AI-Powered Remote Sensing Training Course

Participants can select imagery and supporting datasets for a defined analytical objective.

• Sentinel-1 SAR products for cloud-affected land and flood observation
• Sentinel-2 multispectral bands for vegetation, water and built-surface analysis
• Landsat Collection 2 imagery for long-term environmental change assessment
• Commercial optical imagery choices: resolution, licensing, tasking and archive access
• Drone imagery specifications for asset inspections and localised mapping
• Ancillary vector, elevation, weather and field-survey datasets for model context

Participants can organise spatial datasets with consistent metadata, projections and lineage.

• Coordinate reference systems selected for imagery alignment and area measurement
• Raster grids aligned through pixel size, origin and extent controls
• Metadata records covering source, acquisition date, processing and licence conditions
• STAC catalogues used to discover and manage earth observation assets
• File formats compared: GeoTIFF, Cloud Optimized GeoTIFF, GeoPackage and Parquet
• Data lineage documented from raw acquisition through analytical output publication

Participants can prepare analysis-ready imagery while documenting quality limitations.

• Radiometric calibration converts digital numbers into reflectance or backscatter values
• Atmospheric correction methods compared for multispectral surface reflectance products
• Cloud, shadow and snow masking using quality bands and spectral tests
• SAR speckle filtering balances noise reduction against spatial detail preservation
• Mosaics and temporal composites built from compatible scenes and acquisition windows
• Visual and statistical checks identify striping, gaps, haze and misregistration

Participants can derive and select geospatial features that support model performance.

• NDVI, NDWI and NBR indices derived from calibrated multispectral bands
• Texture measures calculated using gray-level co-occurrence matrices and neighbourhood windows
• Elevation, slope and aspect added from digital elevation models
• Seasonal statistics generated from image collections and observation time series
• SAR polarisation ratios used to distinguish surface structure and moisture
• Feature correlation screening reduces redundancy before supervised model training

Participants can create representative labelled datasets with auditable quality controls.

• Class taxonomy defined using mutually exclusive labels and operational definitions
• Stratified sampling allocates observations across classes, regions and seasons
• Training polygons digitised with minimum mapping units and boundary rules
• Reference labels reconciled using field records, imagery interpretation and independent sources
• Class imbalance diagnosed through sample counts, prevalence and spatial coverage
• Label audits record interpreter agreement, ambiguous cases and correction decisions

Participants can build baseline classifiers for pixel and object-based mapping tasks.

• Random Forest classification trained using spectral, texture and terrain features
• Support Vector Machines compared for limited samples and high-dimensional features
• Gradient boosting applied to structured remote sensing feature tables
• Object-based image analysis segments imagery before feature extraction and classification
• Training, validation and test partitions separated by spatial sampling strategy
• Baseline maps exported with class codes, confidence values and metadata

Participants can evaluate model performance using statistically defensible spatial validation methods.

• Confusion matrices report producer accuracy, user accuracy and overall accuracy
• Precision, recall and F1 scores interpret class-specific model behaviour
• Spatial cross-validation prevents overly optimistic accuracy from neighbouring samples
• Area-adjusted accuracy estimates correct mapped area using reference sample weights
• Confidence intervals calculated for accuracy metrics and class-area estimates
• Error maps reveal geographic patterns, omitted classes and problematic boundaries

Participants can design scalable image-processing workflows using cloud geospatial services.

• Google Earth Engine collections filtered by location, date and cloud percentage
• Server-side processing concepts distinguish deferred computation from local data handling
• Image collection reducers create median, percentile and quality mosaics
• Batch exports configured with projection, scale, region and file naming rules
• GDAL commands automate reprojection, clipping, translation and raster statistics
• Processing logs capture inputs, parameters, runtime and output locations

Participants can structure reproducible Python workflows for remote sensing model development.

• Rasterio reads georeferenced imagery while preserving transforms and coordinate systems
• GeoPandas joins vector labels, attributes and spatial sample locations
• NumPy arrays reshape raster bands into machine-learning feature matrices
• Scikit-learn pipelines combine preprocessing, training and model evaluation stages
• Configuration files separate model parameters from reusable processing code
• Version-controlled repositories record scripts, environments and analytical decisions

Participants can explain and configure convolutional neural networks for geospatial imagery.

• Convolution, pooling and feature maps extract hierarchical image patterns
• Image tiles generated with overlap, padding and consistent label alignment
• Data augmentation applies rotations, flips and spectral perturbations appropriately
• Transfer learning adapts pretrained encoders to limited geospatial training datasets
• Training curves diagnose overfitting, underfitting and unstable optimisation behaviour
• GPU memory, batch size and tile dimensions determine training feasibility

Participants can design segmentation workflows for detailed pixel-level mapping.

• U-Net architecture predicts class labels for every image pixel
• Segmentation masks prepared with class codes, nodata rules and boundary conventions
• Intersection over Union measures overlap between predicted and reference regions
• Dice loss addresses foreground-background imbalance in sparse target classes
• Post-processing removes isolated pixels and applies minimum mapping-unit thresholds
• Segmentation outputs converted into polygons with retained confidence attributes

Participants can plan object-detection models for discrete features in aerial and satellite imagery.

• Bounding-box annotations define object extent, class and difficult-instance flags
• YOLO detection models balance inference speed and small-object performance
• Non-maximum suppression removes duplicate detections using overlap thresholds
• Mean average precision evaluates detection quality across confidence thresholds
• Small-object detection improved through resolution, tiling and targeted sampling
• Detected assets linked to GIS inventories using spatial joins and identifiers

Participants can derive temporal indicators from repeated earth observation measurements.

• Time-series compositing reduces cloud effects while retaining seasonal signal patterns
• Harmonic regression models recurring seasonal variation in vegetation indices
• Phenological metrics estimate green-up, peak growth and senescence dates
• Temporal smoothing methods compare moving averages, Savitzky-Golay and Whittaker filters
• Anomaly scores identify departures from historical seasonal baselines
• Missing observations handled through interpolation, compositing and quality flags

Participants can detect, validate and communicate meaningful spatial change between observations.

• Image differencing highlights spectral changes between aligned multi-date scenes
• Post-classification comparison quantifies transitions between mapped land-cover classes
• Change Vector Analysis combines magnitude and direction of multiband change
• SAR coherence change identifies disturbance following floods, earthquakes and construction
• Threshold selection tested against reference observations and false-alert tolerances
• Change maps distinguish confirmed events, uncertainty zones and no-data areas

Participants can identify and report model limitations that affect remote sensing decisions.

• Feature importance ranks predictors within tree-based classification models
• SHAP values explain individual predictions and unexpected class assignments
• Geographic bias assessed across regions, seasons, sensors and demographic proxies
• Domain shift examined when models move between locations or acquisition conditions
• Uncertainty surfaces communicate confidence beyond single categorical map outputs
• Model cards document intended use, limitations, data and validation evidence

Participants can plan controlled deployment of remote sensing models within operational processes.

• Model deployment options compared: batch processing, APIs and analyst-assisted review
• Human-in-the-loop triage routes low-confidence predictions for expert assessment
• Data governance assigns ownership, retention, access and licensing responsibilities
• Monitoring thresholds flag accuracy drift, input changes and processing failures
• Reproducibility controls specify environments, versions, parameters and random seeds
• Operational standard procedures define review, escalation and model retraining triggers

Participants can communicate AI remote sensing results for technical and non-technical decisions.

• Map layouts communicate extent, date, legend, uncertainty and source provenance
• Executive briefings connect detected patterns with operational actions and priorities
• Accuracy statements present limitations without overstating model certainty or coverage
• Dashboards distinguish monitoring indicators, alerts, trends and confirmed events
• Technical reports document methods, assumptions, validation evidence and reproducibility details
• Decision logs record how imagery findings influenced inspections, funding or interventions

Participants can build, review and present a complete AI remote sensing project for an operational monitoring need.

• Project scope defines decision question, users, area, timing and target classes
• Imagery and reference data selected using documented suitability and licensing criteria
• Preprocessing pipeline produces analysis-ready inputs with recorded quality checks
• Model workflow combines features, labels, training choices and validation design
• Results package includes maps, accuracy evidence, uncertainty and change findings
• Peer review challenges assumptions, errors, governance controls and deployment recommendations