Deep Learning for Remote Sensing Training Course

Participants can inspect, organise and align geospatial imagery and reference data for model development.

• Coordinate reference systems, map projections and georeferencing error sources
• Raster grids, vector geometries, footprints and pixel coordinate conventions
• GDAL workflows for reading, warping, mosaicking and clipping imagery
• QGIS inspection of metadata, band order, no-data values and extents
• Cloud and shadow masks assessed before downstream image preparation
• Spatial joins connecting labels, parcels, assets and administrative boundaries

Participants can create labelled datasets that support valid geospatial model training and testing.

• Training, validation and test splits separated by space and time
• Spatial autocorrelation and leakage diagnosed in geospatial sampling schemes
• Stratified sampling for rare classes, regions and seasonal conditions
• Polygon-to-pixel label rasterisation with boundary and mixed-pixel rules
• Label taxonomies documented through classes, definitions, exclusions and examples
• Inter-annotator agreement checked using confusion matrices and targeted review

Participants can prepare analysis-ready image patches while retaining geospatial integrity.

• Radiometric calibration from digital numbers to reflectance or backscatter
• Atmospheric correction assumptions documented for multispectral surface-reflectance image products
• Band stacking, resampling and alignment without unintended spectral distortion
• Patch extraction with overlap, padding and geospatial index retention
• Normalisation statistics fitted only on training imagery and labels
• Augmentation policies for rotations, flips, crops, noise and illumination

Participants can structure a reproducible deep-learning project for geospatial imagery.

• Python project structure separating data, models, configuration and outputs
• PyTorch tensors, datasets and data loaders for geospatial image patches
• Reproducible environments pinned with package versions, seeds and hardware settings
• Configuration files controlling paths, classes, hyperparameters and experiment identifiers
• Training logs capturing loss, metrics, checkpoints and data provenance
• GPU memory, batch size and mixed-precision training constraints

Participants can train and diagnose a baseline convolutional neural network for remote-sensing imagery.

• Convolution, pooling, receptive fields and feature hierarchies in image models
• Backpropagation, stochastic gradient descent and Adam optimisation compared
• Learning-rate schedules, weight decay and early stopping criteria
• Cross-entropy, focal loss and class weighting for imbalanced labels
• Baseline CNN classifier trained on labelled multispectral image patches
• Validation curves interpreted for underfitting, overfitting and unstable optimisation

Participants can build and assess a multispectral land-cover classification model.

• Multispectral band selection guided by target materials and sensor physics
• Transfer learning with pretrained encoders and adapted input channels
• Patch-level land-cover classification with probability calibration and class thresholds
• Confusion matrices, producer accuracy and user accuracy for mapped classes
• Macro F1, balanced accuracy and Cohen's kappa selected appropriately
• Misclassified patches reviewed against labels, boundaries and acquisition conditions

Participants can create dense pixel-level maps using semantic-segmentation architectures.

• Semantic segmentation outputs converted into georeferenced class probability rasters
• U-Net encoder-decoder architecture selected for dense pixel labelling
• DeepLabV3+ atrous convolutions for multi-scale context modelling
• Dice loss and intersection-over-union for segmentation evaluation
• Tile-edge artefacts reduced through overlap, padding and blending
• Building footprints and crop parcels extracted from segmentation masks

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

• Object detection distinguishes instances, classes and bounding-box localisation
• YOLO annotation format, anchor-free heads and non-maximum suppression
• Intersection-over-union thresholds and mean average precision interpretation
• Small-object detection improved through resolution, tiling and scale augmentation
• Oriented bounding boxes for ships, vehicles and elongated assets
• Detected assets filtered using confidence scores and spatial business rules

Participants can design deep-learning workflows for multi-date imagery and change mapping.

• Multi-date image stacks aligned across sensors, seasons and acquisition geometries
• Phenological signatures modelled from vegetation indices and reflectance sequences
• Temporal CNNs and recurrent networks for sequence classification
• Change-detection labels defined with reference dates and persistence rules
• Bi-temporal Siamese networks compare before-and-after image representations
• False change caused by clouds, illumination and registration error

Participants can incorporate SAR imagery and fused sensor inputs into deep-learning workflows.

• SAR amplitude, phase, polarisation and speckle interpreted for modelling
• Sentinel-1 preprocessing: orbit files, calibration, terrain correction and filtering
• Speckle filters balanced against edge preservation and target visibility
• Optical-SAR fusion architectures combine complementary sensor evidence
• Flood extent mapping handles cloud cover using SAR backscatter
• Cross-sensor co-registration verified before pixel-level fusion

Participants can assess transformer-based approaches against convolutional baselines for imagery tasks.

• Vision Transformers partition imagery into patches and positional embeddings
• Self-attention captures long-range spatial relationships between image regions
• Foundation-model embeddings adapted with limited labelled remote-sensing data
• Fine-tuning strategies compare frozen encoders, adapters and full updates
• Compute budgets, sample size and accuracy guide architecture selection
• Transformer predictions compared against CNN baselines using consistent splits

Participants can produce defensible accuracy evidence for geospatial deep-learning outputs.

• Spatial cross-validation blocks prevent optimistic estimates from nearby samples
• Temporal holdouts test performance under new seasons and acquisition dates
• Area-adjusted accuracy estimation weights samples by mapped class extent
• Confidence intervals reported for accuracy estimates and area calculations
• Error maps reveal geographic clusters, boundary failures and regional bias
• Independent reference data assessed for lineage, date and positional quality

Participants can assess model confidence and communicate uncertainty for operational decisions.

• Softmax probabilities distinguished from calibrated estimates of prediction confidence
• Temperature scaling and reliability diagrams assess probability calibration
• Monte Carlo dropout estimates epistemic uncertainty in image predictions
• Entropy maps identify pixels requiring review or additional reference labels
• Grad-CAM saliency inspected for spurious spectral and contextual cues
• Human review rules combine uncertainty thresholds with operational risk

Participants can convert model predictions into usable map products and operational layers.

• Sliding-window inference generates wall-to-wall predictions over large scenes
• Cloud-Optimized GeoTIFFs store tiled outputs for efficient map access
• Raster-to-vector conversion applies thresholds, smoothing and minimum mapping units
• Post-processing rules enforce topology, adjacency and domain-specific constraints
• Web map layers publish predictions, confidence surfaces and metadata
• Prediction outputs linked to asset registers and field inspection queues

Participants can define controls for reproducible model delivery, monitoring and retraining.

• Data versioning records source scenes, labels, preprocessing and split definitions
• Model cards document intended use, limitations, metrics and training data
• Experiment tracking compares parameters, code revisions and validation results
• Data drift monitors sensor changes, seasonality and geographic expansion
• Retraining triggers defined from error rates, drift signals and business impact
• Access controls protect imagery, labels and sensitive asset locations

Participants can address legal, ethical and assurance requirements for remote-sensing AI services.

• Licensing terms checked for imagery redistribution, derivatives and commercial use
• Personal data risks assessed for high-resolution imagery and linked records
• Bias analysis tests representation across regions, land types and communities
• Human oversight requirements matched to decision consequence and uncertainty
• Audit trails preserve model versions, inputs, outputs and reviewer actions
• Procurement specifications state accuracy, latency, documentation and acceptance tests

Participants can plan a remote-sensing deep-learning initiative around measurable business value.

• Use-case scoping connects remote-sensing outputs to decisions, owners and timing
• Baseline workflows quantified for cost, delay, coverage and error rates
• Value hypotheses specify avoided inspections, prioritised visits or earlier alerts
• Data acquisition plans identify sensors, licences, refresh rates and costs
• Acceptance tests define geographic scope, accuracy thresholds and delivery formats
• Implementation roadmaps assign pilots, approvals, monitoring and scale decisions

Participants can review a geospatial AI pipeline and communicate its evidence, limitations and decisions.

• End-to-end pipeline architecture connects ingestion, preprocessing, training, inference and publication
• Failure-mode analysis ranks data, model, infrastructure and adoption risks
• Peer review criteria assess reproducibility, validity, usability and governance
• Technical reports communicate methods, metrics, limitations and operational recommendations
• Map layouts present classes, confidence, scale bars and source attribution
• Stakeholder briefings distinguish model capability, uncertainty and decision responsibility

Participants can build, review and present a complete deep-learning remote-sensing solution with mapped outputs, validation evidence and an implementation plan.

• Problem brief defines user decision, area, classes, sensors and success metrics
• Geospatial dataset assembled with documented labels, splits and preprocessing choices
• Selected model trained, tuned and evaluated against spatially independent reference data
• Prediction map produced with uncertainty layers and post-processing rules
• Model card records provenance, performance, limitations and intended operational use
• Review presentation defends technical choices, risks, deployment and monitoring plan