Label Semantics and Image Features aware Remote Sensing Sample Retrieval from Multi-Source Datasets for AI-enabled Remote Sensing Monitoring

XingTao Ren ^1,2 Yan Ma ^1* Yixin Zhou ^1,2

• ¹ Aerospace Information Research Institute, Chinese Academy of Sciences (CAS), Beijing, China
• ² University of Chinese Academy of Sciences, Beijing, Beijing, China

With the advent of remote sensing (RS) big data, deep learning (DL) has shown remarkable potential in uncovering intricate features and detecting environmental changes from vast RS imagery, particularly for intelligent environmental monitoring, such as coastal land cover classification and urbanization analysis. Accordingly, the remote sensing imagery sample datasets have become crucial in ensuring robust training models with satisfactory performance across various AI(Artificial Intelligence)-enabled remote sensing applications. However, the demand for abundant and diverse RS imagery samples has surged, leading to data scarcity as a critical challenge for large-scale AI-enabled RS applications. Moreover, the inconsistencies in labeling semantic categories among sample datasets, coupled with the limits of the commonly used single-label representation of datasets, have posed a huge barrier to fully leveraging the training samples across diverse remote sensing sample datasets. Additionally, existing RS datasets are dispersed across platforms with supplier-specific data structures, making it difficult to gather suitable samples for training. To address these challenges, we propose an intelligent RS sample retrieval approach that integrates label semantics and visual features for AI-enabled RS monitoring. It measures similarity distances between samples and label category clusters to dynamically build tailored datasets, overcoming single-label limits and label disparities. Moreover, it also establishes a dynamic RS label category system that is capable of dynamically expanding distinct categories from new sample datasets through label semantic similarity mapping to resolve label inconsistencies across sample datasets. Supported by a virtual distributed storage system, it enables in-memory sample discovery and integration across clouds, leveraging multi-source RS datasets with limited interoperability. Finally, the comparative performance experiments have also confirmed the effectiveness and efficiency of this approach.