We present a proof-of-concept toolkit focused on remote sensing data fusion literature that turns research articles into searchable, ontology-backed records and a lightweight knowledge graph. The system ofers an interactive, Gradio-based interface featuring searchable cards and a simple subgraph preview. Uncertainty is incorporated as basic, human-editable tags. A small, binary text-classification component (BERT/RoBERTa) supports triage of abstracts into “fusion-related” vs. “other remote sensing” to aid curation. We demonstrate the end-to-end pipeline and provide preliminary classifier metrics on a small labeled set, emphasizing the system's scope as a practical starting point. Code, ontology, and configuration are released for reproducibility. This work supports the AI-for-science goal of developing transparent, interpretable, and uncertainty-aware AI tools for scientific knowledge integration.
Data fusion—the integration of heterogeneous data sources into a unified representation—is critical
for enhancing coverage, resolution, and interpretability across diverse domains. A widely adopted
structure for organizing fused data is the knowledge graph, which represents entities and their semantic
relationships as nodes and edges, enabling queryable, interconnected representations [
Building on these developments, we present a lightweight, end-to-end toolkit that transforms remote sensing data fusion literature into searchable records and an ontology-backed knowledge graph with basic uncertainty annotations. Implemented via a Gradio based interface 1, the system supports search, graph previews, and triage using a binary text classifier. Specifically, we ofer: • An ontology-driven schema linking papers, datasets, and fusion methods in the remote sensing domain. • A lightweight binary classifier (BERT/RoBERTa) for abstract-level triage. • Basic uncertainty tagging to support interpretability.
• A simple, reproducible interface for search and visualization, with released artifacts for extension. The 2nd International Workshop on Artificial Intelligence for the Science of Science (AI4SciSci 2025), co-located with ACM/IEEE-CS
JCDL 2025 CEUR Workshop
ISSN1613-0073
In the context of digital libraries, such lightweight AI tools can accelerate domain curation while preserving human oversight. By emphasizing interpretability and incremental uncertainty annotation, our approach aligns with AI4SciSci’s vision of transparent and reproducible AI systems for scientific knowledge integration.
A modular data pipeline was designed to integrate structured data handling, semantic enrichment, machine learning, and lightweight visualization. Figure 1 illustrates the high-level architecture of the proposed toolkit. The architecture is organized into five layers: data ingestion, storage, semantic enrichment, AI-based classification , and user interaction. This layered design supports flexible querying, ontology-based linkage, and editable uncertainty tagging across datasets and fusion methods. At this stage, uncertainty metadata are qualitative and manually assigned rather than formally modeled or propagated computationally.
The pipeline begins with data ingestion, where tabular datasets (CSV and Excel) describing research
papers, data fusion techniques, and datasets are pre-processed using pandas and stored in a PostgreSQL
database. These structured data are exported to an RDF-compatible schema through an OWL ontology
layer implemented with OWLReady22, enabling semantic relationships between papers, datasets, fusion
methods, and associated uncertainty tags [
To assist document triage, a small BERT-based classifier [
An OWL ontology was designed to provide semantic structure for literature-level representation of
remote sensing data fusion. Rather than aiming for comprehensive domain coverage, the ontology was
intentionally kept lightweight and application-driven, with the goal of supporting search, linking, and
human-editable uncertainty tagging within a small curated corpus. This design was informed by prior
work on scholarly knowledge graphs [
The core classes include FusionMethod, Dataset, and Uncertainty, with supporting classes such
as Paper and Publisher. These are connected through properties including integrates, usesData,
isPublishedIn, and hasUncertainty. A numeric attribute, hasConfidenceLevel, allows storage
of confidence-related indicators extracted from text or assigned by curators. Uncertainty modeling
was motivated by existing research on scientific uncertainty detection and integrated modeling [
The ontology was implemented using the owlready2 library and serialized in RDF format for integration with the knowledge graph pipeline. Structural consistency was verified using the Owlready2 reasoner to ensure valid class and property definitions. In addition, a subset of entities and relations was manually reviewed by domain experts to confirm semantic correctness. No external ontology alignment or quantitative quality metrics were applied in this prototype. Figure 2 illustrates the ontology structure and its key relationships.
Two domain-specific knowledge graphs were constructed to support exploration of data fusion and
remote sensing literature [
A graph representation was then generated using the NetworkX library. Nodes correspond to entities defined in the ontology (e.g., Paper, Dataset, FusionMethod), while directed edges encode semantic relations such as integrates and usesData. Only valid entity links were instantiated to preserve referential integrity.
Each graph instance includes optional uncertainty annotations inherited from ontology metadata. 2https://owlready2.readthedocs.io/ 3https://networkx.org 4https://www.gradio.app
Graphs were exported in GraphML format and rendered for interactive inspection using pyvis 5. In
parallel, the relational backend supports keyword- and attribute-based queries. Separate graphs were
maintained for data fusion and general remote sensing to align with the binary classifier’s scope and
reduce semantic drift between domains [
Although most literature tagging in the system was performed manually, machine learning was
incorporated to assist abstract-level triage by identifying whether a paper is related to data fusion or general
remote sensing. Following human-supervised knowledge curation practices [
We fine-tuned a pre-trained BERT model [
A total of 120 abstracts were independently annotated by two domain experts, achieving strong
agreement (Cohen’s = 0.86 ) [
Predictions with confidence below 0.6 were flagged and reviewed by human annotators through the Gradio interface, enabling iterative correction and refinement. Although no classical baselines (e.g., TF–IDF with SVM) were included in this prototype, future work will incorporate comparative evaluations and larger-scale validation.
The user interface was implemented using Gradio, chosen for its simplicity and flexibility in building interactive web applications. The interface provides four main views: Search, Graph, Classify, and Ontology. Users can query the system by research topic, fusion method, or dataset name, and results are presented as styled cards with DOI links, metadata, abstracts, and embedded uncertainty tags.
Uncertainty is visualized using color-coded highlights and tooltips to improve interpretability. The ontology view presents a structured outline of the knowledge model, while the graph view renders ifltered subgraphs for interactive exploration. The classification tab allows users to submit abstracts and receive automated predictions regarding their relevance to data fusion. Figure 3 shows an overview of the interface.
Overall, the interface serves as a lightweight digital library front end [
The evaluation of the developed system was conducted across multiple dimensions, including search accuracy, classification reliability, interface responsiveness, and overall system robustness. The preliminary results demonstrate that the integrated architecture performs efectively in delivering dynamic, context-aware domain insights with associated uncertainties.
The search engine embedded in the user interface mitigates common user query issues through exact
and partial matching via Python’s difflib.get_close_matches() function, misspelling corrections,
etc. Test cases were designed to validate core functionalities such as searching by paper title, dataset
name, and fusion method using related keywords. In all tested scenarios, the system returned card-styled
results with enriched metadata, abstracts, and DOI references that were deemed relevant by expert
evaluators. Across 20 representative test queries spanning fusion methods, datasets, and publication
titles, the system achieved full accuracy on exact matches and an average relevance of 92% for partial
matches, as assessed by two domain experts. Each retrieved record displayed associated uncertainty
tags [
To assess the efectiveness of transformer-based models for abstract triage, both BERT-Base-Uncased and RoBERTa-Base were evaluated on the labeled abstract set. Table 1 summarizes the performance results.
The fine-tuned BERT model achieved an accuracy of 0.966 and an F1 score of 0.963 on the validation subset, outperforming the RoBERTa variant. Precision was perfect for the fusion class, while recall was slightly lower, indicating conservative classification behavior. These results suggest that transformerbased classifiers are efective for abstract-level domain filtering even with limited data.
Classification outputs were linked to ontology instances such as FusionMethod and Dataset, enabling automated tagging during knowledge graph construction. Low-confidence predictions were routed to human annotators via the interface for confirmation or correction, preventing classification errors from propagating into the graph.
Because the dataset is small and manually curated, results should be interpreted as evidence of feasibility rather than generalization performance. Nonetheless, the findings confirm that lightweight ifne-tuning can meaningfully support expert-driven literature organization.
The interface was evaluated for usability and performance. Pagination mechanisms were tested by injecting simulated datasets to ensure stable loading and filtering. Visual transitions, layout adjustments, resizing, and mobile responsiveness were validated. Functional buttons such as “View Fusion Possibilities” and “View Data Fusion Recommendations” successfully directed users to relevant contextual sections, activated embedded logic, and met anticipated use case scenarios. A small pilot evaluation with three research assistants (familiar with remote sensing data) confirmed that the color-coded uncertainty indicators improved interpretability, allowing users to distinguish between confirmed and tentative links in the graph. Participants highlighted the system’s transparency and ease of use, though they also noted the need for richer visual encoding of uncertainty in future iterations. Figure 4 shows a sample knowledge graph query highlighting multiple aspects of uncertainty within search results.
This work demonstrates the feasibility of a lightweight toolkit that integrates ontology-based representation, uncertainty tagging, and human-in-the-loop machine learning for organizing remote sensing data fusion literature. The system prioritizes interpretability and usability over full automation, supporting expert-guided exploration rather than autonomous knowledge extraction.
The ontology was intentionally designed as a minimal, application-driven schema to enable linking,
ifltering, and navigation across papers, datasets, and fusion methods. Structural consistency was verified
using a reasoner, and concepts were refined through iterative curation. Uncertainty is represented as
qualitative metadata to support interpretability rather than formal inference, consistent with prior work
on uncertainty modeling in scientific knowledge systems [
The classifier serves as a triage mechanism, with expert validation of low-confidence predictions to prevent error propagation. While evaluation is limited in scale, the results indicate that lightweight ifne-tuning can efectively support expert workflows.
Limitations include the small dataset, absence of large-scale benchmarking, and lack of automated uncertainty reasoning. These reflect the prototype nature of the system and motivate future work on ontology alignment, uncertainty modeling, and scaling.
This paper presented a Gradio-based toolkit for curating remote sensing data fusion literature into searchable, ontology-backed records with editable uncertainty annotations. A lightweight transformerbased classifier supports abstract-level triage, with expert validation integrated to preserve reliability and interpretability.
Future work will focus on expanding ontology coverage, introducing uncertainty provenance tracking, and incorporating baseline model comparisons. Additional directions include structured user evaluations, multi-label classification, and scaling the framework to larger literature collections. Interoperability with external knowledge sources will also be explored.
Artifacts (code, ontology, and configuration) are available at https://doi.org/10.6084/m9.figshare. 29914361 to support reproducibility and community engagement. We extend our sincere gratitude to Timothy Mugambi, Nancy Odhiambo, and Braiden Betway for their invaluable contributions to this research project. This work was supported by seed funding from the College of Computing at Grand Valley State University.
During the preparation of this manuscript, the author(s) used Generative AI tools for limited purposes such as grammar and language refinement. All technical content, interpretation of results, and conclusions were created by the author(s). The author(s) reviewed and edited the output as needed and take(s) full responsibility for the accuracy, originality, and integrity of the work in compliance with the CEUR Generative AI policy.