Elastic Deep Neural Network for Adaptive Feature Extraction in Remote Sensing Image Retrieval
TL;DR
An Elastic deep neural network for adaptive feature extraction in satellite image retrieval that dynamically adjusts node activation, connection scaling, network width, and network depth to generate discriminative feature representations while maintaining computational efficiency.
Abstract
The rapid growth of large-scale remote sensing image archives has increased the demand for accurate and efficient Content-Based Image Retrieval (CBIR) systems. Existing deep learning approaches commonly employ fixed feature extraction architectures, which may not effectively capture the diverse visual characteristics of aerial scenes. To address this limitation, this paper presents an Elastic deep neural network (EDNN) for adaptive feature extraction in satellite image retrieval. This network dynamically adjusts node activation, connection scaling, network width, and network depth to generate discriminative feature representations while maintaining computational efficiency. The proposed framework is validated on the publicly available Aerial Image Dataset (AID) using widely adopted retrieval metrics. It achieves an F1-score of 97.3%, a Recall of 97.5%, an Average Normalized Modified Retrieval Rank (ANMRR) of 0.052, mean Average Precision (mAP) of 97.0%, and Precision of 97.2%, while utilizing only 3.24 million trainable parameters and requiring an average retrieval time of 18 ms per image. The experimental findings indicate that the proposed EDNN delivers accurate and computationally efficient remote sensing image retrieval, demonstrating its potential for large-scale Earth observation repositories.