Background & Summary

The Qilian Mountains National Park(QMNP), located at the convergence of the Qinghai–Tibet and Inner Mongolia Plateaus, functions as a critical ecological barrier and a nationally designated biodiversity conservation area in northwestern China1,2. Within this landscape, the desert steppe emerges as an ecologically sensitive and biodiverse system that remains highly susceptible to environmental perturbation3. In recent years, population outbreaks of small rodents—most notably the great gerbil (Rhombomys opimus) and the midday gerbil (Meriones meridianus)—have exacerbated vegetation degradation, diminished rangeland productivity, and disrupted ecosystem stability4,5,6.

In response, large-scale chemical control campaigns have been implemented to suppress rodent populations and restore vegetation7,8,9. While effective in the short term, these interventions carry substantial risks for non-target species, particularly native predators, and may trigger cascading ecological effects such as trophic imbalances and biodiversity loss8,10,11,12. Despite these concerns, there remains a paucity of empirical data documenting the spatiotemporal dynamics of rodents, their predators, and other sympatric fauna in arid grassland ecosystems.

To help fill this gap, we employed infrared-triggered camera traps—now widely recognized as indispensable tools for non-invasive wildlife monitoring across extensive spatial and temporal scales13,14,15. These systems are particularly effective for detecting cryptic, nocturnal, and fossorial species, making them ideal for monitoring small mammals and their ecological interactions in desert landscapes. Nonetheless, camera-trap datasets from the desert steppe in QMNP remain scarce.

Here, we present a high-frequency, expert-curated camera-trap dataset from representative desert steppe habitats, situated within a core ecological region of the northern Qilian Mountains in the park. This dataset encompasses vertebrate detections across multiple taxonomic groups and is enriched with detailed metadata, including timestamps, georeferenced camera coordinates, and species-level annotations—together providing a robust platform for ecological analysis and computational research.

By making this dataset publicly accessible, we aim to advance research in community ecology, species distribution modeling, conservation planning, and machine learning-based wildlife identification16,17. The dataset offers crucial insights to inform biodiversity-conscious rodent management and supports evidence-based conservation strategies in fragile dryland ecosystems.

Methods

Study area

This study was conducted in the desert steppe ecosystem of Sunan Yugur Autonomous County, located in the northern Qilian Mountains of Gansu Province, China (38°30′–39°15′N, 99°00′–100°02′E) (Fig. 1). The region experiences an arid continental climate, with a mean annual temperature of approximately 4 ± 0.5 °C and annual precipitation averaging 260 ± 10 mm, most of which occurs between May and August. Vegetation is characterized by drought-tolerant shrub–grassland mosaics dominated by Sympegma regelii, Salsola passerina, Allium polyrhizum, and Reaumuria soongorica.

Fig. 1
figure 1

Location of the study area in Sunan Yugur Autonomous County, Qilian Mountains National Park, Gansu, China.

Full size image

This ecosystem supports a characteristic assemblage of desert-steppe vertebrates, including small mammals such as the midday gerbil and great gerbil, as well as their predators such as the Pallas’s cat (Otocolobus manul) and red fox (Vulpes vulpes). Arid-adapted birds like the isabelline wheatear (Oenanthe isabellina) and ground tit (Pseudopodoces humilis) are also commonly observed. This community exemplifies the biodiversity and trophic interactions typical of Central Asian desert ecosystems18.

Camera trap deployment

From December 2020 to January 2022, 22 infrared-triggered camera traps (model: Ltl-6210PLUS, Shenzhen Ltl Acorn Electronics Co., Ltd.) were systematically deployed across representative shrub–grassland habitats in Sunan Yugur Autonomous County. Site selection was stratified according to rodent population density, which was directly assessed through field surveys prior to camera deployment. To maximize detection efficiency, cameras were preferentially placed in areas with relatively high rodent density. Each camera was mounted 0.5–1.0 m above ground level and positioned to face active rodent burrows or foraging trails, maximizing the probability of wildlife detection19. No baits, including acoustic, olfactory, or food attractants, were used during deployment.

Cameras were equipped with passive infrared (PIR) sensors with a detection range of 18 m and infrared flash illumination effective up to 35 m (850 nm standard lens version). Lenses included either a standard 55° field of view or a 100° wide-angle option. Each trigger event captured three consecutive images with a trigger speed of 0.8 s and no delay between successive triggers (0 s interval), allowing for continuous 24-hour monitoring. Maintenance visits were conducted approximately every two months to ensure operational consistency, including battery replacement and image retrieval.

Image processing and filtering

Captured images were reviewed by trained personnel to exclude empty frames, environmental false triggers (e.g., vegetation movement, precipitation), and unidentifiable animals. A total of 28,881 images were retained as valid wildlife detections.

Each valid image was annotated with standardized metadata, including camera ID, timestamp, and species identification. To reduce temporal autocorrelation and ensure event independence, detections of the same species at the same location were defined as separate events only if they were at least 30 minutes apart—a widely accepted threshold in camera trap studies20,21,22.

Species identification and quality control

Species identification was conducted to the finest taxonomic resolution possible by expert annotators, guided by A Field Guide to the Mammals of China and A Field Guide to the Birds of China. Taxonomy followed the latest frameworks from the Checklist of Mammals in China (2021 edition) and the Classification and Distribution Checklist of Birds of China (3rd edition)23,24,25,26.

Because many small mammal species exhibit cryptic morphology that challenges visual identification in infrared images, supplementary live trapping was conducted both prior to and following camera deployment27. Captured specimens were identified using external features and cranial morphology, serving as physical references for image-based classification.

To ensure annotation consistency and minimize observer bias, all images underwent two rounds of independent screening by separate teams. Any discrepancies were resolved through consensus discussions. This dual-stage validation protocol achieved high inter-observer agreement and ensured reliable species-level classification, enhancing the overall integrity of the dataset28.

Data Records

The dataset is publicly available on Zenodo29,30, under the following DOIs: Valid species images and documentation (https://doi.org/10.5281/zenodo.15730802) and independent detection events and documentation (https://doi.org/10.5281/zenodo.15733158). To support data exploration and reuse, we provide two summary tables: Table 1 describes camera-level sampling effort and output, and Table 3 summarizes species-level detection frequencies.

Table 1 Camera trap locations and survey results.
Full size table

The dataset comprises three integrated components: a validated image set (valid_species_images.zip) containing 28,881 infrared images of wildlife together with a metadata file (metadata.csv); an independent detection event table (independent_events.zip) summarizing 3,828 temporally independent wildlife detections; and a documentation bundle (documentation_and_summaries.csv) providing detailed site coordinates, species detection summaries, and camera operation statistics.

Comprehensive documentation is included to support effective use of the dataset. This includes a README file offering an overview of the dataset and usage instructions, a detailed data dictionary defining metadata fields and file structures, summaries of camera operation effort and species detections to aid analytical planning, and precise camera coordinates in WGS84 georeferenced format to enable spatial analyses.

It comprises a manually validated dataset of infrared camera-trap images and associated metadata collected in the desert steppe of Sunan Yugur Autonomous County, Gansu Province, China. The data reflect spatiotemporal occurrences of vertebrate species captured during a high-frequency monitoring effort. All records underwent dual independent expert reviews to ensure annotation accuracy and consistency. A subset of the dataset represents independent detection events based on standardized temporal thresholds, following common practices in camera-trap research.

All validated entries share a unified metadata schema (Table 2), provided as UTF-8 encoded CSV files to ensure cross-platform compatibility and ease of integration into ecological and computational workflows.

Table 2 Metadata structure for validated wildlife images and independent events.
Full size table
Table 3 Summary of detected species, including Latin names, common names, total images, and independent detections.
Full size table

These records are intended to support a wide range of applications, including species distribution modeling, diel and seasonal activity analysis, biodiversity monitoring, and the development of machine learning models for automated wildlife classification Fig. 2.

Fig. 2
figure 2

Example infrared images of carnivore species captured by camera traps. (a) Felis bieti, (b) Lynx lynx, (c) Otocolobus manul, (d) Vulpes vulpes, (e) Meles leucurus, (f) Mustela nivalis.

Full size image

Technical Validation

To ensure the accuracy and reliability of the dataset, a series of validation procedures was implemented throughout all stages of data collection and processing, including camera functionality, image annotation, species identification, and metadata verification.

Camera operation and data consistency

All 22 infrared-triggered camera traps remained continuously operational from December 2020 to January 2022. Field teams conducted routine maintenance at two-month intervals, including battery replacement and data retrieval, to minimize equipment failure and ensure stable image acquisition throughout the monitoring period.

Image screening and annotation reliability

Image review and annotation were independently conducted by two expert teams from the Engineering Technology Research Center for Rodent Pest Control on Alpine Grasslands, under the National Forestry and Grassland Administration. Annotators systematically removed empty frames, false triggers (e.g., caused by vegetation movement or lighting artifacts), and images containing unidentifiable subjects. Discrepancies between the two teams were resolved through consensus discussions, resulting in a high inter-observer agreement rate of 96%. This dual-review workflow ensured consistency, reduced observer bias, and enhanced annotation reliability.

Species identification accuracy

All species annotations followed authoritative field guides and national taxonomic standards. To support the accurate classification of morphologically similar rodents, live trapping was conducted both prior to and following camera deployment. Captured specimens provided reliable reference material, particularly for distinguishing nocturnal or cryptic species with subtle morphological features. This approach substantially improved the taxonomic resolution and accuracy of rodent identification within the image dataset.

Metadata integrity and file structure

Metadata quality was verified through internal cross-checks to ensure completeness, accuracy, and traceability. Each image file was systematically named to reflect its source camera station, and all metadata were stored in UTF-8 encoded CSV format for compatibility across platforms. Standardized metadata fields—including species name, timestamp, and camera ID—enabled seamless integration into ecological models, activity pattern analyses, and machine learning pipelines.