VideoARD: An Analysis-Ready Multi-Level Data Model for Remote Sensing Video

Highlights

What are the main findings?

• A multi-level VideoARD model formalizes scene–object–event entities with standardized metadata and provenance.
• A spatiotemporal VideoCube links frame facts to spatial, temporal, product, quality, and semantic dimensions to enable OLAP-style queries and cross-sensor integration.

What is the implication of the main finding?

• The standardized representation streamlines workflows, reducing preprocessing burden and improving reproducibility for detection, tracking, and event analysis.
• Benchmarks show lower query latency and resource usage with non-inferior or slightly improved task accuracy versus frame-level baselines.

Abstract

Remote sensing video (RSV) provides continuous, high spatiotemporal earth observations that are increasingly important for environmental monitoring, disaster response, infrastructure inspection and urban management. Despite this potential, operational use of video streams is hindered by very large data volumes, heterogeneous acquisition platforms, inconsistent preprocessing practices, and the absence of standardized formats that deliver data ready for immediate analysis. These shortcomings force repeated low-level computation, complicate semantic extraction, and limit reproducibility and cross-sensor integration. This manuscript presents a principled multi-level analysis-ready data (ARD) model for remote sensing video, named VideoARD, along with VideoCube, a spatiotemporal management and query infrastructure that implements and operationalizes the model. VideoARD formalizes semantic abstraction at scene, object, and event levels and defines minimum and optimal readiness configurations for each level. The proposed pipeline applies stabilization, georeferencing, key frame selection, object detection, trajectory tracking, event inference, and entity materialization. VideoCube places the resulting entities into a five-dimensional structure indexed by spatial, temporal, product, quality, and semantic dimension, and supports earth observation OLAP-style operations to enable efficient slicing, aggregation, and drill down. Benchmark experiments and three application studies, covering vessel speed monitoring, wildfire detection, and near-real-time three-dimensional reconstruction, quantify system performance and operational utility. Results show that the proposed approach achieves multi-gigabyte-per-second ingestion under parallel feeds, sub-second scene retrieval for typical queries, and second-scale trajectory reconstruction for short tracks. Case studies demonstrate faster alert generation, improved detection consistency, and substantial reductions in preprocessing and manual selection work compared with on-demand baselines. The principal trade-off is an upfront cost for materialization and storage that becomes economical when queries are repeated or entities are reused. The contribution of this work lies in extending the analysis-ready data concept from static imagery to continuous video streams and in delivering a practical, scalable architecture that links semantic abstraction to high-performance spatiotemporal management, thereby improving responsiveness, reproducibility, and cross-sensor analysis for Earth observation.

1. Introduction

RSV refers to time-sequenced imagery captured by aerial, spaceborne, or ground-based sensors that preserves both spatial and temporal continuity, enabling the observation of dynamic phenomena and moving objects in the observed scene [1,2]. Unlike static remote sensing imagery, which provides only discrete snapshots, RSV allows analysis of temporal changes, object behavior, and event dynamics, making it particularly valuable for real-time monitoring, environmental surveillance, disaster response, and urban management [3,4,5].

With the rapid advancement of remote sensing technologies, RSV has become increasingly accessible and widely adopted. As illustrated in Figure 1, remote sensing videos offer continuous spatiotemporal capture, enabling the sub-second observation of moving targets. This capability provides new opportunities for automatic object recognition, behavior analysis, and event detection [6].

A widely used term in this context is Full Motion Video (FMV), which specifically denotes high-frame-rate video capturing continuous scene evolution, typically at 24–30 frames per-second or higher, and is extensively employed in applications such as object detection, target tracking, and dynamic activity analysis [7,8,9,10,11,12,13]. FMV can be considered a subset of RSV, emphasizing both the temporal resolution and real-time applicability of the data. Techniques developed for FMV processing are generally applicable to RSV, providing a foundation for advanced video analytics in remote sensing.

However, the massive volume, structural complexity, and diverse sources of video data significantly increase the cost of data management and processing. As shown in Figure 2, the absence of unified preprocessing and quality evaluation standards has left a large portion of these data underutilized, remaining as so-called “dark data” [14].

Against this backdrop, the concept of ARD offers a valuable framework for the efficient use of remote sensing imagery. ARD ensures that data meet minimum analytical requirements through standardized preprocessing steps—such as calibration, registration, and atmospheric correction—allowing users to perform immediate, downstream analysis without additional preparation [15,16,17,18]. While several organizations (e.g., USGS, ESA) have developed ARD standards for optical and radar imagery [19,20,21,22,23], there remains a lack of a systematic ARD framework for remote sensing video. As a result, preprocessing is often tedious, semantic information is difficult to extract, and cross-source data integration poses significant challenges.

To address these challenges, this paper introduces VideoARD, a multi-level semantic abstraction model for analysis-ready remote sensing video data. The model abstracts and represents content at three semantic levels: scene-level for ground feature distribution, object-level for individual object characteristics, and event-level for dynamic activity patterns. It further leverages metadata fields and the OGC Moving Features API [24,25] to support structured storage and efficient data access. To manage large-scale video datasets, we also propose VideoCube which is a spatiotemporal cube framework. It supports Earth Online Analysis Process (EOLAP) operations such as slice, trim, roll-up, and drill-down, enabling multidimensional, high-throughput query and analysis. The main contributions of this study are summarized as follows:

1.

Compared to directly applying algorithms on raw videos, VideoARD performs a one-time preprocessing step to generate reusable analysis-ready data, significantly reducing redundant computation in downstream tasks.

2.

The multi-level semantic abstraction enables accurate extraction of scene, object, and event information, improving detection and recognition performance.

3.

The spatiotemporal cube framework organizes data across varying resolutions and time scales, greatly enhancing querying and statistical efficiency.

4.

The effectiveness and scalability of VideoARD are validated through three representative case studies—vessel speed monitoring, forest fire detection, and 3D scene reconstruction—by comparing it with baseline methods in terms of latency, accuracy, and resource consumption.

The remainder of this paper is organized as follows. Section 2 reviews existing ARD standards and the current landscape of remote sensing video data management. Section 3 presents the proposed VideoARD model and the implementation of the VideoCube framework. Section 4 evaluates system performance through application scenarios and benchmark comparisons. Section 5 discusses limitations and outlines potential directions for future research. Section 6 concludes the paper.

2. Related Works

This section provides a critical review of foundational work in two key areas: (1) the standardization of ARD for earth observation and (2) management techniques for remote sensing video data. We identify existing gaps and challenges that underscore the need for the proposed VideoARD framework.

2.1. Definition and Standardization Progress of ARD

The concept of ARD was first introduced by the Committee on Earth Observation Satellites (CEOS) to standardize preprocessing workflows for earth observation data. The goal is to convert raw observational data into products ready for immediate analysis, thereby reducing users’ reliance on low-level calibration and correction steps. Typically, ARD products must meet strict criteria, including sub-pixel geometric accuracy, atmospheric correction to surface reflectance, cloud, and shadow masking, as well as standardized projection and resolution, enabling automated analysis across time, space, and multiple sensors.

Key developments in the international standardization of ARD are outlined below:

• CEOS Analysis-Ready Data for Land (CARD4L): Developed by the CEOS Working Group on Calibration and Validation (WGCV), CARD4L [26] defines ARD product specifications for land remote sensing, covering optical, infrared, and radar sensors. It imposes strict requirements on geometric accuracy, radiometric calibration, quality masking, and metadata to ensure comparability and interoperability across satellites and sensors [27].
• USGS Landsat Collection 2 ARD: The USGS released Collection 2 ARD [28] products for the Landsat series provide tiled-surface reflectance data embedded with SpatioTemporal Asset Catalog (STAC) metadata and quality masks, supporting cloud-based analysis and downloads.
• Sentinel-2 Level-2A: Although not explicitly labeled as ARD, the European Space Agency’s Sentinel-2 L2A products under the Copernicus program implement atmospheric correction, scene classification masks, and standardized projections, consistent with ARD principles. These products are distributed on cloud platforms such as CREODIAS in an ARD-compatible manner [18,29,30].
• Open Data Cube (ODC): The ODC initiative offers a data cube framework for ARD distribution and management, enabling the loading and cataloging of ARD products from diverse sources and bands [27]. It supports spatiotemporal linked retrieval and analysis, facilitating large-scale time series remote sensing research.

It can be observed that there are currently no ARD standards or products specifically designed for remote sensing videos. Therefore, specialized ARD extension standards are required to support temporal coherence and dynamic semantic abstraction in remote sensing video data. As summarized in

Table 1. Comparison of temporal, semantic, and preprocessing characteristics across representative ARD products and standards.

Product/Standard	Data Type	Provider	Temporal Support	Semantic Granularity	Video Preprocessing
Landsat C2 U.S.	Optical (30 m)	USGS	Multi-temporal (since 1982)	Pixel-level (reflectance, QA)	×
CARD4L	Optical, SAR	CEOS	Time-series compatible	Pixel-level (reflectance, backscatter)	×
CEOS-ARD *	Thematic (Land, Ocean)	CEOS	-	Thematic-level (domain classes)	Conceptual
Sentinel-2 L2A	Optical (10–60 m)	ESA	5-day revisit	Pixel-level (surface reflectance)	×
Sentinel-1 RTC	SAR (C-band, 20 m)	ESA	6-day revisit	Pixel-level (terrain-corrected backscatter)	×
Planet ARPS	Optical (3 m)	Planet Labs	Near-daily	Pixel-level (four-band reflectance)	×
Open Data Cube	Multi-source	ODC	Multi-temporal archive	Pixel-level (gridded data)	×
VideoARD (Ours)	Optical video	WHU	Sub-second continuous	Scene/object/event-level	Stabilization, SR

Notes: QA = quality assessment; SR = super resolution. The × symbol indicates that the corresponding ARD standard does not explicitly support video preprocessing. ^* CEOS-ARD was proposed in 2021 and is currently under public review. WHU = Wuhan University.

, we compare our proposed ARD framework with existing mainstream standards, products, and frameworks of ARD in terms of temporal handling, semantic abstraction, and preprocessing requirements. The results demonstrate the advantages of our approach in achieving higher temporal fidelity, richer semantic representation, and reduced preprocessing overhead.

2.2. Management and Analysis of Remote Sensing Video Data

Remote sensing video data exhibit unique challenges stemming from their temporal continuity and massive data volume. High-frame-rate sequences often contain substantial temporal redundancy, which increases storage and computational burdens when raw frames are directly processed [31,32,33]. Additionally, differences in spatial resolution, field of view, and revisit frequency among satellites, unmanned aerial vehicles (UAVs), and airborne sensors further complicate the integration and standardized management of multi-platform video sources [32,33]. Effective analysis in dynamic environments, e.g., wildfire propagation, traffic movement, or maritime surveillance, requires modeling temporal correlations across video frames, which is essential for capturing evolving spatiotemporal patterns [34].

Existing approaches generally fall into three categories. The first is academic prototype systems, typically built on libraries such as OpenCV and FFmpeg, which focus on fundamental operations like frame extraction or object detection. However, these systems often remain confined to isolated lab environments, lacking scalability and standardization. The second is commercial cloud platforms (e.g., AWS Ground Station [35], Google Earth Engine [36]), which have introduced distributed video analysis capabilities, including segmentation and tracking. These solutions, however, rely on closed APIs with limited transparency in terms of internal data models and semantic representations [37,38,39,40,41]. For the third category, recent research emphasizes multi-level semantic modeling by combining scene-level features with object-level motion to construct spatiotemporal graph structures. Studies in applications such as traffic monitoring and disaster prediction suggest that hierarchical semantic abstraction improves both analysis accuracy and computational efficiency [42,43,44,45,46,47].

A common strategy for semantic abstraction in remote sensing videos involves organizing content into three hierarchical levels: the scene level (environmental context), object level (detected and tracked entities), and event level (higher-order spatiotemporal activities). This framework supports structured annotation and enhances the indexing and retrieval of meaningful video content.

To support scalable organization and efficient querying of multi-source remote sensing video data, recent advances have incorporated data warehouse principles and OLAP techniques [34,48,49]. A typical approach involves the construction of spatiotemporal data cubes, where video frames are indexed by spatial and temporal dimensions. These cubes enable interactive querying through roll-up, drill-down, slicing, and dicing operations. Complementarily, spatiotemporal indexing techniques—such as R-tree [46] and quadtree [47] structures enhanced with temporal stamps—facilitate rapid access to scene-level content. In support of interoperability, interface standards like the OGC Moving Features API and STAC extensions are being explored for encoding semantic labels and geospatial metadata [50].

Although these efforts are still nascent, they lay the foundation for systems like VideoARD and VideoCube, which integrate multi-level semantics with scalable spatiotemporal models, and thus represent a promising direction for advancing video-based remote sensing analysis.

3. VideoARD Framework

3.1. VideoARD: A Multi-Level Data Model

The VideoARD framework is proposed to address the intrinsic complexity of remote sensing video data, which typically exhibit high spatial resolution, high temporal frequency, and a variety of acquisition conditions. Traditional video archives store raw or minimally processed frames, making analytical tasks inefficient due to repeated preprocessing, inconsistent metadata, and the absence of a unified semantic representation. The objective of the VideoARD model is to ensure that each video segment can be used directly for analytical tasks without additional preparation. This is achieved through a formalized multi-level semantic abstraction, covering scene-level, object-level, and event-level representations, each meeting defined minimum and optimal configurations of VideoARD.

3.1.1. Characteristics of Remote Sensing Video Data and ARD Readiness Criteria

As shown in Figure 3, remote sensing videos differ from conventional imagery primarily in their temporal continuity and the diversity of acquisition platforms such as satellites, UAVs, and surveillance and mobile terminals. These characteristics result in several technical challenges, including significant temporal redundancy, platform-dependent distortions, and heterogeneous spatial coverage.

In the context of ARD, readiness is determined by whether the data meet specific quality and completeness criteria. For remote sensing videos, we define a minimum configuration as the set of attributes required for basic analytical usability, such as consistent georeferencing, corrected radiometry, and essential metadata fields. The optimal configuration represents the highest level of readiness, including fine-grained semantic annotations, multi-temporal calibration, and harmonized sensor models. Only when each of the three semantic levels (scene, object, event) meets at least the minimum configuration can the video segment be classified as ARD-compliant.

3.1.2. Semantic Abstraction Framework

The core of the VideoARD model is the semantic abstraction process, which transforms raw spatiotemporal data into structured entities that are immediately analyzable.

$$V=〈F_1,F_2,F_3,\cdots ,F_T〉$$

(1)

Let a remote sensing video be denoted as Equation (1), where $F_T$ is the t-th frame, accompanied by its associated metadata $M_T$ (geolocation, timestamp, platform orientation, IMU/GPS data, sensor parameters, and so on).

We define three sets:

• S: Scene-level entities, each representing a spatial–temporal segment of the video.
• O: Object-level entities, each corresponding to a tracked physical object.
• E: Event-level entities, each denoting high-level semantic phenomena involving one or more objects and/or scenes.

According to the level of data preprocessing, remote sensing videos are divided into four analysis-ready levels, as shown in

Table 2. VideoARD data processing level.

Data Level	Product Name	Description
L0	Original video	Video products representing surface radiance values after geometric correction, radiometric correction, video stabilization, and super-resolution reconstruction, accompanied by essential auxiliary information
L1	Scene-level VideoARD	Processed video imagery mapped onto a unified temporal and spatial scale, ensuring integrity and consistency of variables, supplemented with scene-level information
L2	Object-level VideoARD	Extends L2 data by incorporating auxiliary information that describes object-level information within the video
L3	Event-level VideoARD	Builds upon L3 data by further adding auxiliary information that characterizes event-level information in the video

.

This process is formalized through three mapping functions:

• $f_S$: Mapping from the original video to scene entities.
• $f_O$: Mapping from the video to object entities.
• $f_E$: Mapping from scenes and objects to event entities.

Equation (2) defines the mappings that formalize the transitions between levels.

$$\left\{\begin{array}{cc}{f}_S:& V⟶S\\ f_O:& V⟶O\\ f_E:& (O,S)⟶E\end{array}\right.$$

(2)

As shown in

Table 3. Scene-level VideoARD metadata summary.

Metadata Parameter	Description	Minimum Configuration	Optimal Configuration
Video keyframe encoding	Video keyframe sequence number	Video keyframe ID	The ID of the video keyframe after thinning. The ID reflects the time of the video keyframe in the video.
Video spatial range	The spatial range of the surface covered by the video data	All video frames use a unified spatial range	Each video frame has a unique spatial range, which is manually verified
Video scene information	Scene information in the video	Type of static ground objects in the video	Manually verified background information in the video, consisting of object type, spatial range, and limit value
Scene information storage	Saves extracted scene information	Metadata fields	Scene information file

,

Table 4. Object-level VideoARD metadata summary.

Metadata Parameter	Description	Minimum Configuration	Optimal Configuration
Object detection	Detect objects in video frames	Object axis aligned bounding boxes	Object oriented bounding box
Object positioning	Extract the location of objects in video frames Objects’ relative position within the video frame	Objects’ actual surface coordinates
Object classification	Classify detected objects	Primary classification	Secondary classification with human interpretation assistance
Object information storage	Save extracted object information	Metadata fields	Object information files

and

Table 5. Event-level VideoARD metadata summary.

Metadata Parameter	Description	Minimum Configuration	Optimal Configuration
Object tracking	Tracks the same object in different video frames	Provides different object numbers	Eliminates interference, ensuring the same object has the same number, while different objects have different numbers
Trajectory tracking	Extracts the trajectory of a object within a video frame	A set of relative positions of the object within the video	Provides a set of real-world coordinate points corresponding to the time
Abnormal behavior limits	Scenario-specific abnormal behaviors	Default restricted behaviors	User-configured restricted behaviors
Abnormal behavior perception	Determines whether the object’s behavior in the video is abnormal	Detects default abnormal behaviors	Detects user-configured abnormal behaviors
Event information storage	Record the detected abnormal behavior information	Records default abnormal event information, including event type, time period, spatial scope	Associated objects, in metadata. Records complete abnormal event information, including event number, event type, time period, spatial scope, associated objects, object trajectory, and object status, in an abnormal event file

, we had summarized the minimum and optimal configurations for each level, including mandatory metadata fields, preprocessing steps, and semantic attributes.

As shown in Figure 4, at the scene level, abstraction aims to encapsulate spatial coverage, temporal span, quality assessment, and references to reduced-resolution previews for rapid inspection. At the object level, abstraction involves detecting, tracking, and describing individual moving entities, storing their trajectories in a temporally consistent and geospatially calibrated form. At the event level, abstraction captures higher-order patterns or interactions among objects and scenes, such as vessel clustering, wildfire spread, or large-scale human activities.

Figure 5 shows an example of scene/object/event information stored using the OGC moving features API. It can be observed that scene and object can be linked to event through the trajectory ID. Appendix A shows two real-world instances of moving features in the Jilin-1 satellite video.

By constructing geographic entities at three distinct level, namely scene, object, and event, the proposed approach maximally captures the spatiotemporal semantic information inherent in remote sensing video data and enables the extraction of user-relevant video content. Compared with traditional geographic information models that primarily focus on the data itself, this method enhances data usability, reduces the need for additional user processing, and introduces a novel retrieval dimension in the form of semantic-based video search, which can significantly improve retrieval efficiency.

The advantage of this framework lies in its modularity and composability. Each level can be processed and stored independently while remaining semantically linked, which supports both hierarchical querying and cross-level reasoning.

3.1.3. Multi-Level Model Design and Implementation

The multi-level structure of VideoARD is implemented through a combination of metadata schemas, preprocessing algorithms, and persistence mechanisms. Scene entities are stored with spatial and temporal indexes, object entities are indexed by trajectory identifiers and attributes, and event entities are stored with explicit references to the objects and scenes from which they were derived.

Formally, a scene s is represented as follows:

$$s=〈i{d}_s,A_s,T_s,E_s,C_s,Q_s,U_s〉$$

(3)

where $A_s$ is the spatial coverage, $T_s$ is the temporal range, $E_s$ is the spatial extent, $C_s$ is the category of the major background in the scene, $Q_s$ is the quality descriptor, and $U_s$ is a URI to a quicklook product.

An object o is represented as follows:

$$o=〈i{d}_o,C_o,T_o,{\tau }_o,A{t}_o〉$$

(4)

where $C_o$ is the class label, $T_o$ is the existence time range, ${\tau }_o$ is the trajectory, and $A{t}_o$ is a set of derived attributes.

An event e is represented as follows:

$$e=〈i{d}_e,typ{e}_e,T_e,geo{m}_e,P_e,con{f}_e〉$$

(5)

where $typ{e}_e$ is the event category, $T_e$ is the temporal range, $geo{m}_e$ is the spatial extent, $P_e$ is the set of participating objects, and $con{f}_e$ is the confidence score.

As shown in Figure 6, the scene categories are derived from the classification of surface features defined in the international standard ISO 19144-2 “Geographic Information—Land Cover Metadata [51]”. To ensure interoperability and semantic consistency with existing geographic information systems, this standard was restructured for the domain of remote sensing video analysis, with a focus on highlighting category distinctions frequently emphasized in academic research and application scenarios. After reorganization, the scene classification is arranged into eight primary categories and several secondary subcategories, in order to balance expressive power with indexing efficiency.

The object categories are based on the commonly used semantic system from the COCO dataset [52]. Considering the unique characteristics of remote sensing videos in terms of perspective, scale, and object representation, the original categories were merged and simplified, resulting in seven major categories and seventeen subcategories. This design preserves sufficient semantic granularity to support complex analyses while significantly reducing the engineering complexity of annotation and recognition.

The event categories were designed based on expert recommendations. In view of the cross-disciplinary expansion needs of VideoARD, the initial version of the framework defines only five basic event types, while also providing extensible interfaces that allow users or future work to incorporate additional domain-specific event types as needed. The complete list of scene, object, and event categories is provided in

Table 6. Scene, object, and event categories.

Level 1	Level 2	Level 3
Scene	Artificial surfaces	Building, Bridge, Transportation, Railway, Tunnel, Vegetation, City Furniture, Land Use, Relief, Road, City Object Group
	Cultivated Areas	Cultivated Areas
	Waterbodies	Ocean, Sea, Lake, River, Pond, Swamp
	Permanent Snow/Ice	Snow/Ice
	Forest Areas	Forest
	Grassland	Grassland
	Wetlands	Wetlands
	Bare Areas	Bare Areas
Object	Person	Person
	Vehicle	Bicycle, Car, Motocycle, Bus, Truck, Train
	Ship	Commercial Vessels, Military Vessels, Leisure Vessels, Special Purpose Vessels, Workboats, Support Vessels
	Airplane	Commercial Aircraft, Cargo Aircraft, Military Aircraft, General Aviation Aircraft
	Animal	Dog, Cat, Cow, Horse, Elephant, Sheep, Bear, Giraffe, Zebra, …
	Micro-Object	Sports Ball
	Others	Chair, Dining Table, TV, Bed, …
Event	Motion	Stationary, Constant Speed, Acceleration, Deceleration, Turning, Loitering, Abnormal Path
	Interaction	Meeting, Collision, Separation, Following, Parallel Movement
	State Change	Appearance Change, Disappearance, Emergence
	Environmental Monitoring	Wildfire, Flood, Landslide, Earthquake, Tsunami, …
	Statistical Aggregated	Object Count Threshold Reached, Abnormal Object Density, …

.

Overall, the adoption of standardized sources combined with application-oriented restructuring enables the three-layer semantic system to strike a balance among interoperability, feasibility, and scalability; the scene layer is used for rapid filtering and scheduling, the object layer for detection and semantic interpretation of trajectories, and the event layer for modeling high-level behaviors and processes.

This structured representation ensures reproducibility and traceability, as each entity retains references to its source data and the processing steps applied.

3.2. VideoCube: A Spatiotemporal Data Cube for VideoARD Management

The efficient storage, retrieval, and analysis of multi-level ARD entities require a dedicated spatiotemporal management structure. VideoCube is designed to fulfill this role by organizing scene, object, and event entities into a multidimensional array indexed by space, time, and semantic level.

3.2.1. Architecture and Implementation of VideoCube

Figure 7 depicts the architecture of the VideoCube, which partitions data along one fact, named Video Fact, and five primary dimensions: spatial, temporal, semantic, product, and quality. Each cell in the cube stores references to relevant ARD entities, along with aggregated statistics to support OLAP queries and analysis.

1.

Video Frame Fact: VideoARD tiles, generated by temporal segmentation and spatial gridding, serve as measures of the video frame fact, linked with temporal, spatial, product, and semantic dimensions.

2.

Temporal Dimension: To harmonize varying temporal resolutions, this dimension is designed with precision higher than the maximum among data sources and is linked to the frame fact.

3.

Spatial Dimension: Using a unified grid reference, heterogeneous spatial resolutions are standardized into a 2D grid, where each cell corresponds to a spatial member.

4.

Product Dimension: This records product metadata, including source, type, and processing level, and is linked to the frame fact.

5.

Quality Dimension: This captures scene-level quality indicators (e.g., scene classification, object detection accuracy, cloud masking) and is aggregated into the product dimension.

6.

Semantic Dimension: This encodes higher-level video semantics, including scenes, objects, and events, and is linked to the frame fact.

Spatial partitioning reduces the cost of spatial filtering, temporal slicing allows direct access to relevant intervals, and semantic partitioning ensures that queries can be restricted to the desired level of abstraction. Columnar storage formats are used for object and event attributes to accelerate aggregation queries.

In terms of physical design, VideoCube ensures high-performance storage and efficient management of VideoARD through its database architecture and storage structures, thereby laying the foundation for subsequent high-performance analytical services. A complete database design is provided in Appendix B.

3.2.2. EOLAP Operations on VideoCube

VideoCube supports a set of EOLAP query and analysis operations tailored to remote sensing video analytics.

(1) EOLAP Query Operations: As shown in

Table 7. EOLAP query operations on VideoCube.

Query Type	Functional Description
Slice	Selects a single member from a specified dimension, producing a sub-cube restricted to that member.
Trim	Selects multiple members from a specific dimension to form a subset of the data cube.
Roll-up	Aggregates measures along a dimension to produce a new cube with coarser granularity.
Drill-down	Performs the inverse of roll-up, retrieving data at finer granularity from an aggregated cube.

, EOLAP query operations mainly include the classical OLAP operations such as slice, trim, roll-up, and drill-down [53,54,55]. However, OLAP primarily targets business intelligence data cubes, while the proposed EOLAP extends these query operations to Earth observation data.

Figure 8 is an example of performing EOLAP query operations on VideoCube along the time dimension. It is obvious that new video cubes can be obtained through EOLAP query operation. Optimization strategies include the use of precomputed views for frequently requested aggregations and incremental decoding for drill-down operations, which minimizes the need to reprocess raw frames.

(2) EOLAP Analysis Operations: Traditional spatial OLAP (SOLAP) systems extend general OLAP to support multidimensional analysis of spatially referenced but static or slowly changing datasets, typically focusing on map-based exploration, aggregation, and visualization of spatial measures. In this work, we build on this idea and introduce the EOLAP analysis operations that are specifically tailored to RSV and high-temporal-resolution EO data. Beyond conventional SOLAP, EOLAP natively incorporates moving objects, events, and sensor/geometry metadata into the cube, and provides domain-specific operations such as trajectory-based slicing, event materialization, and frame-to-scene temporal aggregation. While general OLAP mainly supports statistical analysis and SOLAP adds geometric and scalar operations, neither can handle complex spatiotemporal and semantic analyses involving events and causal relationships. VideoCube overcomes this limitation by introducing advanced EOLAP-based semantic analysis operations for Earth observation data. One key function is trajectory join, which retrieves pairs of trajectories within a defined spatial or temporal window based on criteria such as intersection, proximity, or similarity. This allows the system to identify encounters, collisions, or coordinated formations. Another important function is event causality tracing, which leverages a three-level index of events, objects, and scenes to rapidly uncover the precursors of specific events. For instance, the system can link the occurrence of a fire to preceding variations in temperature, brightness, or object activity. Query execution and optimization are also integral to the framework. An EOLAP query first employs VideoCube metadata to determine appropriate voxel resolutions and materialized views. Attribute-based filtering is then applied in columnar storage, with trajectory indexes enabling partial decoding. For complex event inference, the system can adopt incremental decoding, processing only the key frames associated with relevant trajectories, which substantially reduces computational costs. To maintain interactive performance, frequently used complex queries may be accelerated through materialized templates, asynchronous aggregation, and multi-level caching.

The VideoARD framework and its VideoCube management structure provide a theoretically grounded and practically efficient approach to organizing remote sensing video data for immediate analysis. The multi-level abstraction ensures that analytical tasks can target the appropriate semantic granularity without redundant preprocessing. The spatiotemporal data cube model enables scalable retrieval and OLAP-style analysis, while maintaining traceability through explicit references between data entities and their sources.

4. Experimental Evaluation

This section presents a systematic evaluation of the VideoARD framework with respect to remote sensing video organization, analysis efficiency, and semantic representation capabilities. The evaluation is conducted through quantitative comparative experiments and validated using real-world application scenarios.

4.1. Efficiency Benchmarking

The efficiency of data access and query processing is a critical factor in evaluating the practical value of the proposed VideoARD model and its companion spatiotemporal management framework, VideoCube. In real-world remote sensing applications, data analysis often needs to be performed over long temporal sequences, heterogeneous sources, and under multi-user concurrent workloads. Therefore, the primary objective of this experiment is to determine whether the proposed method can sustain low latency and high throughput under large-scale workloads while maintaining analysis readiness.

4.1.1. Experimental Environment and Dataset

The experimental evaluation was conducted on a distributed computing platform specifically optimized for large-scale remote sensing video processing. The hardware configuration, summarized in

Table 8. Experimental hardware environment.

Node Role	Operating System	CPU	Memory	Storage
Master	CentOS 7.9.2009	Intel Xeon CPU E5-2692 v2 @ 2.20 GHz	94 GB	PB-scale
Worker #0	CentOS 7.9.2009	Intel Xeon CPU E5-2692 v2 @ 2.20 GHz	128 GB
Worker #1	CentOS 7.9.2009	Intel Xeon CPU E5-2692 v2 @ 2.20 GHz	128 GB
Worker #2	CentOS 7.9.2009	Intel Xeon CPU E5-2692 v2 @ 2.20 GHz	128 GB

, consists of one master node and three worker nodes, each equipped with high-performance multi-core CPUs and large-capacity memory to ensure the computational throughput required for near-real-time analysis. The system adopts a petabyte-scale storage infrastructure to accommodate the continuous ingestion and long-term retention of high-resolution video streams.

The software environment, summarized in

Table 9. Experimental software environment.

Software	Version	Purpose
MySQL	5.5.62	Storage and retrieval of transactional data from the legacy system
PostgreSQL	9.4.4	Storage and retrieval of remote sensing video metadata
Hadoop	2.7.4	Distributed file system for large-scale data storage
HBase	1.4.13	Distributed storage for structured scene-level and object-level data
MinIO	8.3.3	Distributed object storage for unstructured video data
JDK	1.8.0_131	Java runtime environment
Tomcat	8.5.57	Application server for system service deployment
Apache	2.4.10	Web server for system front-end delivery
FFmpeg	6.0	Video preprocessing
OpenCV	4.8.0.76	Object detection and tracking
YOLO	v8	Object detection, classification, and tracking in video streams

, integrates relational databases, spatial extensions, distributed file systems, and object storage frameworks. PostgreSQL with the PostGIS extension enables efficient management of spatial metadata associated with remote sensing video, supporting fine-grained geographic queries on frame- and object-level metadata. Hadoop Distributed File System (HDFS) and HBase are employed for distributed indexing of structured video frame and object information, providing high-throughput retrieval capabilities. For unstructured video files, the system adopts a distributed object storage approach using MinIO, allowing for flexible and scalable management of heterogeneous video data formats. This hybrid storage strategy ensures a balanced optimization between query latency for structured metadata and sustained I/O throughput for raw video content retrieval.

This integrated architecture enables seamless interoperability between spatial databases and distributed storage backends. The combination of PostgreSQL/PostGIS with HDFS and HBase significantly reduces query latency for spatiotemporal retrieval operations while maintaining scalability for high-throughput ingestion. The object storage layer provided by MinIO ensures that large video files are managed independently of their metadata, which allows for optimized caching and load balancing strategies.

The datasets used in the experiments are shown in

Table 10. Experimental datasets.

Datasets	Content
VideoARD#1	21 Jilin-1 satellite videos and 5 Shuangqing-1 satellite videos
VideoARD#2	50.4 GB of UAV fire videos and 42.5 GB of UAV forest videos

. They consist of multiple satellite and drone videos. The data sources used in the examples below are all from the datasets in Table 10.

4.1.2. Benchmarking Methodology

Traditional remote sensing video data services are typically built on distributed file systems and object storage. To demonstrate the feasibility and efficiency of the approach proposed in this study, we conduct a comprehensive evaluation covering three aspects: experimental design, experimental results, and result analysis. The comparison focuses on performance in terms of query time, data throughput, and resource utilization between our method and traditional approaches.

The experiments are conducted using a controlled comparison design, with details of the experimental plan, evaluation metrics, datasets, and procedures described as follows.

(1) Experimental Groups: The experimental group applies the proposed service method (Group A), while the control group (Group B) uses a video data service based on the HDFS combined with the MinIO object storage service.

(2) Evaluation Metrics:

• Query time (ms): The time difference between a user request and the system’s data response.
• Data throughput (Mbps): The maximum data volume the system can process.
• Resource utilization: CPU consumption.

(3) Experimental Data: The datasets and experimental environment used in this section are described in Section 4.1.1.

(4) Experimental Procedure: A load-testing tool (JMeter) is employed to simulate concurrent access by multiple users, with four levels of concurrency: 10, 50, 100, and 500 users. The experiments are carried out in two steps: data query and object tracking. For each concurrency level, query time, data throughput, and resource utilization are measured and recorded.

4.1.3. Results and Observations

After conducting the benchmark tests, the experimental results are shown in Figure 9.

As illustrated in Figure 9a, the response time of the experimental group increased from 200 ms to 600 ms as the number of concurrent users rose, whereas the control group’s response time increased from 250 ms to 1200 ms. The control group exhibited both higher latency and a faster growth rate in response time compared with the experimental group. Figure 9b shows that as concurrency increased, the experimental group’s throughput initially rose from 90 Mbps to 1000 Mbps and then stabilized, while the control group’s throughput decreased from 85 Mbps to 50 Mbps. Thus, the experimental method demonstrated higher and more stable throughput. Figure 9c indicates that CPU utilization increased for both groups with higher concurrency, but the experimental group showed slower and more stable growth, whereas the control group’s CPU usage rose faster and was less stable.

The results indicate that the proposed service method is more efficient than traditional approaches. Several factors, summarized in Table 11, likely contribute to this improvement:

1.

Enhanced data modeling: The proposed method incorporates a multi-dimensional hierarchical structure on top of distributed storage, making it better suited for complex analytical tasks. In contrast, object storage methods focus primarily on block-based storage and redundancy, without supporting complex data-dimensional models.

2.

Structured and fine-grained management: The proposed method extends data across spatial, temporal, and semantic dimensions, facilitating rapid queries and supporting complex applications. Object storage, however, is oriented toward unstructured data management and lacks support for complex queries.

3.

High concurrency support: By using distributed architecture and multi-dimensional data partitioning, the proposed method supports parallel processing and high-concurrency access. The VideoCube design leverages load balancing and parallel mechanisms to maintain low latency under heavy load. While traditional file systems and object storage can handle concurrency, their performance can degrade for complex queries or large-scale concurrent access.

4.

Optimized query efficiency: The proposed method emphasizes efficient multi-dimensional query and analysis through the indexing, caching, and pre-processing of key frames. Traditional methods primarily optimize for data transfer speed and storage scalability rather than complex query performance.

5.

Application-specific design: The method is tailored to complex remote sensing video scenarios, enabling multi-dimensional queries and rapid analytics. Object storage solutions are better suited for general-purpose tasks such as video-on-demand or data backup, but are less effective for real-time analysis or multi-dimensional retrieval.

Table 11. Comparison of experimental groups.

Comparison Dimension	Group A (Ours)	Group B
Architecture	Multi-level analytical model combining spatiotemporal cubes and semantic dimensions; supports complex queries and analytics	Distributed file system or object storage architecture; emphasizes storage and data transfer efficiency
Data Management	Multi-level management (scene, object, event); structured and fine-grained	Simple file-block and metadata-based management
Concurrency Support	Multi-dimensional partitioning and parallel processing; suitable for high-concurrency, multi-dimensional queries	High concurrency capability, but limited support for complex queries
Performance Optimization	Preprocessing key frames and semantic data; caching and indexing improve access speed	Relies on block storage and redundancy; low efficiency for complex queries
Application Scenarios	Multi-user remote sensing video analysis, dynamic analytics, event detection	Simple queries and storage, e.g., video-on-demand, backup

Table 11. Comparison of experimental groups.

Comparison Dimension	Group A (Ours)	Group B
Architecture	Multi-level analytical model combining spatiotemporal cubes and semantic dimensions; supports complex queries and analytics	Distributed file system or object storage architecture; emphasizes storage and data transfer efficiency
Data Management	Multi-level management (scene, object, event); structured and fine-grained	Simple file-block and metadata-based management
Concurrency Support	Multi-dimensional partitioning and parallel processing; suitable for high-concurrency, multi-dimensional queries	High concurrency capability, but limited support for complex queries
Performance Optimization	Preprocessing key frames and semantic data; caching and indexing improve access speed	Relies on block storage and redundancy; low efficiency for complex queries
Application Scenarios	Multi-user remote sensing video analysis, dynamic analytics, event detection	Simple queries and storage, e.g., video-on-demand, backup

In summary, compared with traditional distributed file system and object storage approaches, the proposed method offers superior data management and complex query support, enabling efficient multi-dimensional, dynamic, and concurrent data analysis. Traditional object storage is better suited for general-purpose video storage, providing scalability but limited analytical capability.

The results demonstrate that the proposed approach achieves multi-gigabyte-per-second ingestion under parallel feeds, sub-second scene retrieval for typical queries and second-scale trajectory reconstruction for short tracks.

4.2. Application Cases Studies

Having established the superior system-level performance of VideoARD through controlled benchmarks, we now evaluate its practical utility through three real-world application scenarios. These cases span different semantic requirements and operational constraints, thereby exercising the scene-level, object-level, and event-level functionalities of the proposed framework. Each case begins with a concise statement of the operational objective, specifies the input modality and the VARD configuration used, describes the processing chain, and then reports system-level performance and practical observations.

4.2.1. Case 1: Vessel Speed Monitoring (Object-Level Case)

(1) Objective: The operational objective is continuous monitoring of vessel speed within a port area, supporting near-real-time detection of abnormal motion and automated alerting. This application primarily exercises the object-level semantics of VideoARD while requiring scene-level consistency to ensure geolocation correctness.

(2) Data and VideoARD Configuration: As shown in Figure 10, the case uses Jilin 1 satellite videos covering Victoria Harbour, Hong Kong. In Figure 10a, the satellite video data are from the Jilin-1 satellite, captured on 24 September 2023, at 13:00, covering the Victoria Harbour area in Hong Kong. The video spans from 13:33:05 to 13:34:35 on the same day, with a spatial extent from 114.131018°E to 114.175469°E in longitude and from 22.283304°N to 22.305214°N in latitude. Figure 10b shows Google satellite imagery updated in September 2023, covering a spatial extent from 114.078582°E to 114.229129°E in longitude and from 22.255089°N to 22.339743°N in latitude.

The object-level entities are required to satisfy the minimum configuration for basic monitoring. The preprocessing pipeline enforces spatial and temporal registration, video stabilization, per-frame georeferencing, and key frame extraction. Object detection and classification is performed on preprocessed key frames, followed by multi-frame association and trajectory smoothing.

(3) Processing and Results: A dedicated cross-video small object tracking algorithm, based on an improved YOLOv8 [56] and DeepSORT [57], was employed to address challenges such as motion blur and occlusion common in satellite videos, as detailed in Figure 11. The algorithm integrates a fine-tuned YOLOv8 with SAM [58] for enhanced shape representation of small-scale objects and incorporates a cross-video object retrieval module for object re-identification across videos, significantly improving tracking accuracy while reducing identity switches.

System-level timing for a typical monitoring episode is summarized in

Table 12. Time spent at each stage of Case 1.

Stage	Duration(s)	Variance (±)	Description
Cube Creation	32	1	Including preprocessing and storage
Data Query	3.2	0.25	Query and index reading
Event Creation	15	5	Create the event
Alert Push	3.0	0.25	Triggering and push

, and the visualization results are shown in Figure 12. The results demonstrate that once trajectories are materialized in VideoCube, query times for operational alarms become lightweight and deterministic.

(4) Observation: Comparison to the traditional baseline indicates that VideoARD reduces the time taken to retrieve object trajectories and generate alerts by one order of magnitude in high-reuse scenarios. The principal reason for this is that the baseline recomputes detection and association on demand, whereas VideoARD reuses materialized object entities.

4.2.2. Case 2: Wildfire Detection and Monitoring (Event-Level Case)

(1) Objective: The objective is early detection and continuous tracking of fire fronts and smoke plumes across heterogeneous platforms. This application stresses scene-level semantics because accurate estimation of burn extent and propagation depends on consistent scene quality metadata and on multi-frame spatial coherence.

(2) Data and VideoARD Configuration: Input data include both UAV video and satellite video (Himawari 8) over a forested test site in Australia. As shown in Figure 13a, the Himawari-8 satellite video data cover the period from 1 December 2019 to 4 January 2020, with a spatial extent from 112.131671°E to 152.649245°E in longitude and from 37.498355°S to 10.948692°S in latitude. Figure 13b presents UAV surveillance video data collected over the New South Wales State Conservation Area between 26 December and 30 December 2019, with a spatial extent from 145.664346°E to 145.893686°E in longitude and from 33.082984°S to 32.942492°S in latitude. Figure 13c shows Google satellite imagery updated in September 2023, covering a spatial extent from 112.131671°E to 152.649245°E in longitude and from 37.498355°S to 10.948692°S in latitude.

(3) Processing and Results: The preprocessing pipeline emphasizes stabilization, radiometric normalization, and cloud and smoke masking. Fire detection leverages a frame-level classifier (DEnet [59]) tuned to smoke and high-temperature signatures. The visualization results are shown in Figure 14, and Figure 15 presents the sequence of frames showing detected smoke masks and the temporal expansion of the burnt area.

Representative timings for the wildfire monitoring workflow are reported in

Table 13. Time spent at each stage of Case 2.

Stage	Duration(s)	Variance (±)	Description
Cube Creation	10	1	Including preprocessing and storage
Data Query	3.2	0.25	Query and index reading
Event Creation	8	5	Create the event
Alert Push	2.9	0.2	Triggering and push

.

(4) Observation: The experiment demonstrates that the scene materialization time is modest. Once scene-level overviews and per-frame masks are generated, area aggregation and temporal trend queries are executed rapidly. The VideoARD approach reduces the time required for repeated queries such as cumulative burned area estimation, which are common during active fire operations.

4.2.3. Case 3: Near-Real-Time 3D Scene Reconstruction (Scene-Level Case)

(1) Objective: The objective is to produce dense three-dimensional models of an urban segment from a collection of UAV video streams in near real time. This application primarily exercises scene-level indexing combined with selective extraction of multi-view frame sets that satisfy geometric diversity and overlap constraints.

(2) Data and VideoARD Configuration: The input data includes both UAV videos and Google satellite imagery. As shown in Figure 16a, the remote sensing video data were acquired by a UAV over the Qilihe District of Lanzhou City on 4 February 2023. The video spans from 13:12:05 to 13:32:01 local time on the same day, covering a spatial extent from 103.708586°E to 103.768924°E in longitude and from 36.071303°N to 36.098563°N in latitude. Figure 16b presents Google satellite imagery updated in September 2023, covering a spatial extent from 103.570312°E to 103.939384°E in longitude and from 36.015922°N to 36.159638°N in latitude.

The key contribution of VideoARD is that the cube provides fast access to scene candidates and to meta-information that identifies frames with complementary viewpoints and sufficient baseline for photogrammetric reconstruction.

(3) Processing and Results: The preprocessing stage generates scene-level summaries that include frame geolocation, view vectors, and coarse image quality indicators. A view selection step then queries VideoCube to return a candidate set of frames. The subsequent dense reconstruction uses the Danesfield algorithm [60], an end-to-end system for extracting 3D models from multi-view imagery, which includes semantic segmentation of buildings and texture mapping. The visualization results are shown in Figure 17, and Figure 18 shows the results of the 3D scene reconstruction.

Representative timings for the near-real-time 3D scene reconstruction workflow are reported in

Table 14. Time spent at each stage of Case 3.

Stage	Duration(s)	Variance (±)	Description
Cube Creation	10	1	Including preprocessing and storage
Data Query	3.5	0.25	Query and index reading
Event Creation	305	150	Create the event
Alert Push	910	1	Triggering and push

.

(4) Observation: VideoARD reduces the pre-modeling search effort significantly. In conventional practice, large volumes of frames must be manually filtered or exhaustively considered for matching. With VideoCube, the retrieval of geometrically useful frames is an indexed operation, which reduces human effort and I/O overhead. The dominant cost remains the photogrammetric processing, which is inherently compute intensive.

4.2.4. Case-Level Summary and Practical Remarks

Across the three cases, a consistent pattern is observed. Scene and object materialization introduce an upfront cost that is justified when queries are frequent, when multi-user access is expected, or when downstream analytics reuse the same semantic entities for different tasks.

As shown in Figure 19, stacked bar charts showing cube creation, data querying, event creation, and alert pushing durations for the vessel speed monitoring, wildfire detection and monitoring, and near-real-time 3D reconstruction cases.

4.2.5. Accuracy Consistency Evaluation

Our objective prioritizes efficiency while maintaining task accuracy. We therefore adopt a non-inferiority protocol with an absolute margin $\mathrm{\Delta }=1\%$ against strong baselines using the same algorithm. A small, stratified subset in each case is manually annotated to isolate pipeline effects from algorithm changes. The tested tasks were conducted under the same experimental conditions as described in Section 4.1.

To evaluate the accuracy of the proposed method, we use the following metrics:

IDSW: An IDSW occurs when a tracker wrongfully swaps object identities or when a track was lost and reinitialised with a different identity. Formally, an IDSW is a True Positive (TP) which has a predicted ID (prID) that is different from the prID of the previous TP, which has the same ground truth ID (gtID). IDSWs only measure association errors compared to the single previous TP, and do not count errors where the same prID swaps to a different gtID (ID Transfer).

MOTA: Matching is performed at a detection level. Bijective (one-to-one) mapping is constructed between predict detections (prDets) and ground truth detections (gtDets) in each frame. Any prDets and gtDets that are matched (correct predictions) become TPs. Any remaining prDets that are not matched (extra predictions) become False Positives (FPs). Any gtDets that are not matched (missing predictions) become False Negatives (FNs). prDets and gtDets can only be matched if they are adequately spatially similar. MOTA measures three types of tracking errors: the detection errors of FNs and FPs, as well as the association error of IDSW. The final MOTA score is calculated by summing up the total number of these errors, dividing by this result by the number of gtDets, and subtracting from one.

$$MOTA=1-{\displaystyle \frac{\left|FN\right|+\left|FP\right|+\left|IDSW\right|}{gtDet}}$$

(6)

IDF1: Unlike MOTA, which matches at a detection level, IDF1 calculates a bijective (one-to-one) mapping between the sets of ground truth trajectories (gtTrajs) and predicted trajectories (prTrajs). It defines new types of detection matches. Identity True Positives (IDTPs) are matches on the overlapping part of trajectories that are matched together. Identity False Negatives (IDFNs) and Identity False Positives (IDFPs) are the remaining gtDets and prDets, respectively, from both non-overlapping sections of matched trajectories, and from the remaining trajectories that are not matched.

$$IDF1={\displaystyle \frac{\left|IDTP\right|}{\left|IDTP\right|+0.5\ast \left|IDFN\right|+0.5\ast \left|IDFP\right|}}$$

(7)

Precision/Recall: We use $\mathrm{IoU}\ge =0.5$:

$$\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}},\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(8)

Z-RMSE: After alignment and correspondence, measure only the height error:

$$\begin{array}{cc}\hfill \mathrm{Z}\text{-}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}& =\sqrt{\frac{1}{N}\sum _{i=1}^N{\left(\mathrm{\Delta }{z}_i\right)}^2}\hfill \\ \hfill \mathrm{\Delta }{z}_i& =z_i^{\mathrm{est}}-z_i^{\mathrm{ref}}\hfill \end{array}$$

(9)

where N is the number of matched points; i indexes a matched pair; $z_i^{\mathrm{est}}$ and $z_i^{\mathrm{ref}}$ are the estimated and reference heights; and $\mathrm{\Delta }{z}_i$ is their height residual.

H-RMSE: Measure only the planar (X–Y) error:

$$\begin{array}{cc}\hfill \mathrm{H}\text{-}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}& =\sqrt{\frac{1}{N}\sum _{i=1}^N\left[{\left(\mathrm{\Delta }{x}_i\right)}^2+{\left(\mathrm{\Delta }{y}_i\right)}^2\right]}\hfill \\ \hfill \mathrm{\Delta }{x}_i& =x_i^{\mathrm{est}}-x_i^{\mathrm{ref}}\hfill \\ \hfill \mathrm{\Delta }{y}_i& =y_i^{\mathrm{est}}-y_i^{\mathrm{ref}}\hfill \end{array}$$

(10)

where $x_i^{\mathrm{est}},y_i^{\mathrm{est}}$ and $x_i^{\mathrm{ref}},y_i^{\mathrm{ref}}$ are the estimated and reference planar coordinates; $\mathrm{\Delta }{x}_i,\mathrm{\Delta }{y}_i$ are the planar residuals; and N and i are as above.

Table 15. Accuracy comparison between the baseline pipeline and VideoARD.

Case	Algorithms	Metric	Group A	Group B	$\mathbf{\Delta }$
#1	YOLOv8+DeepSORT	MOTA	59.6	58.7	+0.9
		IDF1	64.9	64.1	+0.8
#2	DEnet	Precision	85.3	84.5	+0.8
		Recall	83.4	82.7	+0.7
#3	Danesfield	Z-RMSE	1.29	1.32	−0.03
		H-RMSE	1.91	2.06	−0.15

Notes: Comparison conducted under a non-inferiority margin of $\mathrm{\Delta }=1\%$ (absolute). For percentage metrics (MOTA, IDF1, Precision, Recall, F1), $\mathrm{\Delta }$ is in percentage points (pp). For Z-/H-RMSE, $\mathrm{\Delta }$ is in meters (negative is better).

summarizes the accuracy comparison. The detailed experimental designs for Group A (VideoARD) and Group B (baseline) are presented in Table 11. Within the non-inferiority margin ($\mathrm{\Delta }=1\%$) for percentage metrics, Group A maintains accuracy and shows small positive deltas across object tracking and wildfire tasks. For 3D reconstruction, Group A achieves slightly improved vertical error (Z-RMSE: $-0.03$ m) and horizontal error (H-RMSE: $-0.15$ m), indicating enhanced planimetric consistency. Combined with substantially lower query latency, steadier data throughput, and reduced resource utilization (Section 4.1.3), these results confirm that the proposed data model and pipeline deliver significant efficiency gains without degrading, and in some cases improving, analytical quality.

5. Discussion

This section synthesizes the experimental findings and identifies avenues for methodological refinement and future research. The discussion addresses limitations discovered in the experiments, the generalizability of the VideoARD approach, and targeted directions where advances are most likely to increase practical utility.

Scalability and incremental update mechanisms: The experiments show that the initial cost of scene and semantic materialization is the primary engineering burden. To serve continuously evolving data streams, a robust incremental update mechanism is necessary, one that can update voxel-level aggregates and semantic indices without full cube reconstruction. Research efforts should focus on fine-grained update policies, change-detection-driven refresh schedules, and bounded staleness models that allow controlled freshness trade-offs.

Edge-assisted preprocessing and bandwidth management: The majority of per-scene computational cost arises from video stabilization, geometric rectification, and object detection. Pushing parts of the preprocessing pipeline toward edge nodes or sensor-side processors reduces transport volume and central computational load. Future work should quantify how edge-assisted processing affects overall accuracy, latency, and cost when integrated with VideoCube. Additionally, adaptive compression schemes guided by semantic importance can substantially reduce bandwidth consumption while preserving analysis fidelity.

Adaptive materialization policies: The decision to materialize a scene or an object should be informed by predicted query likelihood, task criticality, and storage constraints. Developing a predictive cost model and an associated policy engine will improve resource utilization. Such a model requires historical query logs, semantic popularity profiles, and cost estimates for recomputation versus storage. Machine-learned policies that incorporate these inputs can outperform static heuristics.

Cross-sensor fusion and provenance: The value of VideoARD increases when multiple sensors with complementary characteristics are integrated. Fusion across satellite video, UAV video, and ground cameras improves detection robustness and reduces occlusion induced uncertainty. Key scientific challenges include time alignment of heterogeneous streams, resolution harmonization, and provenance tracking that preserves lineage for every semantic entity. Transparent provenance representation is essential for downstream scientific interpretation and for auditability in operational contexts.

Semantic enrichment and uncertainty quantification: Event detection often relies on thresholding or deterministic rules, which may lack robustness in ambiguous conditions. Augmenting event entities with uncertainty measures and supporting the probabilistic fusion of evidence will make VideoARD outputs more useful for decision making. Research into calibrated confidence estimates for detectors and trackers, and the propagation of such uncertainty through aggregate queries, is an important direction.

User-centric interfaces and interactive workflows: The experiments show that VideoCube accelerates data preparation and enables interactive analysis. To maximize adoption, it is necessary to develop user interfaces that expose semantic entity versioning, lineage, and quality indicators. Interactive workflows that allow analysts to request additional refinement on demand, for example raising a minimum configuration object to best configuration upon request, will increase operational flexibility.

6. Conclusions

This study introduces VideoARD, a multi-level analysis-ready data model for remote sensing video, and VideoCube, a spatiotemporal organization and query infrastructure that operationalizes the model. The central idea is to transform high-volume video streams into semantically rich entities at scene, object, and event levels and to apply those entities in a multidimensional structure optimized for Earth observation analytics. The experimental evaluation demonstrates that this approach delivers substantial reductions in query latency and substantial improvements in concurrent throughput when compared to a baseline that performs repeated frame-level processing.

From an engineering perspective, the principal contribution is a practical architecture that balances preprocessing cost against downstream responsiveness. The experimental cases show that the initial materialization cost becomes economical as soon as queries are repeated, as soon as multiple users share the same dataset, or as soon as multiple analytics reuse the created semantic entities. From a scientific perspective, the contribution is an explicit formalization of multi-level readiness, which clarifies what metadata and semantic attributes are necessary for reliable analysis at each layer.

Limitations remain. The upfront cost of materialization and the storage overhead of persisting semantic entities require careful operational planning. Additional research is required to automate incremental updates and to integrate edge-level preprocessing without loss of provenance. Uncertainty quantification and cross-sensor fusion are promising directions that will broaden the impact of VideoARD for complex monitoring and response tasks.

In closing, VideoARD and VideoCube provide a feasible path toward practical, scalable, and reproducible remote sensing video analytics. The architecture is sufficiently modular to incorporate new detection algorithms, to adopt alternative storage back ends, and to integrate automated policy engines for materialization. Future work that refines incremental update logic, extends cross-sensor fusion capabilities, and formalizes uncertainty propagation will increase both the scientific value and the operational readiness of video-based Earth observation systems.