Remote Sensing Classification and Feature Extraction Training Course

Participants can define data requirements and acceptance criteria for a classification project.

• Mapping objectives translated into classes, minimum mapping units, and accuracy targets
• Imagery selection using seasonality, cloud cover, acquisition date, and resolution
• Data specifications for Sentinel-2, Landsat, commercial imagery, and aerial surveys
• Class definitions written with inclusion rules, exclusions, and examples
• Ancillary datasets assessed for provenance, scale, currency, and positional accuracy
• Project metadata recorded using ISO 19115 geographic information principles

Participants can prepare optical imagery for consistent feature extraction and classification.

• Atmospheric correction effects on reflectance values and cross-date comparability
• Cloud, shadow, snow, and water masking using quality-assurance bands
• Geometric correction checked through control points and image-to-image registration
• Mosaicking scenes with radiometric balancing, overlap inspection, and seam control
• Resampling methods compared: nearest neighbour, bilinear interpolation, cubic convolution
• Clipping analysis-ready rasters to project boundaries and processing extents

Participants can create and assess spectral features that separate target classes.

• Normalized Difference Vegetation Index calculated for vegetation condition mapping
• Normalized Difference Built-up Index applied to impervious surface discrimination
• Normalized Difference Water Index used for open-water feature extraction
• Band ratios selected from class spectral-response differences and sensor bands
• Principal component analysis used to reduce correlated multispectral information
• Tasseled Cap transformation interpreted through brightness, greenness, and wetness components

Participants can derive spatial context features that improve class separation.

• Grey-level co-occurrence matrix texture measures calculated from image neighbourhoods
• Contrast, homogeneity, entropy, and variance matched to target feature patterns
• Digital elevation models converted into slope, aspect, curvature, and hillshade
• Terrain position index used to distinguish ridges, valleys, and plains
• Multi-scale feature windows tested against object size and image resolution
• Feature correlation screened before model training and variable selection

Participants can create representative reference data for supervised classification and validation.

• Training polygons digitised within homogeneous areas and verified against reference sources
• Stratified random sampling allocated across rare, common, and confused classes
• Sample independence maintained between model training and accuracy assessment datasets
• Class imbalance identified through sample counts, prevalence, and spatial distribution
• Reference labels documented with source, interpreter, date, and confidence fields
• Field observations, orthophotos, and authoritative records compared for label quality

Participants can configure baseline pixel-based classifiers and interpret their outputs.

• Maximum likelihood classification built from class statistics and probability thresholds
• Minimum distance classification compared with spectral angle mapper outputs
• Class separability measured using Jeffries-Matusita distance and transformed divergence
• Decision boundaries inspected through scatterplots, histograms, and spectral profiles
• Post-classification filtering removes isolated pixels using majority neighbourhood rules
• Pixel-based outputs reviewed against class boundaries, mixed pixels, and spatial noise

Participants can train, tune and compare machine learning classifiers for imagery.

• Random Forest trained with spectral, texture, terrain, and ancillary predictors
• Support vector machines configured using kernel choice, C, and gamma parameters
• Gradient boosting evaluated for nonlinear class boundaries and predictor interactions
• Cross-validation used to estimate model performance before final deployment
• Feature importance ranked using permutation measures and model-based scores
• Hyperparameter search structured through grid search and validation partitions

Participants can extract meaningful image objects and classify them using object attributes.

• Multiresolution segmentation controlled through scale, shape, and compactness parameters
• Segmentation quality assessed against target object size and boundary fidelity
• Object attributes calculated from spectral means, geometry, texture, and context
• Rule sets combine thresholds, spatial relationships, and class hierarchy logic
• Parcel, roof, road, and water objects classified from image segments
• Object-based outputs compared with pixel classification for fragmented landscapes

Participants can assess when convolutional neural networks are suitable for feature extraction.

• Convolutional neural network architecture interpreted through kernels, layers, and feature maps
• Image chips prepared with labels, augmentation, and balanced class representation
• Semantic segmentation distinguished from image classification and object detection
• Transfer learning applied using pretrained image models and local training data
• Overfitting monitored through learning curves, validation loss, and held-out samples
• Deep learning outputs reviewed for explainability, error patterns, and operational constraints

Participants can validate classifications and report accuracy using accepted statistical measures.

• Confusion matrices constructed from independent reference samples and predicted classes
• Producer's accuracy, user's accuracy, precision, recall, and F1 score calculated
• Overall accuracy interpreted alongside class prevalence and error consequences
• Area-adjusted accuracy estimation used for stratified map validation designs
• Confidence intervals calculated for accuracy estimates and mapped area totals
• Error matrices reported with class definitions, sampling design, and limitations

Participants can produce defensible change products from aligned multi-date imagery.

• Image co-registration verified before comparing pixels, objects, and mapped features
• Post-classification comparison identifies transitions between defined land-cover classes
• Image differencing applied to spectral indices and surface reflectance bands
• Change vector analysis measures magnitude and direction across feature space
• Thresholds selected using reference samples, histograms, and change objectives
• Transition matrices summarise gains, losses, persistence, and change uncertainty

Participants can use radar backscatter and texture features in classification workflows.

• SAR backscatter explained through wavelength, polarisation, incidence angle, and surface properties
• Speckle filtering compared using Lee, Frost, and refined Lee filters
• Radiometric terrain correction applied before mountainous-area backscatter interpretation
• VV, VH, HH, and HV polarisation features selected for target classes
• Interferometric coherence interpreted for surface stability and disturbance detection
• Optical and SAR features combined for cloud-prone monitoring requirements

Participants can derive terrain and three-dimensional features from LiDAR and elevation data.

• Point clouds classified into ground, vegetation, building, and noise returns
• Digital terrain models separated from digital surface models and canopy models
• Canopy height models calculated from surface and terrain elevation differences
• Building footprints extracted using height thresholds, planarity, and object geometry
• Slope, roughness, and local relief models derived for terrain classification
• LiDAR metrics integrated with imagery for forest and infrastructure mapping

Participants can select suitable techniques for high-dimensional hyperspectral classification tasks.

• Hyperspectral signatures interpreted through absorption features and narrow spectral bands
• Noise bands removed using sensor diagnostics and signal-to-noise assessment
• Minimum noise fraction transformation used for dimensionality reduction
• Continuum removal applied to isolate diagnostic mineral and vegetation absorptions
• Spectral angle mapper used for library-based material identification
• Endmember selection supports linear spectral mixture analysis and sub-pixel mapping

Participants can design repeatable processing chains for scalable image classification work.

• Batch processing designed with parameter files, naming conventions, and processing logs
• Cloud geospatial platforms assessed for data access, processing limits, and governance
• Reusable scripts separate input paths, parameters, functions, and output locations
• Raster processing chains documented through source data, transformations, and outputs
• Version control applied to scripts, class schemas, and model configurations
• Processing costs estimated from imagery volume, compute demand, and storage retention

Participants can package classification outputs for review, publication and operational maintenance.

• Quality checks inspect projection, extent, nodata values, classes, and topology
• Map symbology designed to distinguish classes without implying unsupported certainty
• Raster and vector outputs packaged with metadata, lineage, and accuracy statements
• Acceptance criteria linked to class accuracy, completeness, positional quality, and timeliness
• Stakeholder review sessions structured around sample areas, errors, and decision impacts
• Update schedules defined from sensor revisit, seasonal change, and business requirements

Participants can build, review and present a complete classification and feature-extraction project with validated outputs.

• Project brief defines decision use, study area, classes, outputs, and constraints
• Input imagery prepared using documented preprocessing and quality-control decisions
• Feature stack assembled from spectral, texture, terrain, radar, or object attributes
• Classification model selected through comparative performance and operational suitability
• Validation evidence presented using confusion matrices, error analysis, and confidence intervals
• Delivery package reviewed against metadata, map design, reproducibility, and stakeholder requirements