Machine Learning for GIS Applications Training Course

Participants can assess geospatial data fitness before model development.

• Coordinate reference systems, projections and distance-based feature calculation
• Positional accuracy, temporal currency and attribute completeness assessment
• Topology rules for polygons, lines, networks and administrative boundaries
• Missing values, duplicate records and inconsistent spatial identifiers
• Metadata review using ISO 19115 lineage and quality elements
• Data inventories linking source datasets to prediction requirements

Participants can configure a reproducible environment for geospatial model development.

• Python notebooks structured for data loading, modelling and reporting
• GeoPandas workflows for vector data selection, joins and overlays
• Rasterio operations for reading raster bands, masks and metadata
• Scikit-learn pipelines for repeatable preprocessing and model fitting
• File paths, environments and package versions recorded for reproducibility
• QGIS and ArcGIS Pro outputs integrated with Python analysis

Participants can create training datasets that represent the intended prediction area.

• Sampling frames matched to the prediction population and resolution
• Stratified random sampling for imbalanced land-cover and inspection classes
• Reference labels gathered from surveys, imagery interpretation and administrative records
• Label definitions written with inclusion rules, exclusions and confidence grades
• Class imbalance addressed through weights, resampling and threshold selection
• Training, validation and holdout datasets separated before feature selection

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

• Distance-to-feature variables calculated from roads, rivers and service points
• Buffer summaries derived from land use, population and facility layers
• Spatial joins aggregating counts, densities and proportions by analysis unit
• Network distance measures for accessibility and response-time prediction
• Terrain attributes extracted from digital elevation models and contours
• Feature dictionaries recording definitions, units, sources and transformation rules

Participants can prepare imagery and derive spectral predictors for machine learning.

• Radiometric correction concepts for comparable satellite image reflectance values
• Cloud, shadow and no-data masks applied before pixel classification
• Spectral bands selected for vegetation, water, soil and built surfaces
• NDVI, NDWI and NDBI indices calculated from multispectral imagery
• Texture measures derived using grey-level co-occurrence matrices
• Image composites constructed from seasonal scenes and temporal summaries

Participants can identify patterns, outliers and data relationships before model fitting.

• Distribution plots used to inspect skewness, outliers and missingness
• Correlation matrices screened for redundant numeric predictor variables
• Moran's I used to measure global spatial autocorrelation
• Local Indicators of Spatial Association mapped for clustered values
• Bivariate maps compared against known processes and field knowledge
• Feature leakage identified before model training and performance reporting

Participants can train and compare classifiers for categorical geospatial targets.

• Decision trees fitted for transparent rule-based land classification
• Random Forest classifiers configured using trees, depth and feature sampling
• Support vector machines applied to separable spectral feature spaces
• Gradient boosting models trained for complex non-linear class boundaries
• Probability outputs converted to classes using decision thresholds
• Confusion matrices interpreted with precision, recall and F1 scores

Participants can build regression models and map continuous spatial predictions.

• Linear regression established as a transparent continuous prediction baseline
• Random Forest regression fitted for non-linear environmental response patterns
• Gradient boosting regression tuned for asset condition and risk scores
• Residual maps inspected for spatially structured prediction errors
• Mean absolute error and root mean squared error compared
• Prediction surfaces generated with units, ranges and valid-area masks

Participants can identify spatial groups and unusual observations without labelled outcomes.

• K-means clustering groups locations using standardised spatial attribute profiles
• Hierarchical clustering reveals nested patterns within survey and asset data
• DBSCAN identifies dense clusters while separating spatial noise points
• Principal component analysis reduces correlated multispectral feature dimensions
• Isolation Forest flags unusual records for inspection prioritisation
• Cluster profiles translated into interpretable geographic management categories

Participants can apply object-based methods when pixels do not represent meaningful units.

• Image segmentation creates objects from spectral, shape and scale criteria
• Segment attributes summarise mean bands, texture and geometric properties
• Object-based classification improves mapping of buildings, fields and roads
• Training polygons matched to segments without mixed-object label errors
• Segmentation scale assessed against target feature size and imagery resolution
• Object boundaries reviewed against reference imagery and field evidence

Participants can judge when deep learning is appropriate for imagery-based GIS tasks.

• Convolutional neural networks recognise spatial patterns within image patches
• Semantic segmentation assigns classes to individual image pixels
• Object detection locates vehicles, buildings and infrastructure features
• Transfer learning adapts pre-trained models to limited local labels
• Training data volume, graphics hardware and annotation costs assessed
• Deep learning outputs compared with conventional machine learning baselines

Participants can evaluate models using validation designs appropriate for spatial dependence.

• Random cross-validation compared with spatial block cross-validation results
• Spatial folds designed around grids, regions or temporal periods
• Hyperparameter search performed with grid search and random search
• Learning curves diagnose underfitting, overfitting and insufficient training data
• Calibration plots assess whether predicted probabilities match observed frequencies
• Model selection balances accuracy, interpretability, cost and maintenance needs

Participants can communicate model uncertainty and manage risks in geospatial decision support.

• Prediction uncertainty mapped using probabilities, intervals and ensemble variation
• Error hotspots examined by geography, class, season and data source
• Sampling bias assessed where observations reflect uneven service coverage
• Proxy variables reviewed for discriminatory or inappropriate decision effects
• Human review points defined for high-consequence model recommendations
• Model cards document intended use, limitations and monitoring requirements

Participants can package model outputs for repeatable operational GIS use.

• Batch workflows scheduled for recurring imagery and operational data updates
• Data pipelines separate raw inputs, processed features and approved outputs
• Web maps configured with legends, confidence layers and user guidance
• Map symbology distinguishes predictions, observations and uncertain areas
• Decision thresholds linked to inspections, interventions and escalation routes
• Version control records model releases, datasets and parameter changes

Participants can build, review and present a complete geospatial machine learning project for an operational decision.

• Project scope defines users, prediction target, geography and decision actions
• Training dataset assembled with documented labels, features and quality checks
• Selected model trained using spatial validation and performance metrics
• Prediction map produced with uncertainty layers and decision thresholds
• Peer review tests assumptions, errors, governance controls and usability
• Implementation roadmap presents data updates, ownership and success measures