Nonlinear Feature Normalization for Hyperspectral Feature Transfer
Hyperspectral remote sensing is an important topic for deriving high-level information about the earth's surface. Applications include land cover mapping, precision farming, and the detection of environmental pollution. This is made possible by recording and evaluating narrow-band features that are characteristic of individual materials. External effects, however, lead to nonlinearities in the data and complicate data analysis. These effects include changes in illumination, hard and partial shadows, as well as transmission / multiple reflections by objects in the scene, and anisotropic effects for 3D objects. Correcting these effects is required for robust data analysis. In particular, when comparing multiple data sets a unified representation is required. Physically motivated models for correcting atmospheric in uences are generally used for the pre-processing of hyperspectral data. However, these models do not consider local variations, such as shadows and object geometry. Therefore, this thesis deals with data-driven approaches in the field of Manifold Alignment (MA) and Feature Transfer (FT) to transfer several data sets to a common system. Previous research on these topics has focused primarily on learning the underlying geometry of high-dimensional data and aligning multiple datasets by determining the minimum discrepancy while preserving the individual data structure. Usually, a common domain with very high dimensionality is chosen to facilitate the alignment. The transformation into another domain, however, prevents physical interpretability. Also, inversion of one data set from the common domain to the domain of a target data set is diffcult due to the pre-image problem. The contributions of this thesis can be divided into two categories. The Nonlinear Feature Normalization (NFN) is a data-driven approach to mitigate nonlinear effects in hyperspectral data. NFN is a supervised method and requires training samples for each class in the scene. A new basis for data representation is defined, consisting of one spectral reference signature per class. The training data are then used to individually shift all samples towards the new basis. This significantly reduces the effects of nonlinearities, as shown by comparing classification results before and after the NFN transformation. The NFN is then used to derive the Nonlinear Feature Normalization for Data Alignment (NFNalign). NFNalign transforms multiple data sets to the same basis in the common domain and then applies an inverse transformation to transfer data sets from the common domain to a domain of another data set. Since the dimensionality of the data is not changed during the transformation, it is possible to perform the inversion analytically. The functionality of NFNalign is demonstrated by transforming hyperspectral radiance data to reflection data. Thereby, the pre-processing step of the atmospheric correction can be replaced, shadows and other nonlinearities are corrected, and characteristic features of the spectral signatures are transferred. The quality of the alignment is demonstrated by applying an SVM model trained on a reference data set to the aligned data set. Additional alignment is assessed by applying a classification model trained on a reference data set to a test data set after it has been transformed to the domain of the reference with NFNalign. Further experiments investigate the robustness with regard to noise and errors in the training data as well as the alignment of data with different dimensions. Also, a comparison with common reference methods is performed. Overall, NFN and NFNalign provide a complete framework for hyperspectral data alignment and FT.
Stuttgart, Univ., Diss., 2019