Hyperspectral Remote Sensing Analysis Training Course

Participants can inspect reflectance data and identify quality issues requiring correction or masking.

• Reflectance curves, absorption features and diagnostic spectral regions
• Radiance-to-reflectance conversion and reflectance product selection
• Signal-to-noise ratio assessed across wavelengths and image locations
• Identify dead pixels, striping, saturation and dropped spectral bands
• Water-vapour absorption regions removed before analytical processing
• Metadata review covering acquisition geometry, calibration and processing history

Participants can organise hyperspectral data and spatial references for reproducible analysis.

• ENVI, GeoTIFF and HDF formats with associated metadata
• Coordinate reference systems checked against field and vector reference data
• Subset scenes using areas of interest and analysis masks
• Raster alignment verified before combining hyperspectral and ancillary layers
• Create project folders, naming conventions and processing provenance records
• GDAL inspection commands for dimensions, bands, projections and no-data values

Participants can select and evaluate correction approaches appropriate to source imagery.

• Atmospheric scattering, absorption and adjacency effects in hyperspectral scenes
• FLAASH and QUAC assumptions, inputs and output interpretation
• Empirical line correction using field targets and reflectance references
• Topographic illumination effects assessed using slope, aspect and solar geometry
• Compare corrected spectra against library signatures and field observations
• Document correction parameters, exclusions and residual quality concerns

Participants can quantify noise sources and apply appropriate signal-enhancement treatments.

• Random noise, structured noise and spectral smile characteristics
• Dark-current effects and detector-specific artefacts identified from imagery
• Savitzky-Golay smoothing with window-size and polynomial-order selection
• Median filters and spatial smoothing without excessive boundary loss
• Noise-equivalent delta reflectance interpreted for sensor performance assessment
• Compare denoised spectra while retaining narrow diagnostic absorption features

Participants can transform spectra to reveal material-specific features and reduce nuisance variation.

• Continuum removal isolates absorption depth, position and feature width
• First and second derivatives highlight subtle spectral absorption changes
• Vector normalization reduces brightness variation between comparable spectra
• Standard normal variate correction addresses multiplicative scatter effects
• Band ratios and normalized difference indices target material responses
• Transformation choice matched to physical question and sensor characteristics

Participants can investigate spectral variability and identify candidate discriminating features.

• Spectral plots compare targets, backgrounds and reference library records
• Mean, variance and coefficient-of-variation profiles reveal unstable bands
• Spectral angle mapper measures directional similarity between spectra
• Euclidean distance and correlation quantify alternative spectral relationships
• Boxplots expose outliers within labelled training and validation samples
• Scatterplots examine band-pair separation before feature engineering decisions

Participants can apply and interpret PCA for spectral redundancy reduction.

• Covariance and correlation matrices as PCA input choices
• Mean centring and scaling decisions before component extraction
• Eigenvalues, eigenvectors and explained variance interpreted correctly
• Scree plots determine retained components using stated selection criteria
• Component loadings identify influential wavelengths and spectral regions
• PCA score images inspected for materials, gradients and artefacts

Participants can use MNF to separate signal-bearing components from noise-dominated components.

• MNF transformation combines noise whitening with principal component rotation
• Noise covariance estimated from local differences and homogeneous areas
• Eigenvalue ordering distinguishes coherent information from residual noise
• MNF component images reviewed for spatial structure and striping
• Retained MNF bands selected using eigenvalues, imagery and spectra
• Inverse MNF transformation reconstructs denoised hyperspectral image products

Participants can assess nonlinear and statistically independent transformations for complex scenes.

• Independent component analysis separates non-Gaussian source signals statistically
• FastICA assumptions, convergence behaviour and component interpretation
• Kernel PCA captures nonlinear spectral relationships using kernel functions
• Radial basis and polynomial kernels compared through validation results
• Parameter tuning controls component stability, separation and overfitting risk
• Choose PCA, MNF, ICA or kernel PCA by analytical objective

Participants can select compact informative band sets using quantitative ranking criteria.

• Filter methods rank bands using variance, correlation and mutual information
• Sequential forward selection builds compact predictive feature subsets
• Recursive feature elimination removes weak variables from supervised models
• Jeffries-Matusita distance quantifies class separability between reference samples
• Transformed divergence identifies discriminating bands for material classes
• Redundant bands removed using correlation thresholds and physical interpretation

Participants can prepare spectral references and define endmember strategies for target materials.

• USGS spectral library records matched to sensor wavelength sampling
• Resampling library spectra using sensor response functions
• Field spectra cleaned for illumination variation and calibration drift
• Pure pixels, endmembers and mixed pixels defined for image analysis
• Target and background signatures selected from imagery and field evidence
• Library metadata records material composition, measurement conditions and confidence

Participants can extract endmembers and estimate material fractions in mixed pixels.

• Pixel purity index identifies spectrally extreme candidate endmembers
• N-FINDR extracts simplex vertices representing pure spectral sources
• Vertex component analysis estimates endmembers from mixed hyperspectral pixels
• Linear spectral unmixing estimates fractional abundances under mixing assumptions
• Fully constrained least squares enforces nonnegative fractions summing to one
• Residual images reveal missing endmembers and model mismatch patterns

Participants can construct traceable feature stacks suited to classification and detection tasks.

• Combine selected bands, components, indices and texture measures
• Gray-level co-occurrence metrics describe local spatial texture patterns
• Feature scaling applied consistently between training and production data
• Masks exclude shadows, clouds, water and irrelevant land-cover areas
• Training samples linked to feature tables and spatial reference geometry
• Feature-stack documentation records source bands, transforms and parameter values

Participants can compare supervised classifiers using reduced hyperspectral feature sets.

• Support vector machines classify high-dimensional features with margin optimization
• Random forests rank variables and handle nonlinear class boundaries
• Gaussian maximum likelihood classification requires class distribution assumptions
• Training sample design addresses representativeness, balance and spatial autocorrelation
• Cross-validation estimates model performance before map production
• Confusion matrices compare classifier errors by target material class

Participants can identify spectral groupings and anomalous targets without complete labels.

• K-means clustering partitions spectra using centroid-based similarity
• ISODATA clustering adjusts cluster membership, merging and splitting iteratively
• Hierarchical clustering reveals spectral group relationships at multiple scales
• Reed-Xiaoli detector identifies statistically unusual spectral pixels
• Matched filtering detects known target signatures against variable backgrounds
• Cluster interpretation combines spectra, spatial context and field knowledge

Participants can validate hyperspectral outputs and communicate confidence, limitations and fitness for use.

• Independent validation samples separated from model-development reference data
• Overall accuracy, producer accuracy and user accuracy reported correctly
• F1 score and balanced accuracy address imbalanced target classes
• Spatially blocked validation reduces optimistic estimates from nearby samples
• Uncertainty maps communicate low-confidence predictions and ambiguous materials
• Technical reports record assumptions, methods, limitations and reproducibility information

Participants can plan repeatable hyperspectral processing for operational monitoring work.

• Standard operating procedures define inputs, parameters, outputs and acceptance criteria
• Version control tracks scripts, models, training data and processing changes
• Quality assurance checks verify projections, masks, ranges and output completeness
• Batch processing designs support repeated scenes and consistent feature generation
• Decision thresholds linked to field inspection, escalation and reporting actions
• Data governance addresses licensing, retention, access controls and provenance

Participants can build, review and present a complete hyperspectral feature-extraction analysis for a defined mapping objective.

• Define the mapping objective, target classes and acceptance criteria
• Prepare reflectance imagery, masks and documented quality-control evidence
• Select transformations, reduced components and discriminating spectral features
• Build a feature stack with endmembers or labelled reference samples
• Validate outputs using independent data, error metrics and uncertainty evidence
• Present maps, workflow rationale, limitations and operational recommendations