Spatial Machine Learning Training Course

Participants can assemble compatible vector, raster and tabular data for modelling.

• Vector geometries, raster grids and point observations for predictive modelling
• Coordinate reference systems selected for distance, area and direction calculations
• Reprojecting layers without introducing alignment or resolution errors
• Spatial extent, pixel size and temporal coverage assessed before integration
• GeoPackage, GeoTIFF, Cloud Optimized GeoTIFF and Parquet formats compared
• Metadata records documenting source, date, resolution, licence and processing

Participants can assess training data quality and create a defensible label specification.

• Label definitions tied to observable events, classes and dates
• Positional accuracy, completeness and attribute consistency checks
• Class imbalance measured using prevalence, confusion patterns and sampling counts
• Stratified sampling designs for rare classes and heterogeneous landscapes
• Field observations matched to imagery acquisition dates and spatial support
• Data lineage captured from source collection through training-table creation

Participants can identify spatial structure, outliers and candidate explanatory relationships.

• Moran's I used to measure global spatial autocorrelation
• Local Indicators of Spatial Association reveal clusters and spatial outliers
• Kernel density estimation maps concentration patterns from point events
• Variograms describe distance-dependent similarity in continuous observations
• Bivariate maps expose relationships between targets and candidate predictors
• Distribution diagnostics identify skewness, missingness and implausible measurement values

Participants can derive predictive features from vector layers and spatial relationships.

• Distance-to-feature variables calculated for roads, rivers and facilities
• Buffer summaries aggregate land use, population and asset attributes
• Spatial joins attach polygon and neighbourhood characteristics to observations
• Network distance separates travel access from straight-line proximity
• Terrain derivatives generated from digital elevation models
• Neighbourhood statistics capture local density, diversity and configuration

Participants can derive spectral, temporal and texture features from raster imagery.

• Spectral bands sampled consistently at labelled training locations
• NDVI, NDWI and NDBI indices derived from multispectral imagery
• Cloud masking and compositing rules applied before feature extraction
• Texture measures calculated using gray-level co-occurrence matrices
• Multi-temporal image stacks capture seasonality and change signals
• Raster resampling methods matched to categorical and continuous variables

Participants can structure a repeatable workflow from raw geospatial data to model-ready tables.

• Python notebooks organised into ingestion, transformation, modelling and reporting stages
• GeoPandas and Rasterio workflows handle vector and raster inputs
• Parameter files record paths, dates, thresholds and processing choices
• Train, validation and test datasets stored with immutable identifiers
• Quality-control checks flag duplicate geometries and invalid attribute ranges
• Versioned outputs preserve reproducibility for model review and reruns

Participants can build and assess baseline classification models for discrete mapped outcomes.

• Logistic regression establishes an interpretable classification benchmark
• Decision trees identify nonlinear splits and interaction candidates
• Confusion matrices report class-specific classification errors
• Precision, recall, F1 score and ROC-AUC compared
• Probability thresholds selected against intervention capacity and error costs
• Prediction probabilities mapped before converting results into operational classes

Participants can model continuous outcomes and map prediction error across an area.

• Linear regression baseline models continuous spatial response variables
• Random forest regression captures nonlinear environmental and operational drivers
• Mean absolute error, RMSE and R-squared interpreted together
• Residual maps identify geographic patterns missed by fitted models
• Quantile regression forests estimate prediction intervals by location
• Continuous prediction surfaces classified without concealing original model values

Participants can tune and compare ensemble models for spatial prediction tasks.

• Random forests tuned using tree count, depth and feature sampling
• Gradient boosting models configured with learning rate and regularisation
• XGBoost handles nonlinear interactions and missing predictor values
• Grid search and random search compared for hyperparameter selection
• Feature importance checked for instability under correlated predictors
• Model comparison tables record performance, complexity and operational suitability

Participants can design validation schemes that estimate geographic model transferability.

• Random cross-validation contrasted with spatially blocked cross-validation
• Spatial blocks sized using autocorrelation range and deployment geography
• Leave-one-region-out validation tests geographic transfer to new areas
• Temporal holdouts separate historical training data from future deployment data
• Data leakage traced through duplicates, buffers and overlapping imagery tiles
• Validation reports distinguish interpolation performance from extrapolation performance

Participants can identify meaningful spatial groupings where labels are absent or incomplete.

• K-means clusters standardised environmental and socioeconomic feature profiles
• DBSCAN identifies dense clusters while labelling isolated observations as noise
• Hierarchical clustering supports interpretable geographic typologies
• Principal component analysis reduces correlated spectral and census variables
• Cluster validity assessed using silhouette scores and map coherence
• Segment profiles translated into service, risk or land-management categories

Participants can select and evaluate deep learning approaches for image-based mapping tasks.

• Convolutional neural networks learn spatial patterns from image patches
• Image patch extraction preserves labels, boundaries and class balance
• Data augmentation uses rotations, flips and brightness transformations
• Transfer learning adapts pretrained image models to local imagery
• Overfitting diagnosed through learning curves and held-out tile performance
• GPU requirements, training time and model complexity assessed realistically

Participants can distinguish pixel, object and segmentation approaches for mapped features.

• Semantic segmentation assigns classes to every image pixel
• U-Net architecture combines encoder features with decoder localisation
• Intersection over Union measures overlap between predicted and reference masks
• Object-based image analysis groups pixels before feature classification
• Boundary errors examined for buildings, roads, fields and watercourses
• Post-processing removes isolated pixels and enforces minimum mapping units

Participants can explain model behaviour, communicate uncertainty and identify responsible-use controls.

• Permutation importance ranks predictors by contribution to model performance
• Partial dependence plots show average feature-response relationships
• SHAP values explain individual predictions and local variation
• Uncertainty maps distinguish low confidence from low predicted probability
• Bias audits test performance differences across places and population groups
• Model cards document intended use, limitations, data and review controls

Participants can specify how a spatial model will be delivered, monitored and maintained.

• Batch prediction workflows generate scheduled raster and vector outputs
• Web map layers publish predictions with legends and metadata
• Decision rules connect scores to inspections, alerts and resource allocation
• Data drift monitored through feature distributions and imagery changes
• Performance monitoring compares new observations against predicted outcomes
• Retraining triggers defined by accuracy decline, changed labels or new coverage

Participants can build, review and present a complete spatial machine learning project with documented recommendations.

• Problem framing specifies target, geography, users and decision actions
• Data inventory records sources, labels, resolutions and quality limitations
• Feature pipeline produces documented vector, raster and temporal predictors
• Spatial validation design tests performance against realistic deployment conditions
• Model comparison selects outputs using accuracy, uncertainty and interpretability
• Presentation communicates maps, limitations, controls and implementation recommendations