting technological innovation and facilitating industrial applications in this domain.

6. Conclusions

A generalised and practically oriented dual-temporal dataset plays a crucial role in applications such as disaster situation analyses. Landslides and debris flows frequently occur in a coupled manner within disaster chains and exhibit similar visual characteristics in optical imagery. Hence, we constructed the Composite Geological Hazards Dataset (CGHD), which primarily covered landslides and debris flows triggered by earthquakes and heavy rainfall. The dataset spans 14 disaster events across six continents, integrates imagery from four remote sensing platforms, and includes diverse landforms and lithological structures, with a total study area of approximately 15,092.79 km². It provides 3963 pairs of multiresolution pre- and post-disaster images with a spatial size of 1024 × 1024 pixels, effectively addressing the scarcity of high-resolution paired satellite images. Furthermore, we conducted a systematic evaluation of the reliability and accuracy of the dataset using nine change-detection algorithms and further tested nine models trained on CGHD against an unseen dataset to validate their generalisation capability and robustness across diverse geographic environments. The experimental results demonstrate that CGHD exhibits a high recognition accuracy and good generalisation ability. Therefore, CGHD not only provides a solid data foundation for geological hazard research but also offers important support for advancing intelligent disaster monitoring and prevention.