AI and Machine Learning for Geospatial Analysis Training Course

Participants can assess whether spatial datasets are fit for machine-learning development.

• Coordinate reference systems, datums and transformations checked before integration
• Spatial resolution, temporal resolution and positional accuracy assessed systematically
• STAC metadata concepts for cataloguing imagery and derived assets
• Schema design for features, labels, predictions and model provenance
• Missing values, duplicate geometries and topology errors identified early
• Data lineage recorded from acquisition through processing and publication

Participants can prepare analysis-ready raster stacks from multisource imagery.

• Radiometric calibration and atmospheric correction for optical satellite imagery
• Cloud, shadow and snow masking using quality-assurance bands
• Resampling methods selected for continuous, categorical and spectral datasets
• Raster alignment verified through extent, projection, pixel size and grid origin
• Mosaics and temporal composites created without obscuring source metadata
• Normalized Difference Vegetation Index calculated from red and near-infrared bands

Participants can create representative labelled datasets and sampling plans for model training.

• Reference labels sourced from surveys, imagery interpretation and administrative records
• Class definitions written with observable criteria and exclusion rules
• Stratified random sampling balanced across classes, regions and conditions
• Spatial autocorrelation considered when separating training and validation samples
• Label uncertainty recorded using confidence scores and reviewer agreement
• Sampling bias diagnosed through maps, frequency tables and coverage checks

Participants can identify informative geospatial predictors before model training.

• Distribution plots reveal skew, outliers, missingness and class imbalance
• Spectral indices derived for vegetation, water, built surfaces and burn severity
• Terrain derivatives generated from elevation models: slope, aspect and curvature
• Texture measures calculated using grey-level co-occurrence matrices
• Proximity features measured from roads, rivers, facilities and boundaries
• Multicollinearity screened using correlation matrices and variance inflation factors

Participants can organise reproducible Python workflows for spatial data and model development.

• Python environments structured with packages, notebooks, scripts and configuration files
• GeoPandas operations join attributes, filter features and manage geometries
• Raster arrays loaded, masked and transformed for model-ready inputs
• Pandas tables connect sampled pixels, labels and engineered predictors
• Functions separate data preparation, training, evaluation and map export
• Version control principles protect scripts, parameters and project documentation

Participants can train and apply baseline supervised classifiers to geospatial data.

• Classification pipelines split features, labels and held-out evaluation samples
• Logistic regression establishes interpretable baseline models for binary targets
• k-nearest neighbours classified using scaled spatial and spectral predictors
• Support vector machines tuned through kernels, margins and penalty parameters
• Class weights address imbalanced labels without duplicating rare observations
• Prediction rasters generated with consistent feature ordering and metadata

Participants can build and tune tree-based models for classification and regression mapping.

• Decision trees split predictors using impurity reduction and stopping rules
• Random forests estimate stable predictions from bootstrapped decision trees
• Gradient boosting improves residual errors through sequential weak learners
• Hyperparameters tuned using search grids and spatial validation folds
• Feature importance compared using permutation tests and split-based measures
• Regression maps created for continuous variables such as yield or exposure

Participants can select appropriate neural-network approaches for geospatial imagery tasks.

• Neural-network layers transform imagery through weights, activations and loss functions
• Convolutional filters detect edges, textures, shapes and spatial arrangements
• Training, validation and test partitions prevent misleading performance estimates
• Data augmentation expands labelled imagery through rotations, flips and crops
• Transfer learning adapts pretrained image models to local classes
• GPU requirements, batch sizes and epochs influence training feasibility

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

• Semantic segmentation assigns a class to every image pixel
• Object-based image analysis groups pixels using scale, shape and compactness
• U-Net architecture supports boundary-sensitive segmentation of geospatial imagery
• Object detection identifies bounding boxes for roofs, vehicles and infrastructure
• Intersection over Union measures overlap between predicted and reference objects
• Post-processing removes speckle, fills gaps and regularises mapped boundaries

Participants can find spatial clusters, anomalies and unexplained patterns without complete labels.

• k-means clustering groups observations by spectral and environmental similarity
• DBSCAN identifies dense clusters while separating noise and isolated observations
• Principal component analysis reduces correlated bands into informative components
• Isolation forests flag unusual asset, sensor and land-cover signatures
• Cluster maps interpreted against known land use and operational context
• Anomaly thresholds calibrated using review capacity and false-alert costs

Participants can detect and interpret meaningful changes across repeated geospatial observations.

• Image differencing compares calibrated observations from two acquisition dates
• Change vector analysis measures multiband magnitude and directional change
• Time-series compositing reduces cloud contamination and seasonal noise
• Seasonality separated from disturbance using trend and residual analysis
• Breakpoint detection identifies abrupt changes in vegetation or surface condition
• Change alerts prioritised by magnitude, persistence, location and asset criticality

Participants can assess geospatial model performance using valid spatial evaluation designs.

• Confusion matrices report correct classifications, omissions and commission errors
• Precision, recall, F1 score and balanced accuracy compared by class
• Root mean squared error evaluates continuous prediction surfaces
• Spatial block cross-validation reduces optimistic estimates from nearby samples
• Area-adjusted accuracy estimates account for unequal mapped class proportions
• Uncertainty maps communicate confidence, disagreement and extrapolation risk

Participants can explain model behaviour and investigate geographically concentrated errors.

• Partial dependence plots show predictor effects across modelled ranges
• SHAP values attribute individual predictions to input feature contributions
• Residual maps reveal geographic patterns in model underperformance
• Error strata examined by season, sensor, terrain and settlement type
• Feature leakage identified when predictors contain target-related future information
• Model cards document purpose, training data, limitations and intended users

Participants can plan scalable processing for large imagery collections and prediction outputs.

• Tiled processing manages memory when predicting large raster extents
• Cloud-Optimized GeoTIFF files support efficient range requests and visualisation
• Google Earth Engine concepts support image collections and server-side computation
• Chunking strategies balance processing speed, memory use and edge effects
• Distributed processing risks include inconsistent environments and hidden data dependencies
• Cost estimates consider storage, computation, egress and model retraining frequency

Participants can specify controls for repeatable geospatial model training and updates.

• Data versioning links each model run to exact source datasets
• Experiment tracking records parameters, metrics, code versions and artefacts
• Model registries control approved versions, owners and release status
• Automated checks test geometry validity, schemas and raster alignment
• Drift monitoring detects changes in inputs, labels and model accuracy
• Retraining triggers defined through performance thresholds and operational events

Participants can convert model outputs into usable maps, alerts and action priorities.

• Prediction outputs translated into risk tiers, work queues and service areas
• Threshold selection balances false alarms, missed events and review capacity
• Priority scores combine model probability, consequence and intervention cost
• Map symbology distinguishes predictions, confidence, observations and restricted information
• Human review steps address ambiguous cases and high-consequence decisions
• Feedback loops capture field outcomes for future model improvement

Participants can identify governance controls for responsible geospatial AI implementation.

• Location-data sensitivity assessed for people, communities, assets and ecosystems
• Purpose limitation defined before collecting personal or high-resolution location data
• Bias risks examined across geographic coverage, labels and protected groups
• Human oversight assigned for enforcement, eligibility and safety-related decisions
• Retention schedules specify storage periods, access rights and deletion controls
• Audit trails preserve model inputs, approvals, outputs and reviewer decisions

Participants can prepare a feasible project plan for a production geospatial AI initiative.

• Project charters define decisions, users, scope, success measures and constraints
• Data inventories identify availability, ownership, refresh rates and quality gaps
• Work breakdown structures sequence labelling, engineering, modelling and deployment
• Acceptance criteria specify minimum accuracy, latency and map usability requirements
• Risk registers cover data access, bias, compute capacity and operational adoption
• Stakeholder communication plans schedule technical reviews and decision briefings

Participants can build, review and present a complete geospatial machine-learning analysis with documented operational recommendations.

• Problem statement connects a mapped prediction to a defined operational decision
• Prepared dataset combines imagery, vector predictors, labels and provenance records
• Selected model justified against baseline performance and data constraints
• Spatial validation results reported with class-level metrics and uncertainty maps
• Error analysis identifies geographic limitations, bias risks and review requirements
• Deployment plan defines owners, refresh cycles, controls and decision outputs