Metric Learning Methods For Hyperspectral Remote Sensing Imagery

With the characteristic of the high spectral resolution and continuous spectral curve,hyperspectral images(HSI)contain large number of spectral bands for each pixel,and convey abundant spectral information,which can be used to distinguish subtle spectral differences even between the similar materials.Therefore,it has unique advantages to detect and classify certain materials with hyperspectral data.Although it has been found that there are many methods and technologies in HSI have made some progress based on the application and practice,the limitations of the existing methods and the characteristics of the HSI lead to several problems are still not well solved.Thus,this thesis aims to develop metric learning methods for hyperspectral remote sensing imagery based on the latest methods in machine learning and pattern recognition area.In particular,we adopt metric learning technology as the mainline and then combine random forest and ensemble learning into our algorithms.(1)This dissertation aims to investigate the intrinsic characteristic of the HSI,and summarize the challenges in the field of hyperspectral remote sensing analysis.The status quo and existing problems of this dissetrtation have been introduced in the first place.Besides,possible breakthroughs and innovations by combining the metric learning methods with machine learning area have been described.Then,this dissertation reviews the development history of metric learning methods,and mainly introduces the advances of metric learning methods for hyperspectral remote sensing analysis.(2)The current target detectors for HSI depend on the specific statistical hypothesis test,and the detectors may only perform well under certain conditions.How to develop a proper metric to measure the separability between targets and backgrounds becomes the key issue for target detection.This dissertation proposes an efficient maximum margin metric learning(MMML)method for hyperspectral target detection.MMML algorithm can obtain metric matrix to map the original data into the metric feature subspace without specific statistical hypothesis test,especially when the number of target samples are very small or the targets are difficult to detect.Finally,with the help of MMML algorithm,we can maximally separate the target samples from the background ones.(3)Hyperspectral target detection has problem that the number of target pixels and background pixels is imbalanced,and the number of target samples is very limited.Most of the existing metric learning methods are based on a single metric and cannot directly handle multi-feature representations.This dissertation proposes random forest metric learning(RFML)method,which use random forests as the underlying representation of the metric learning for detecting the desired targets.RFML method needs fewer adjusted parameters,and allows a more flexible and more scalable metric than the Mahalanobis-based methods to deal with limited numbers of target samples or situations where the targets are difficult to detect.Besides,RFML method combines the efficiency of a single-metric method and the accuracy of multiple metrics through considering both the relative position and the absolute pairwise position,which adds additional feature spaces to guarantee that the method can adapt to the heterogeneous distribution of high-dimensional data,and differentiates target pixels from background pixels to the greatest degree.(4)The high-dimensional data space generated by hyperspectral sensors introduces challenges for the conventional data analysis techniques.Popular dimensionality reduction techniques usually assume a Gaussian distribution and are sensitive to the parameters,which may not be in accordance with real life and cannot separate between-class samples with high similarity from the local area.We utilize locally adaptive decision constraints for the labeled training samples per class to considere the locality of data distribution and make decision based on a threshold and the changes between the distances before and after metric learning.Thus,this dissertation presents the locally adaptive dimensionality reduction metric learning(LADRml)method for guaranteeing the distance between different classes to the greatest degree.(5)In the high-dimensional feature space,the between-class samples lying in a neighborhood area have a higher similarity and different samples having different effectiveness,which leads to useful data existing primarily in a subspace and therefore may be misclassified.Traditional dimensionality reduction methods are likely to obtain a good performance only under certain conditions.Moreover,most of existing methods can make use of limited spectral information,which is not suitable for the real conditions.Classical metric learning methods usually construct a global transformation and have problems handling high-dimensional data,adjusting parameters,and working well with a large number of training samples.Considering that global metric learning method is not appropriate for all training samples,this dissertation proposes an ensemble discriminative local metric learning(EDLML)method for HSI classification to learn the local discriminative distance metrics from weighted training samples and the relative neighbors,and then these local metrics are then aligned in a global framework via ensemble learning.By applying EDLML algorithm,we can effectively maximize the between-class distances while minimizing the within-class distances,which further prove that metric learning is effective.