NR-IQA for UAV hyperspectral image based on distortion constructing, feature screening, and machine learning
Wenzhong Tian,Arturo Sanchez-Azofeifa,Za Kan,Qingzhan Zhao,Guoshun Zhang,Yuzhen Wu,Kai Jiang
DOI: https://doi.org/10.1016/j.jag.2024.104130
IF: 7.5
2024-09-11
International Journal of Applied Earth Observation and Geoinformation
Abstract:Assessing the quality of UAV-HSIs (Unmanned aerial vehicle hyperspectral images) is crucial for evaluating sensor performance, identifying distortion types, and measuring data inversion accuracy. Due to the absence of reference images, UAV-HSI quality assessment leans towards no-reference image quality assessment (NR-IQA), offering versatile applications. NR-IQA methods of remote sensing images using machine learning techniques have emerged, however, NR-IQA methods for UAV-HSIs containing multi-type and multiple distortions have not been developed. This paper introduces an NR-IQA method for UAV-HSI, employing machine learning techniques. We summarize and simulate distortion types in UAV-HSIs, constructing a quality assessment dataset based on 23 original high-quality and 806 simulated degraded UAV-HSIs. Extracting 129 features encompassing texture, color, transform domain, structural, and statistical aspects, we form seven feature sets through random and filtered feature selection algorithms. Ten machine learning quality assessment models are trained using this dataset and feature sets. The results showed that the model with the highest evaluation accuracy was extra trees (ET) ( R 2 = 0.928, RMSE = 0.326, RPD = 3.601), using feature set 1 that fuses Tamura texture, color, wavelet transform, and mean subtracted contrast normalized (MSCN) coefficient for a total of 11 features, the PLCC and SROCC of its predicted and true quality scores reached 0.963 and 0.925, respectively. In addition, the random forest (RF), gradient boosting decision tree (GBDT), generalized regression neural network (GRNN), and extreme learning machine (ELM) also had high evaluation accuracies ( R 2 > 0.9 and RPD > 2.5). These findings underscore the applicability of our proposed machine learning-based NR-IQA method to assess the quality of the UAV-HSIs containing noise, blur, strip noise, and multiple distortions. Additionally, this study serves as a reference for selecting features and models for other hyperspectral image quality assessments.
remote sensing