Research On Remote Sensing Image Processing With Machine Learning And Its Applications

With the development of remote sensing technology, the amount of acquired remote sensing images is growing explosively; also the dimension of corresponding spectral signatures is expanded. Labeling these massive data manually is tedious and time-consuming. So applying machine learning into auto-processing remote sensing images is a reasonable choice. The main research work including in the paper is about classification algorithm design and spectral feature selection.A domain adaptation method deals with problems where the marginal probability distribution in the target domain is different from but correlated to the one in the source domain. Meanwhile, classification tasks of both domains are the same. This phenomenon occurs frequently in classifications of remote sensing data, e.g., when data are collected in the same area but at different time (different weather conditions/different atmosphere circumstances) or when data are acquired by the same sensor with the same class label set but from different locations. Traditional machine learning algorithms cannot handle this problem in a satisfactory manner. In this paper, we propose a rationale input-output consistency where samples in the same cluster and defined by spectral signatures (input space) should always have the same class label (output space) if they are accurately classified. With this rationale, samples of high confidence (with high probability accurately classified) in the target domain are selected to redefine a new prediction function. Since two related domains can have different probability distributions, the data in the source domain which cannot follow the distribution of the target domain are deleted from the original training data set. Therefore, the proposed algorithm is denoted as input-consistent-output domain adaptation (ICODA) and works in an iterative way. After selection of highly-confident target domain samples and deletion of partial source domain data, a new training data set is used to retrain a new prediction model. The proposed ICODA algorithm was evaluated on two hyperspectral data sets-Botswana and KSC. Experimental results demonstrate much better classification accuracies when compared to a traditionally used supervised classifier.Meanwhile, Derivatives of spectral reflectance signatures can capture salient features of different land-cover classes. Such information has been used for supervised classification of remote sensing data along with spectral reflectances. In the paper, we study how supervised classification of hyperspectral remote sensing data can benefit from the use of additional derivatives of spectral reflectance without the aid of other techniques, such as dimensionality reduction and data fusion. An empirical conclusion is given based on a large amount of experimental evaluations carried out on three real hyperspectral remote sensing data sets. The experimental results show that when a training data set is of a small size or the quality of the data is poor, the use of additional first order derivatives along with original spectral features can significantly improve classification accuracies when using classifiers which can avoid the "curse of dimensionality," such as the SVM algorithm.Also, massive remote sensing data can be free download from the Internet. These data can be used for unsupervised feature learning via deep neural network. In this paper, we applied deep belief network on AVIRIS Gulf of Mexico oil spill data set for oil leak detection. The experimental classification results quite match with RGB images exported from corresponding bands.