Convolutional neural networks (CNNs) have been shown to be a powerful tool for image classification. Recently, they have been adopted into the remote sensing community with applications in material classification from hyperspectral images. However, CNNs are time-consuming to train and often require large amounts of labeled training data. The widespread use of CNNs in the image processing and computer vision communities has been facilitated by the networks that have already been trained on large amounts of data. These pretrained networks can be used to initialize networks for new tasks. This transfer of knowledge makes it far less time-consuming to train a new classifier and reduces the need for a large labeled data set. This concept of transfer learning has not yet been fully explored by those using CNNs to train material classifiers from hyperspectral data. This paper provides an insight into training hyperspectral CNN classifiers by transferring knowledge from well labeled data sets to data sets that are less well labeled. It is shown that these CNNs can transfer between completely different domains and sensing platforms, and still improve classification performance. The application of this work is in the training of material classifiers of data acquired from field-based platforms, by transferring knowledge from publicly accessible airborne data sets. Factors, such as training set size, CNN architectures, and the impact of filter width and wavelength interval, are studied.

This paper proposes novel autoencoders for unsupervised feature-learning from hyperspectral data. Hyperspectral data typically have many dimensions and a significant amount of variability such that many data points are required to represent the distribution of the data. This poses challenges for higher-level algorithms which use the hyperspectral data (e.g., those that map the environment). Feature-learning mitigates this by projecting the data into a lower-dimensional space where the important information is either preserved or enhanced. In many applications, the amount of labelled hyperspectral data that can be acquired is limited. Hence, there is a need for feature-learning algorithms to be unsupervised. This work proposes unsupervised techniques that incorporate spectral measures from the remote-sensing literature into the objective functions of autoencoder feature learners. The proposed techniques are evaluated on the separability of their feature spaces as well as on their application as features for a clustering task, where they are compared against other unsupervised feature-learning approaches on several different datasets. The results show that autoencoders using spectral measures outperform those using the standard squared-error objective function for unsupervised hyperspectral feature-learning.

Hyperspectral imagery of a vertical mine face acquired from a field-based platform is used to evaluate the effects of different conditions of illumination on absorption feature parameters wavelength position, depth and width. Imagery was acquired at different times of the day under direct solar illumination and under diffuse illumination imposed by cloud cover. Imagery acquired under direct solar illumination did not show large amounts of variability in any absorption feature parameter; however, imagery acquired under cloud caused changes in absorption feature parameters. These included the introduction of a spurious absorption feature at wavelengths > 2250 nm and a shifting of the wavelength position of specific clay absorption features to longer or shorter wavelengths. Absorption feature depth increased. The spatial patterns of clay absorption in imagery acquired under similar conditions of direct illumination were preserved but not in imagery acquired under cloud. Kaolinite, ferruginous smectite and nontronite were identified and mapped on the mine face. Results were validated by comparing them with predictions from x-ray diffraction and laboratory hyperspectral imagery of samples acquired from the mine face. These results have implications for the collection of hyperspectral data from field-based platforms.

Direct measurement of depth of the ~1900nm feature (F1900) caused by molecular water is not possible from hyperspectral data acquired under natural illumination. Proxy measures for depth were evaluated to estimate the true depth of F1900 from such data. Suitable proxy measures were identified using data acquired by a high-resolution spectrometer (under artificial and natural illumination) and by a hyperspectral imaging sensor (under artificial illumination), with different amounts of simulated atmospheric noise. The best performing proxy measures were used to estimate depth from hyperspectral imagery acquired in the field under natural illumination. Proxy measures comprised a ratio of reflectance, continuum-removed depth at a single wavelength and continuum-removed depth integrated over several wavelengths. For high resolution spectra, a ratio of reflectance at 2017nm and 1967nm (Ratio2017/1967) and the continuum-removed depth at 1967nm (CRD1967) were strongly correlated with depth R2=0.98–0.99 and R2=0.95–0.96, respectively. For hyperspectral imagery acquired in the laboratory, Ratio2017/1967 and CRD1967 were strongly correlated with depth across all levels of atmospheric noise (R2=0.94–0.98 and 0.94–0.97, respectively). Proxy measures using integrated depth performed relatively poorly (R2=0.65–0.86). In independent tests, depth of F1900 was consistently overestimated by CRD1967 but not by Ratio2017/1967. Validation of field imagery was done by comparing depth predicted by the proxy measures with depth measured from laboratory imagery of collocated samples from the field. Ratio2017/1967 derived from a polynomial fit to the data between 1961 and 2134nm (RatioP2017/1967) most closely approximated measured depth. Ratio2017/1967, derived from the original data either under or overestimated depth. Measures of continuum-removed depth generated from the original data (CRD1967) and from a polynomial fit to the original data (CRDP1967) overestimated the depth of F1900. This study showed that the depth of F1900 can be predicted from data acquired under natural illumination however, the choice of proxy measure can have a significant impact upon quantitative estimates