HyperSee Features

Sample Spectra Selection

One of the most important steps in extracting information from hyperspectral images is the proper selection of data subsets from the massive data available within even a single image hypercube. Which regions are most important, and how can they be easily identified and selected? HyperSee offers a wide range of data visualization and selection tools. Spatial scene selection tools include point, cross, and square cursors; straight and freehand lines; and rectangle regions of interest. For proper object or feature identification and modeling it is often necessary to select representive regions from a complex spatial shape. HyperSee provides image background or feature thresholding tools which can be intuitively combined with spatial selection tools to easily select irregular shaped objects. Spectral plotting windows enable interactive image exploration, displaying plots of selected sets of spectra or their means.

Spectral Libraries

In addition to data exploration purposes, spectral selection sets can be accumulated and used for regression model building. Libraries of selected sample spectra can be easily maintained with HyperSee. Each library file contains a collection of spectra representing different classes, which can then be used for either PLS or PLS-DA regression model building. Sample spectra for each class can originate from either a single or multiple hypercube data sources, and are easily archived for future use. Data from an individual image, class, or selected region can be deleted or appended to at any time.

Spectral Pre-Processing

Changes in instrumentation (thermal drift, sensor sensitivity, or noise), lighting conditions, or even sample preparation and presentation - can all contribute to spectral data variations. Numerical spectral pre-processing filters can sometimes be applied to reduce these effects. HyperSee offers a wide range of pre-processing filters: first and second order normalizations; smoothing, first and second derivatives based on different filter window sizes; linear detrending; and standard normal variate (SNV) transformations. Up to three different filters can be applied sequentially. Regression models constructed and saved from pre-processed spectra contain filter information necessary for transformation and prediction of any future spectral image dataset.

Model Building & Validation

Ideally, multivariate model training spectra should include all possible sources of sample variance. How should these training sample sets be selected from the hundreds or thousands of hyperspectral image spectra? The classical approach of computing mean spectra for each sample class can be used, however this does not provide sufficient robustness. HyperSee provides several additional options for sub-sampling: randomized selections from either spatial locations or Principal Component Analysis (PCA) scatter plots of each data class. To provide additional model robustness, repeated subsampling and model building can be performed (bootstrapping). How can the performance of these models be evaluated? Tables of class prediction statistics are available in HyperSee, however hyperspectral imaging provides additional performance tools: image prediction maps. Color coded spatial image maps can be produced which indicate classification or quantification results. In the case of classification, maps indicating training and test samples not correctly classified can also be created.

Model Optimization

One of the key questions regarding multivarite regression models based on PLS or PLS-DA is: How many latent variables or factors should be included in the model? And in the case of PLS-DA: How should the class membership threshold be determined? HyperSee provides traditional diagnostics such as plots of regression vectors and PRESS plots of prediction errors - RMSEC and RMSEP. HyperSee provides additional innovative tools for examining and optimizing model performance. Levels of underfitting or overfitting may be determined from PRESS plots of additional model statistics related to regression vector jaggedness and correlation. The use of interactive cursors with the simultaneous display of prediction value histograms and image maps can assist with the selection of optimal classification threshold values. HyperSee also creates classification confusion tables of both training and test datasets, which may be examined for inter-class relationships.

In the case of quantitative analysis, PLS regression model results may be similarly viewed as both spatial prediction maps, and histogram distributions of prediction values. Data displays may be autoscaled, or set to arbitrary values defined by the user. Individual PLS models may be displayed color coded to sample prediction values. For multiple PLS models, gray scale levels may be mapped to red, green, and blue (RGB) to produce pseudo-color prediction maps.

Data File Formats

HyperSee supports image files acquired from all HyperPro imaging systems, and supports ENVI compatible as well as Malvern - MatrixNIR image files. All PLS and PLS-DA models may be saved and recalled for later use in prediction of independent hyperspectral images. All spectral data sets, models, and prediction maps may be exported as MATLAB compatible files.