A Three-Layer AI-Driven Image Filtering for Efficient LEO Satellite Remote Sensing

Low-earth orbit (LEO) satellites are considered essential tools for remote sensing (RS) and Earth observation (EO) applications thanks to their high spatial resolution images. However, limited on-board resources (storage, downlink throughput, and bandwidth) make it challenging to manage the large volume of data collected. Real-time data management is needed to avoid wasting on-board compute and downlink budget, especially when dealing with a significant influx of anomalous data images resulting from factors such as redundancy, seasonal changes, sensor properties, and satellite orbit dynamics. This paper introduces a novel three-layer filtering solution designed to optimize satellite image transmission for EO. Incorporating a fusion of machine learning and deep learning methodologies, our data management solution mitigates critical challenges associated with LEO satellite resource constraints. The proposed solution includes a no-reference deep image quality assessment (NoR-DIQA) model using a convolutional neural network architecture to identify and filter out distorted images. Then, it employs a perceptual hashing redundancy detection approach to eliminate duplicated images, and finally, it employs a classification model that categorizes suitable RS and EO applications for each image set before transmission. This framework effectively maximizes data utility and optimizes resource allocation for RS and EO satellites. The proposed framework was tested on different RS and EO datasets, where each layer was benchmarked against established models and consistently achieved superior results (e.g., higher correlation and lower error in image quality assessment, improved redundancy detection, and classification accuracy up to 99.4%), demonstrating the robustness and reliability of the three-layer filtering solution.