Non-contact behavioral study through intelligent image analysis is becoming increasingly vital in animal neuroscience and ethology. The shift from traditional “black box” methods to more open and intelligent approaches is driven by advances in deep learning-based pose estimation and tracking. These technologies enable the extraction of key points and their temporal relationships from sequence images. Such approach is particularly crucial for investigating animal behaviors in outer space, with microgravity, high radiation, and hypomagnetic field. However, the limited image data of space animal and the lack of publicly accessible datasets with ground truth annotations have hindered the development of effective evaluation tools and methods. To address this challenge, we present the SpaceAnimal Dataset—the first multi-task, expert-validated dataset for multi-animal behavior analysis in complex scenarios, including model organisms such as Caenorhabditis elegans, Drosophila, and zebrafish. Additionally, this paper provides evaluation code for deep learning models, establishing benchmarks to guide future research. This dataset will advance AI technology innovation in this field, contributing to the discovery of new behavior patterns in space animals.

Domain adaptation for classification has achieved significant progress in natural images but not in remote-sensing images due to huge differences in data-imaging mechanisms between different modalities and inconsistencies in class labels among existing datasets. More importantly, the lack of cross-domain benchmark datasets has become a major obstacle to the development of scene classification in multimodal remote-sensing images. In this paper, we present a cross-domain dataset of multimodal remote-sensing scene classification (MRSSC). The proposed MRSSC dataset contains 26,710 images of 7 typical scene categories with 4 distinct domains that are collected from Tiangong-2, a Chinese manned spacecraft. Based on this dataset, we evaluate several representative domain adaptation algorithms on three cross-domain tasks to build baselines for future research. The results demonstrate that the domain adaptation algorithm can reduce the differences in data distribution between different domains and improve the accuracy of the three tasks to varying degrees. Furthermore, MRSSC also achieved fairly results in three applications: cross-domain data annotation, weakly supervised object detection and data retrieval. This dataset is believed to stimulate innovative research ideas and methods in remote-sensing cross-domain scene classification and remote-sensing intelligent interpretation.

Near-nadir interferometric imaging SAR (Synthetic Aperture Radar) techniques are promising in measuring global water extent and surface height at fine spatial and temporal resolutions. The concept of near-nadir interferometric measurements was implemented in the experimental Interferometric Imaging Radar Altimeters (InIRA) mounted on Chinese Tian Gong 2 (TG-2) space laboratory. This study is focused on mapping the extent of high mountain lakes in the remote Qinghai–Tibet Plateau (QTP) areas using the InIRA observations. Theoretical simulations were first conducted to understand the scattering mechanisms under near-nadir observation geometry. It was found that water and surrounding land pixels are generally distinguishable depending on the degree of their difference in dielectric properties and surface roughness. The observed radar backscatter is also greatly influenced by incidence angles. A dynamic threshold method was then developed to detect water pixels based on the theoretical analysis and ancillary data. As assessed by the LandSat results, the overall classification accuracy is higher than 90%, though the classifications are affected by low backscatter possibly from very smooth water surface. The algorithms developed from this study can be extended to all InIRA land measurements and provide support for the similar space missions in the future.

In space environments, microgravity, high radiation, and weak magnetic fields induce behavioral alterations in animals, resulting in erratic movement patterns that complicate tracking. These challenges impede accurate behavioral analysis, especially in multi-object scenarios. To address this issue, this study proposes a deep learning-based multi-object tracking (MOT) framework specifically designed for space animals. The proposed method decouples appearance and motion features through dual-stream inputs and employs modality-specific encoders (MSEs), which are fused via a heterogeneous graph network to model cross-modal spatio-temporal relationships. Additionally, an object re-detection module is integrated to maintain identity continuity during occlusions or rapid movements. This approach is validated using public datasets of space-observed Drosophila and zebrafish, with experimental results demonstrating superior performance compared with existing tracking methods. This work highlights the potential of artificial intelligence as a valuable tool in behavioral studies, enabling reliable animal tracking and analysis under extreme space conditions and supporting future research in space life sciences.

Multi-animal pose estimation is a critical technique for enabling fine-grained quantification of group animal behaviors, which holds significant scientific value for uncovering behavioral changes under space environmental factors such as microgravity and radiation. Currently, the China Space Station has conducted a series of space biology experiments involving typical model organisms, including Caenorhabditis elegans (C. elegans), zebrafish, and Drosophila. However, substantial differences in species types, body scales, and posture dynamics among these animals pose serious challenges to the generalization and robustness of traditional pose estimation methods. To address this, we propose a novel, flexible, and general single-stage multi-animal pose estimation method. The method constructs species-specific pose group representations based on anatomical priors, incorporates a multi-scale feature-sampling module to integrate shallow and deep visual cues, and employs a structure-guided learning mechanism to enhance keypoint localization robustness under occlusion and overlap. We evaluate our method on the SpaceAnimal dataset—the first public benchmark for pose estimation and tracking of model organisms in space—containing multi-species samples from both spaceflight and ground-based experiments. Our method achieves AP scores of 72.8%, 62.1%, and 67.1% on C. elegans, zebrafish, and Drosophila, respectively, surpassing the state-of-the-art performance. These findings demonstrate the effectiveness and robustness of the proposed method across species and imaging conditions, offering strong technical support for on-orbit behavior modeling and large-scale quantitative analysis.