Real-time remote sensing for sudden surface anomalies: a review of principles and challenges

Sudden surface anomalies (SSAs)—ranging from earthquakes and floods to wildfires and industrial accidents—pose escalating threats to ecosystems and societies worldwide. With its advantages of wide coverage, high revisit frequency, and rich information content, remote sensing is regarded as an effective tool for rapid detection and early warning of such anomalies. However, significant technical bottlenecks remain in using remote sensing to monitor SSAs, such as delays due to revisit intervals and complex data processing chains and challenges for the integration and joint analysis of multi-source information, resulting in frequent occurrences of missed detections and false alarms. Real-time remote sensing has emerged as a transformative approach for monitoring and early warning of such abrupt events, driven by recent advances in satellite constellations, on-orbit artificial intelligence, and multi-source data fusion. In a review article recently published in Space: Science & Technology, a team led by Qiao Wang, academician of the Chinese Academy of Engineering, from Beijing Normal University systematically synthesized the global landscape of sudden surface anomalies, elucidated their spatiotemporal patterns and underlying drivers, and critically assessed the capabilities and limitations of current remote sensing technologies for rapid detection and assessment, charting the course for future research and system development in real-time remote sensing for sudden surface anomaly monitoring and early warning.

First, the current status of real-time remote sensing for monitoring sudden surface anomalies are analyzed. Using the global disaster event database EM-DAT (Emergency Events Database), established and maintained by the Disaster Epidemiology Research Center at KU Leuven, Belgium, authors systematically compiled spatial distribution and frequency data for various common natural disasters between 2000 and 2024. Based on this, a geographic distribution map of global disaster events (Figure 1) was created. Overall, the broad distribution of global disaster events underscores the complex interplay between climate change and Earth’s dynamic internal processes, further emphasizing the urgent need to strengthen real-time remote sensing monitoring and early warning systems. Remote sensing possesses different response characteristics for various types of sudden anomaly events (as shown in Table 1), presenting increasingly prominent potential in monitoring natural disasters and human-induced emergencies. However, traditional remote sensing data acquisition and processing which relies heavily on manual operations and high-performance computing resources suffers efficiency limitations in emergency response. With the growing complexity and scale of anomaly detection, researchers are increasingly turning to advanced artificial intelligence approaches. According to publication data from the Web of Science core database, cutting-edge technologies, such as on-orbit intelligence, edge computing, and on-orbit processing, are gradually being integrated into the remote sensing research framework. The increasing application of artificial intelligence and machine learning technologies for real-time analysis, on-board processing, and multi-source data fusion on remote sensing platforms is becoming a key driving force for technological innovation and academic progress in this field.

Then, key technical challenges are identified.

(1) real-time atmospheric correction.

Traditional ground-based post-processing methods that rely on physical models can no longer meet the real-time and flexibility requirements. To address this challenge, researchers have begun exploring the potential of machine learning techniques for on-orbit atmospheric correction. These machine learning techniques such as deep regression or statistical simulation methods can automatically fit complex atmospheric-surface coupling relationships, making them particularly suitable for big data environments and offering the potential for near-real-time processing on the satellite platform. However, despite significant progress, these emerging methods still face several challenges. On one hand, model training heavily depends on calibration data, while actual on-orbit data is limited by bandwidth and heterogeneity, making it difficult to obtain. On the other hand, deep models have limited transferability between different satellite platforms and sensors, often requiring retraining for specific platforms. Additionally, the constraints of on-orbit computing resources (such as power consumption, storage space, and radiation environments) also hinder the deployment and operation of complex networks. Overall, continued in-depth research and technological breakthroughs are needed across multiple dimensions, including algorithm efficiency, data acquisition, and model generalization.

(2) model deployment under severe on-orbit resource constraints.

The high computational complexity and the vast number of parameters in deep models have severely constrained their real-time deployment and stable operation on resource-limited platforms such as satellites and edge devices. As a result, model lightweighting techniques have become a core enabler for engineering and real-time intelligent analysis in remote sensing, aiming to balance inference efficiency and accuracy under limited hardware conditions, thereby meeting the practical demands of timely and large-scale remote sensing data processing. Mainstream lightweighting approaches primarily include depthwise separable convolutions, structured pruning, and numerical quantization. Depthwise separable convolutions, which decompose conventional convolution operations into spatial domain depthwise convolutions and channel-wise pointwise convolutions, effectively reduce both model parameters and computational complexity. Structured pruning strategies further compress model size by selectively removing redundant channels, convolutional kernels, or layers, thus facilitating efficient deployment on embedded and satellite-based hardware. In recent years, pruning methods are often combined with attention mechanisms and feature importance evaluation to ensure the retention of critical feature representation capabilities. Numerical quantization focuses on reducing the bit-width of parameters and activations. Common techniques such as 8-bit integer quantization and mixed-precision quantization enable models to fully leverage the high throughput and low bandwidth characteristics of embedded AI chips. In response to the engineering constraints of remote sensing platforms, mainstream on-orbit computing solutions include graphics processing units (GPU), vision processing units (VPU), field-programmable gate arrays (FPGA), and AI-specific chips. To meet the real-time requirements of remote sensing platforms, on-orbit and edge computing hardware has evolved from reconfigurable parallel pipelines toward general-purpose deep learning acceleration while balancing millisecond-level latency, stringent power budgets, and operational robustness. As shown in Figure 2, the landscape can be ordered by computational capability and generation from lower to higher levels.

(3) robust multi-hazard identification across heterogeneous sensor platforms.

Through the collaborative interaction of on-orbit preprocessing, satellite data downlink, and ground-based anomaly analysis on early warning platforms, the star-ground collaborative processing mechanism, with real-time response services at its core, is gradually becoming the mainstream development direction for remote sensing data acquisition and processing. Despite positive progress in specific disaster scenarios, real-time intelligent processing of on-orbit remote sensing still faces numerous technical bottlenecks. First, most existing methods are optimized for a single disaster type or specific indicators, lacking a universal detection framework suitable for multiple disaster types and diverse scenarios. Second, there is a lack of unified response thresholds and composite indices for different types of surface anomalies. Furthermore, in terms of sensor performance, various remote sensing methods still have limitations, restricting the integration of all-weather, full-spectrum monitoring.

In summary, the integration of machine learning and deep learning technologies into real-time remote sensing systems has become a growing trend. To achieve truly efficient and accurate surface anomaly intelligent perception, future research must focus on building a unified system for cross-scenario, multi-sensor, and multi-task fusion.

Finally, the blueprint of next-generation integrated real-time remote sensing systems for anomaly monitoring is further discussed. To institutionalize the sequence of anomaly detection, initial localization, fine detection, and information transmission, this paper proposes an immediate remote-sensing constellation, as shown in Figure 3, that serves as the backbone for orchestration and in-space relay. A communications satellite functions as the relay and tasking hub, with laser intersatellite links binding multispectral, hyperspectral, thermal infrared, SAR and agile satellites into a cooperative network. The BeiDou constellation supplies precise orbit determination and time transfer, establishing a common spatiotemporal frame for coordinated geometry. Agile satellites carry out emergency retasking and targeted revisits, and high-resolution hyperspectral assets deliver localized imaging and spectral feature extraction. On-board processing converts results into anomaly semantics or compact packets that are delivered via intersatellite relay to the ground Telemetry, Tracking and Command system (TT&C) or sent directly to mobile devices through the BeiDou short-message service. On the ground, data are consolidated by the TT&C system and synchronized to a cloud platform for cataloging, storage, sharing and service orchestration. A management system maintains model versioning and parameter uplinks, composes observation plans and commands, and keeps two-way synchronization with the cloud so that space and ground share a consistent policy and knowledge base. User devices access the cloud to subscribe to and review reports and diagnostics, forming a closed loop that spans on-orbit computation, relay transport, BeiDou direct alert, TT&C and cloud governance, and end-user services, enabling minute-level disaster response and real-time early warning.