AI Remote Sensing Image Analysis Training Course

Participants can select sensor data appropriate to a defined monitoring question.

• Spatial, spectral, temporal and radiometric resolution compared for use cases
• Sentinel-2 bands matched to vegetation, water and built-up analysis
• Landsat archives selected for long-term environmental change studies
• Sentinel-1 SAR backscatter interpreted for cloud-prone flood monitoring
• Drone orthomosaics assessed for centimetre-scale asset and site inspection
• Elevation models incorporated for slope, drainage and terrain context

Participants can align spatial datasets into an analysis-ready common reference framework.

• Coordinate reference systems selected using area, distortion and output requirements
• Raster grids aligned by projection, extent, cell size and origin
• Resampling methods matched to continuous reflectance and categorical labels
• Co-registration errors identified through image flicker and control features
• Vector boundaries checked against imagery dates, scale and positional accuracy
• Analysis areas masked to exclude irrelevant land, water and cloud regions

Participants can prepare optical and radar imagery for consistent model input.

• Surface reflectance distinguished from top-of-atmosphere reflectance products
• Cloud, shadow and snow masks applied using quality assurance bands
• Atmospheric correction assumptions checked before multi-date optical comparisons
• SAR speckle filtering balanced against edge preservation and feature loss
• Radiometric normalization applied to reduce cross-scene brightness differences
• Preprocessing logs record source, dates, parameters and output versions

Participants can derive and select meaningful geospatial features for supervised models.

• NDVI, NDWI and NDBI calculated from calibrated multispectral bands
• Band ratios selected to separate water, vegetation, soil and impervious surfaces
• GLCM texture measures derived for surface pattern discrimination
• Digital elevation derivatives generated as slope, aspect and curvature
• Multi-temporal composites summarize seasonal phenology and persistent land cover
• Feature collinearity reviewed before training tree-based and linear models

Participants can create representative, traceable training and validation datasets.

• Class taxonomy defined with mutually exclusive labels and decision rules
• Stratified sampling allocates observations across classes and geographic regions
• Polygon labels converted to pixels without boundary contamination
• Reference imagery dates matched to sensor acquisition and target conditions
• Inter-annotator agreement assessed using labelled sample comparison procedures
• Training, validation and test partitions separated by spatial blocks

Participants can inspect image features and labels for quality issues before modelling.

• Histograms reveal outliers, saturation and class-specific spectral distributions
• Scatterplots compare band combinations and identify separable target classes
• Spatial sampling maps expose geographic gaps and clustered labels
• Class balance quantified before selecting weighting or resampling strategies
• Mislabelled samples identified through imagery inspection and neighbourhood comparison
• Baseline rules established using indices, thresholds and simple decision trees

Participants can train and interpret pixel-based supervised classification models.

• Random forest classification trained with spectral, terrain and texture features
• Gradient boosting compared against random forest using held-out spatial data
• Hyperparameters tuned for tree depth, features and minimum samples
• Class weights applied when target categories are rare or costly
• Probability rasters retained alongside hard class assignments for review
• Feature importance interpreted cautiously where predictors are correlated

Participants can classify meaningful image objects rather than individual pixels.

• Multiresolution segmentation groups neighbouring pixels into candidate image objects
• Scale, shape and compactness parameters tested against target object sizes
• Object attributes calculated from spectral, texture, geometry and context measures
• Rule-based classification combines thresholds, adjacency and expert knowledge
• Parcel and building boundaries used to constrain object extraction
• Object-based outputs compared with pixel classifications for salt-and-pepper reduction

Participants can configure deep-learning workflows for remote sensing imagery tasks.

• Convolutional layers explained through kernels, feature maps and receptive fields
• Image tiles generated with overlap, labels and consistent channel ordering
• Data augmentation applies rotation, flips, brightness shifts and random crops
• Transfer learning adapts pretrained encoders to limited labelled imagery
• Batch size, learning rate and epochs controlled during training
• Training curves diagnose overfitting, underfitting and unstable optimization

Participants can detect and assess discrete objects in high-resolution imagery.

• Bounding-box labels prepared for vehicles, roofs, poles and solar panels
• Intersection over Union measures overlap between predicted and reference boxes
• YOLO detection outputs filtered using confidence and non-maximum suppression
• Small-object detection improved through tiling, resolution and anchor configuration
• False detections reviewed by background type, shadow and viewing geometry
• Detected assets summarized by location, count, type and confidence

Participants can produce pixel-level segmentation maps for continuous geographic features.

• Semantic segmentation distinguishes roads, water, buildings and vegetation pixels
• U-Net architecture links encoder features with decoder reconstruction layers
• Segmentation masks encoded for binary, multiclass and overlapping target definitions
• Dice coefficient and Intersection over Union assess segmentation quality
• Boundary errors inspected around narrow roads, roofs and water edges
• Post-processing removes isolated pixels and fills valid interior gaps

Participants can quantify model performance and diagnose geographically patterned errors.

• Confusion matrices report producer, user and overall classification accuracy
• Precision, recall and F1 score selected for imbalanced detection tasks
• Spatial cross-validation prevents inflated scores from nearby training samples
• Area-adjusted accuracy estimates account for unequal class sampling rates
• Error maps reveal terrain, season, region and class-specific weaknesses
• Confidence intervals reported for accuracy estimates and mapped area totals

Participants can identify and validate meaningful land and asset changes through time.

• Image time series assembled using consistent preprocessing and seasonal windows
• Image differencing detects reflectance and index changes between dates
• Change vector analysis combines magnitude and direction across multiple bands
• Post-classification comparison quantifies transitions between mapped land-cover classes
• BFAST-style temporal decomposition separates trend, seasonality and abrupt breaks
• Change candidates verified against clouds, shadows and acquisition differences

Participants can design rapid image-analysis workflows for flood and disaster assessment.

• SAR backscatter thresholds identify open-water inundation beneath cloud cover
• Pre-event and post-event composites isolate disaster-related surface changes
• Permanent water masks reduce false flood detections near rivers
• Terrain and drainage constraints refine flood extent plausibility
• Exposure layers intersect flood maps with population, roads and facilities
• Rapid products label confidence, acquisition dates and known coverage gaps

Participants can design scalable, reproducible workflows for imagery processing and modelling.

• Google Earth Engine collections filtered by date, geometry and metadata
• Server-side processing distinguished from client-side operations and data export
• Python scripts parameterize imagery queries, preprocessing and model inference
• Batch processing tiles large areas while preserving grid consistency
• Version control records code, parameters, training data and model artefacts
• Processing costs estimated from imagery volume, compute time and storage

Participants can prepare an AI image-analysis model for controlled operational use.

• Inference pipelines specify input checks, processing steps and output formats
• Model cards document purpose, training data, performance and limitations
• Data drift monitored through feature distributions and acquisition conditions
• Concept drift assessed when target appearance changes over time
• Human review queues route low-confidence outputs for analyst verification
• Retraining triggers defined using accuracy decline, new labels and policy changes

Participants can apply governance controls to geospatial AI data, outputs and decisions.

• Privacy risks assessed for high-resolution imagery and sensitive location data
• Data licences checked for training, derivative products and redistribution rights
• Bias sources examined across geography, season, settlement type and class representation
• Purpose limitation documented for surveillance-sensitive imagery analysis applications
• Audit trails retain source imagery, labels, code and approval decisions
• Communication protocols distinguish model estimates from verified ground observations

Participants can plan an AI remote sensing project with defensible technical and delivery controls.

• Project charters define decision use, scope, stakeholders and acceptance thresholds
• Data acquisition plans specify sensors, dates, coverage and reference labels
• Model selection criteria balance accuracy, explainability, compute and maintenance
• Quality assurance gates assign checks before release and operational use
• Map narratives explain findings, uncertainty, confidence and recommended actions
• Implementation roadmaps sequence pilot, validation, deployment and monitoring activities

Participants can build, review and present a complete AI remote sensing analysis project from imagery selection through validated operational recommendations.

• Problem statement finalized with target classes, users and decision criteria
• Imagery and reference data assembled into a documented analysis dataset
• Features or image tiles prepared using justified preprocessing choices
• Classification, detection or segmentation model trained and evaluated spatially
• Error analysis produces limitations, confidence guidance and corrective actions
• Final maps, model card and implementation plan presented for review