Core Concepts of AI/ML for Earth Observation

CoPhil EO AI/ML Training - Day 1, Session 2

Stylianos Kotsopoulos

EU-Philippines CoPhil Programme

Welcome to Session 2

Session Objectives

• Understand what AI/ML means in Earth Observation context
• Learn the end-to-end workflow for ML projects
• Distinguish supervised vs unsupervised learning
• Grasp deep learning and neural network basics
• Explore 2025 AI innovations (foundation models, data-centric AI)

Duration: 2 hours

This session covers fundamental AI/ML concepts tailored to EO applications. Mostly conceptual, but essential foundation for hands-on work in Sessions 3-4 and throughout Day 2-4.

Session Roadmap

Time	Topic	Duration
00-10 min	What is AI/ML?	10 min
10-35 min	EO Workflow & Data Pipeline	25 min
35-60 min	Supervised vs Unsupervised Learning	25 min
60-65 min	☕ Break	5 min
65-90 min	Deep Learning & Neural Networks	25 min
90-110 min	Data-Centric AI & 2025 Updates	20 min
110-120 min	Q&A & Summary	10 min

Timing: 2 minutes

This session is more conceptual than Session 1. Focus on building intuition and mental models. Hands-on practice comes in Sessions 3-4 and Days 2-4.

Why AI/ML for Earth Observation?

Traditional Approach

• Manual interpretation
• Rule-based classification
• Simple thresholds
• Time-consuming
• Hard to scale

AI/ML Approach

• Automated pattern recognition
• Learn from examples
• Complex decision boundaries
• Fast processing
• Scalable to large areas

ML can process years of satellite data in hours!

Machine learning allows us to automatically recognize patterns in satellite imagery without hard-coding rules for every scenario. This is transformative for large-area, time-series analysis.

What is AI/ML?

Defining the Terms

AI, ML, and Deep Learning Hierarchy with EO Applications

• Artificial Intelligence (AI): Broad field of making machines “smart”
• Machine Learning (ML): Subset of AI where algorithms learn from data
• Deep Learning (DL): Subset of ML using neural networks with many layers

AI is the broadest term. Machine Learning is a subset where computer algorithms learn patterns from data without being explicitly programmed. Deep Learning uses neural networks. This comprehensive diagram shows the nested hierarchy with specific algorithms at each level (Random Forest, CNNs, RNNs, etc.) and their Earth Observation applications like land cover classification and time series analysis.

Machine Learning in Simple Terms

Traditional Programming

Rules + Data → Results

• Programmer writes explicit rules
• Fixed logic
• Hard to handle complexity

Machine Learning

Data + Results → Rules

• Algorithm learns rules from examples
• Adaptive
• Handles complex patterns

In traditional programming, we tell computers what to do step-by-step. In ML, we show examples and the algorithm figures out the pattern.

ML in Earth Observation Context

Example: Forest vs Non-Forest

Traditional:

IF NDVI > 0.6 THEN Forest
ELSE Non-Forest

Simple, but breaks easily

Machine Learning:

• Show 1000 examples of forest pixels
• Show 1000 examples of non-forest
• Algorithm learns complex patterns
• Works in diverse conditions

A simple NDVI threshold might work in one region but fail in another. ML can learn the nuanced patterns that distinguish forest from non-forest across different conditions.

The AI/ML Workflow for EO

End-to-End ML Workflow

–>

This is the typical workflow for any ML project in Earth Observation. Understanding these steps is crucial for successful implementation.

Step 1: Problem Definition

Key Questions

• What exactly are we trying to achieve?
• What decisions will this support?
• What level of accuracy is needed?
• What resources are available?

EO Examples

• Map rice paddy extent
• Detect flooded areas after typhoon
• Classify land cover types
• Estimate crop yield
• Monitor deforestation

Clear problem definition = 50% of success

Being clear on the question helps design the solution. “We want to classify land cover in Palawan” is much more actionable than “We want to use AI.”

Step 2: Data Acquisition

Satellite Imagery

• Sentinel-1/2 (covered in Session 1!)
• Landsat
• Planet
• High-resolution commercial
• Multiple dates/seasons

Ground Truth / Labels

• Field surveys
• GPS points
• Existing maps
• Photo interpretation
• Expert knowledge

Challenge: Getting quality labels is often hardest part

Data acquisition includes both satellite images and the ground truth labels needed to train supervised models. The quality and quantity of labels directly impact model performance.

Step 3: Data Preprocessing

For Satellite Imagery:

• Atmospheric correction (use Level-2A!)
• Cloud masking
• Geometric correction
• Radiometric calibration
• Co-registration (multiple sensors)
• Temporal compositing

“Garbage In, Garbage Out” - preprocessing matters!

Well-prepared input data is crucial. Even the best model will fail if fed cloudy, misaligned, or uncorrected images.

Preprocessing Example

Cloud Removal Before and After Comparison

Before Preprocessing: - Clouds present - Atmospheric haze - Different acquisition dates

After Preprocessing: - Clouds masked - Atmospherically corrected - Temporal composite created

Preprocessing transforms raw satellite data into analysis-ready products. This side-by-side comparison shows the dramatic improvement from cloud masking and creating temporal composites - essential steps before any ML analysis.

Step 4: Feature Engineering

What are Features?

• Input variables for the model
• Derived from raw data
• Informative for the task

EO Features

• Spectral bands (Blue, Red, NIR, etc.)
• Spectral indices (NDVI, NDWI)
• Texture measures
• Temporal statistics
• Topography (elevation, slope)

Deep Learning: Often learns features automatically!

For traditional ML like Random Forest, we engineer features. For deep learning, the network learns features automatically from raw pixels.

Common EO Features

Feature Type	Examples	What They Capture
Spectral Bands	B2, B3, B4, B8	Reflectance at different wavelengths
Vegetation Indices	NDVI, EVI, SAVI	Vegetation health, density
Water Indices	NDWI, MNDWI	Water presence, moisture
Texture	GLCM variance, entropy	Spatial patterns
Temporal	Mean, std over time	Phenology, seasonality
Topographic	Elevation, slope, aspect	Terrain characteristics

Different features highlight different aspects of the landscape. Vegetation indices emphasize green biomass, water indices highlight water bodies, etc.

Step 5: Model Selection & Training

Model Selection

Choose based on:

• Problem type (classification vs regression)
• Data size
• Interpretability needs
• Computational resources

Common EO Models

• Random Forest
• Support Vector Machines
• Convolutional Neural Networks
• U-Net (segmentation)
• Recurrent networks (time series)

Model choice depends on your specific problem, available data, and resources. We’ll cover Random Forest on Day 2 and CNNs on Day 3.

Training Process

–>{fig-align=“center” width=“75%”}

1. Split data: Training set (70-80%) & Validation set (20-30%)
2. Feed training data to model
3. Model learns patterns by adjusting internal parameters
4. Validate on unseen validation data
5. Iterate: Adjust model or data if needed

Training involves feeding labeled examples to the model. The model adjusts its internal parameters to minimize errors on the training data.

Step 6: Validation & Evaluation

Why Validate?

• Ensure model generalizes
• Detect overfitting
• Compare different models
• Build confidence

Evaluation Metrics

• Overall Accuracy
• Confusion Matrix
• Precision & Recall
• F1-Score
• Kappa coefficient

Use independent test data - never validate on training data!

Rigorous validation using held-out data ensures the model works on new, unseen examples - not just memorizing training data.

Confusion Matrix Example

What it shows:

• True Positives (correct predictions)
• False Positives (type I error)
• False Negatives (type II error)
• True Negatives

Derived Metrics:

• Precision = TP / (TP + FP)
• Recall = TP / (TP + FN)
• Accuracy = (TP + TN) / Total

The confusion matrix shows where your model is making mistakes. This helps identify which classes are being confused.

Step 7: Deployment

Deployment Options

• Generate full maps
• Near real-time monitoring
• Operational pipelines
• Decision support systems
• Web applications

Considerations

• Model retraining schedule
• Computational requirements
• User interface
• Data updates
• Maintenance plan

If the model is satisfactory, deploy it for operational use. This might mean generating maps for entire regions or setting up automatic processing of new satellite images.

Workflow is Iterative

–>{fig-align=“center” width=“70%”}

• Poor validation? → Go back to data acquisition or model selection
• New data available? → Retrain model
• Requirements change? → Redefine problem
• Continuous improvement is key

Real projects are iterative. You often loop back: if validation is poor, you might need more data, different features, or a different model.

Types of Machine Learning

Main ML Paradigms

–>{fig-align=“center” width=“75%”}

1. Supervised Learning (most common in EO)
2. Unsupervised Learning (exploratory analysis)
3. Semi-supervised Learning (combines both)
4. Reinforcement Learning (less common in EO)

We’ll focus on supervised and unsupervised learning as these are most relevant for Earth Observation applications.

Supervised Learning

What is Supervised Learning?

Definition

• Learning from labeled data
• Known input-output pairs
• Model learns mapping from inputs to outputs
• Like learning with an answer key

–>{width=“100%”}

Requires ground truth labels for training

Supervised learning is the most common in EO. The algorithm is given examples with known outcomes (labels) and learns to predict labels for new, unseen data.

Two Types of Supervised Learning

Classification

• Predict categorical labels
• Discrete classes
• Example outputs: “Forest”, “Water”, “Urban”

Regression

• Predict continuous values
• Numeric outputs
• Example outputs: 25.3 tons/hectare, 15.2°C

Classification assigns data to categories. Regression predicts numeric values. Both require labeled training data.

Classification Examples in EO

Land Cover Classification

• Forest, agriculture, urban, water
• Pixel-wise or object-based
• Multi-class problem

Crop Type Mapping

• Rice, corn, sugarcane
• Seasonal patterns important
• Supports agricultural planning

Land cover classification is the classic EO supervised learning task. Each pixel or region is assigned to a class like forest, water, or urban.

Regression Examples in EO

Biomass Estimation

• Predict tons of biomass per hectare
• Important for carbon accounting
• Uses SAR and optical data

Crop Yield Prediction

• Predict tons per hectare
• Seasonal NDVI time series
• Supports food security planning

Regression tasks predict continuous values like biomass density, crop yield, soil moisture, or sea surface temperature from satellite data.

Common Supervised Algorithms

Algorithm	Strengths	EO Applications
Random Forest	Handles high dimensions, robust	Land cover, crop classification
SVM	Effective in high dimensions	Binary classification, change detection
Neural Networks	Learns complex patterns	Image classification, segmentation
Decision Trees	Interpretable	Quick classifications
k-NN	Simple, non-parametric	Local classifications

Different algorithms have different strengths. Random Forest is popular in EO for its robustness and ability to handle many features.

Supervised Learning Requirements

Essential:

1. Training data with known labels
2. Representative samples covering all classes
3. Sufficient quantity (varies by algorithm)
4. Quality labels (accurate, consistent)
5. Independent validation data

Challenge: Getting quality labels is often the bottleneck!

Supervised learning needs ground truth. For land cover, this might be field surveys, GPS points, or careful photo interpretation. Quality matters more than quantity!

Unsupervised Learning

What is Unsupervised Learning?

Definition

• Learning from unlabeled data
• No known outputs
• Discover hidden patterns
• Like sorting without instructions

–>{width=“100%”}

Useful for exploratory analysis and finding structure

Unsupervised learning finds patterns or groupings inherent in the data without being told what to look for.

Clustering: Main Unsupervised Technique

–>{fig-align=“center” width=“70%”}

• Group similar pixels based on spectral characteristics
• Algorithm decides number of clusters (or you specify)
• Analyst interprets what each cluster means
• Example: “Cluster 3 looks like water, Cluster 7 looks like forest”

K-means clustering is a common unsupervised method. It groups pixels with similar reflectance, but you have to interpret what those groups mean.

Unsupervised EO Applications

Change Detection

• Cluster “before” and “after” images
• Identify changed areas
• No labels needed

Anomaly Detection

• Find unusual pixels
• Potential forest disturbance
• Data quality issues

Initial Exploration

• Quick overview of spectral classes
• Inform supervised approach
• Generate training samples

Dimensionality Reduction

• PCA, t-SNE
• Visualize high-dimensional data
• Feature extraction

Unsupervised methods are useful for quick initial analysis or when you don’t have ground truth labels. Results need interpretation though.

Supervised vs Unsupervised

Aspect	Supervised	Unsupervised
Labels	Required	Not needed
Accuracy	Generally higher	Lower, needs interpretation
Use Case	Precise classification	Exploration, pattern discovery
Effort	High (collecting labels)	Low (no labels)
Output	Predefined classes	Discovered clusters
Control	High (you define classes)	Low (algorithm decides groups)

Supervised methods generally yield more accurate results when good training data is available. Unsupervised is useful when labels are unavailable or for exploratory work.

Which to Choose?

Use Supervised When:

• You have ground truth labels
• Need specific classes
• Accuracy is critical
• Operational application

Use Unsupervised When:

• No labels available
• Exploratory analysis
• Discovering unknown patterns
• Quick initial assessment

In practice: Often combine both approaches!

Most operational EO applications use supervised learning because accuracy and specific class definitions are important. Unsupervised helps with initial exploration.

☕ 5-Minute Break

Stretch Break

Stand up • Grab water • Back in 5 minutes

Timing: 5 minutes

Instructor Actions: - Announce break - Mention we’ll dive into deep learning next - Be available for quick questions

When Resuming: - Quick recap: “We’ve covered ML basics, workflow, supervised/unsupervised. Now: deep learning!”

Introduction to Deep Learning

What is Deep Learning?

Deep Learning = Neural Networks with Many Layers

• Subset of machine learning
• “Deep” refers to multiple layers
• Automatically learns features
• Excels at image analysis
• Data-hungry

–>{width=“100%”}

Deep learning is essentially about neural networks with many layers (dozens or even hundreds). These “deep” networks can capture very complex relationships.

Neural Networks: Building Blocks

–>{fig-align=“center” width=“70%”}

Artificial Neuron:

1. Takes multiple inputs
2. Multiplies each by a weight
3. Adds a bias
4. Applies activation function
5. Produces output

A single neuron is like a logistic regression unit. It takes inputs, applies weights, and uses an activation function to produce an output.

Neural Network Architecture

–>{fig-align=“center” width=“80%”}

Layers:

• Input Layer: Receives data (e.g., pixel values)
• Hidden Layers: Process and transform
• Output Layer: Final prediction

Connections:

• Each neuron connects to next layer
• Weights on connections
• Information flows forward

Neurons are organized into layers. The input layer receives data (pixel values), hidden layers progressively extract features, and the output layer makes predictions.

Key Concepts

Activation Functions

• Introduce non-linearity
• Common: ReLU, Sigmoid, Tanh
• Allow network to learn complex patterns

Weights and Biases

• Parameters the network learns
• Millions of parameters in deep networks
• Adjusted during training

Forward Propagation

• Data flows input → output
• Generate prediction

Activation functions are crucial - they allow neural networks to learn non-linear relationships. Without them, the network would just be linear regression!

How Neural Networks Learn

–>{fig-align=“center” width=“75%”}

1. Forward pass: Input data, get prediction
2. Calculate loss: How wrong is the prediction?
3. Backpropagation: Calculate gradients
4. Update weights: Adjust to reduce error
5. Repeat: Thousands of times (epochs)

Training adjusts weights to minimize error. This happens through backpropagation - computing gradients and updating weights in the direction that reduces loss.

Loss Functions

Classification

Cross-Entropy Loss

• Measures classification error
• Higher penalty for confident wrong predictions
• Standard for multi-class problems

Regression

Mean Squared Error

\[MSE = \frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2\]

• Measures prediction error
• Squared difference from true value

Loss functions quantify “how bad” predictions are. The training process tries to minimize this loss by adjusting weights.

Optimizers

Stochastic Gradient Descent (SGD)

• Basic optimizer
• Updates weights based on gradients
• Learning rate controls step size

Adam Optimizer

• Adaptive learning rates
• Faster convergence
• Most popular for deep learning
• Generally works well

You don’t need to implement these - frameworks do it for you!

Optimizers determine how weights are updated. Adam is the most popular because it adapts learning rates automatically and generally converges faster than SGD.

Convolutional Neural Networks (CNNs)

Specialized for images:

• Convolutional layers: Detect spatial patterns
• Pooling layers: Reduce dimensionality
• Fully connected layers: Final classification
• Automatically learn features (edges, textures, objects)

CNNs are neural networks specialized for grid data like images. They use convolutional layers to automatically extract spatial features.

How CNNs Process Images

–>{fig-align=“center” width=“80%”}

Hierarchical Feature Learning:

• Early layers: Detect edges, simple patterns
• Middle layers: Detect textures, parts
• Later layers: Detect objects, scenes
• No manual feature engineering needed!

CNNs learn increasingly complex features at each layer. Early layers detect edges, later layers detect whole objects. This happens automatically during training!

CNNs in Earth Observation

Applications:

• Image classification
• Object detection (ships, buildings)
• Semantic segmentation (pixel-wise)
• Change detection
• Super-resolution

Advantages:

• Learn features automatically
• Handle spatial context
• State-of-the-art performance
• Transfer learning possible

CNNs have achieved state-of-the-art results in many EO tasks. They can learn to recognize complex patterns without manual feature engineering.

Popular CNN Architectures for EO

Architecture	Year	Key Innovation	EO Use Cases
ResNet	2015	Residual connections	Classification, backbone for detection
U-Net	2015	Skip connections	Semantic segmentation, flood mapping
EfficientNet	2019	Compound scaling	Efficient classification, mobile deployment
DeepLabv3+	2018	Atrous convolution	Land cover segmentation
YOLOv8	2023	Real-time detection	Object detection, ship/vehicle counting

ResNet and EfficientNet are most popular backbones for EO

These are proven architectures widely used in EO. ResNet-50 is often the starting point for transfer learning. U-Net dominates semantic segmentation tasks.

ResNet: Residual Networks

–>{fig-align=“center” width=“70%”}

Key Innovation: Skip Connections

• Allows training very deep networks (50, 101, 152 layers)
• Solves vanishing gradient problem
• Identity mapping preserves information

Common Variants: - ResNet-50 (25M parameters) - ResNet-101 (44M parameters) - ResNet-152 (60M parameters)

EO Applications:

• Pre-trained on ImageNet
• Fine-tune for EO tasks
• Backbone for object detection
• Transfer learning baseline

Performance: - Top-5 error: 3.57% (ImageNet) - Works well with 10k+ images

ResNet revolutionized deep learning by enabling training of very deep networks. Skip connections allow gradients to flow directly through the network.

U-Net for Semantic Segmentation

Architecture: - Encoder (contracting path): Captures context - Decoder (expanding path): Enables precise localization - Skip connections: Combine low & high-level features

Why Dominant in EO: - Works with small datasets (hundreds of images) - Precise pixel-wise predictions - Perfect for segmentation tasks

EO Applications: Flood mapping, land cover, building footprints, crop fields

U-Net is THE architecture for semantic segmentation in EO. Originally designed for biomedical image segmentation, it’s now standard for pixel-wise classification tasks.

Deep Learning Frameworks

TensorFlow / Keras

–>{width=“150px”}

• Google’s framework
• High-level Keras API
• Production-ready
• Large ecosystem

PyTorch

–>{width=“150px”}

• Facebook’s framework
• Pythonic and intuitive
• Popular in research
• Flexible

We’ll use TensorFlow/Keras in this training

You don’t implement backpropagation yourself - frameworks like TensorFlow and PyTorch handle the math. You just define the architecture and provide data.

Deep Learning Considerations

Advantages:

• Automatic feature learning
• State-of-the-art accuracy
• Handles complex patterns
• Scales to big data

Challenges:

• Requires lots of training data
• Computationally intensive (need GPUs)
• Less interpretable (“black box”)
• Harder to debug

Start simple (Random Forest), move to DL when you have data and compute

Deep learning is powerful but data-hungry and computationally expensive. For many EO tasks, simpler models like Random Forest work well with less data.

Benchmark Datasets for EO

Why Benchmark Datasets Matter

1. Standardized Evaluation - Compare algorithms objectively
2. Training Resources - Pre-labeled data for model training
3. Transfer Learning - Pre-train on large datasets, fine-tune locally
4. Research Reproducibility - Enable comparison across studies
5. Community Building - Shared resources accelerate progress

You don’t need to label everything from scratch!

Benchmark datasets are crucial for EO ML. They provide labeled training data and enable fair comparison of methods across research groups worldwide.

EuroSAT: Land Cover Classification

Specifications: - Images: 27,000 labeled patches - Classes: 10 land cover types - Size: 64×64 pixels - Bands: All 13 Sentinel-2 bands - Source: European cities

10 Classes: Annual Crop • Forest • Herbaceous Vegetation • Highway • Industrial • Pasture • Permanent Crop • Residential • River • Sea/Lake

–>{width=“100%”}

Achievement: 98.57% accuracy with CNNs

Why Popular: - Sentinel-2 based - Balanced classes - Easy to use

EuroSAT is one of the most popular benchmarks for EO classification. Based on Sentinel-2, making it highly relevant for operational applications. Great starting point for CNN experiments.

BigEarthNet: Large-Scale Multi-Label

Massive Scale: - Images: 590,326 Sentinel-2 patches - Coverage: 10 European countries - Labels: 43 land cover classes - Multi-label: Multiple classes per image - Multi-modal: Optical + SAR version

Real-World Complexity: - Forest + Water - Urban + Agricultural - Reflects actual landscapes

Why Different:

Unlike EuroSAT (single label), BigEarthNet has multiple overlapping classes - more realistic!

Access: - bigearth.net - TensorFlow Datasets - Papers With Code

BigEarthNet’s multi-label nature makes it more challenging but also more realistic. Essential for semantic segmentation research and testing advanced architectures.

xView: Object Detection Benchmark

Specifications: - Objects: >1 million annotated - Classes: 60 object types - Resolution: 0.3m (WorldView-3) - Area: >1,400 km² - Annotations: Bounding boxes

Object Categories: - Buildings & infrastructure - Vehicles (cars, trucks, aircraft) - Ships & maritime - Storage tanks - Construction equipment

Created for disaster response

Applications: - YOLO training - Faster R-CNN - Small object detection - Infrastructure mapping

xView is THE benchmark for object detection in satellite imagery. Created for disaster response applications, now widely used for testing detection algorithms like YOLO and Faster R-CNN.

Philippine Data Resources

PRiSM (PhilRice) - Rice area maps (wet/dry season) - Planting dates & growth stages - Yield estimates - Since 2014 - https://prism.philrice.gov.ph/

PhilSA Products - Flood extent maps (DATOS) - Mangrove extent mapping - Land cover classifications - Disaster damage assessments

DOST-ASTI Outputs - DATOS rapid flood mapping - Hazard susceptibility maps - AI-powered damage assessment - hazardhunter.georisk.gov.ph

NAMRIA Geoportal - National land cover (2020) - Topographic basemaps - Administrative boundaries - Digital Elevation Models - www.geoportal.gov.ph

Use these as training/validation data - don’t start from scratch!

Philippine agencies have produced operational EO products that can serve as training or validation data for your ML models. Leverage existing work!

Data-Centric AI & 2025 Innovations

Paradigm Shift: Model-Centric vs Data-Centric

–>{fig-align=“center” width=“80%”}

Model-Centric (Traditional)

• Focus on improving algorithms
• Keep data fixed
• Try different models
• Tune hyperparameters

Data-Centric (Modern)

• Focus on improving data
• Keep model fixed
• Clean and augment data
• Better annotations

A lot of early ML progress focused on model algorithms. Data-Centric AI, popularized by Andrew Ng, advocates that improving your data often yields bigger gains.

Why Data-Centric Matters for EO

EO-Specific Data Challenges:

• Cloud contamination
• Atmospheric effects
• Sensor artifacts and noise
• Label uncertainty
• Geographic variability
• Temporal dynamics
• Class imbalance

“Better data beats a cleverer model” in most cases

In EO, the “food” you feed your AI matters more than fancy model tweaks. Cloudy images, mislabeled points, or biased samples can derail any model.

2025 Research: Data Efficiency

–>{fig-align=“center” width=“70%”}

Key Finding (ArXiv 2025):

• Some EO datasets reach optimal accuracy with <20% of temporal instances
• Single band from single modality can be sufficient
• Data efficiency crucial for operational systems
• Quality over quantity

Recent research shows you don’t always need all available data. Smart selection of temporal instances and bands can achieve similar accuracy with much less data.

Four Pillars of Data-Centric AI

1. Data Quality

• Cloud/shadow removal
• Atmospheric correction
• Sensor calibration
• Geometric accuracy

2. Data Quantity

• Sufficient training samples
• Balanced classes
• Data augmentation
• Transfer learning

Four Pillars (Continued)

3. Data Diversity

• Multiple seasons
• Different regions
• Various conditions
• Class variations

4. Label Quality

• Clear definitions
• Consistent protocols
• Expert validation
• Accurate geolocation

These four aspects - quality, quantity, diversity, and labels - determine model success more than architectural choices.

Data Quality

Data Quality in EO

Common Issues:

• Clouds and shadows
• Haze and aerosols
• Sensor artifacts (striping, banding)
• Geometric misalignment
• Radiometric inconsistencies
• Mixed pixels at boundaries

Solutions:

• Use Level-2A products
• Rigorous cloud masking
• Quality flag filtering
• Multi-temporal compositing
• Validation checks
• Document preprocessing

Satellite data can be noisy. Cloud masking, using atmospherically corrected products, and careful preprocessing are essential.

Quality Example: Cloud Masking

Without Cloud Masking

• Clouds misclassified
• Shadows cause errors
• Poor model performance

With Proper Masking

• Clean training data
• Accurate classifications
• Better generalization

One cloudy image can ruin your training data!

Even a few cloudy training samples can teach the model wrong patterns. Rigorous cloud masking is non-negotiable.

Data Quantity

How Much Data Do You Need?

Depends on:

• Model complexity (DL needs more)
• Problem difficulty
• Class separability
• Available features

General Guidelines:

• Traditional ML: 100s to 1000s of samples per class
• Deep Learning: 1000s to 10,000s per class
• Transfer Learning: Can work with 100s per class

Deep learning is data-hungry. Random Forest can work with smaller datasets. Transfer learning (starting from pre-trained models) reduces data requirements.

Data Augmentation

Techniques:

• Rotation (90°, 180°, 270°)
• Flipping (horizontal, vertical)
• Brightness/contrast adjustment
• Adding noise
• Elastic deformations

Result: 10x more training samples from existing data!

Data augmentation synthetically increases dataset size by creating modified versions of existing samples. This helps models generalize better.

Transfer Learning

–>{fig-align=“center” width=“75%”}

Concept:

• Start with model pre-trained on large dataset
• Fine-tune on your specific task
• Requires much less data

EO Applications:

• Use ImageNet pre-trained models
• NASA-IBM Geospatial Foundation Model
• Domain-specific pre-training

Transfer learning leverages knowledge learned on large datasets and adapts it to your specific problem with much less data.

Data Diversity

Why Diversity Matters

Problem: Biased Training

• All samples from one season
• One geographic region only
• Similar conditions
• Result: Model fails elsewhere

Solution: Diverse Training

• Multiple seasons
• Different regions
• Various conditions
• Result: Model generalizes

Models trained on narrow datasets often fail when deployed in different conditions. Diversity in training data leads to better generalization.

Sources of Diversity Needed

Temporal Diversity:

• Different seasons (wet/dry)
• Multiple years
• Phenological stages

Geographic Diversity:

• Different regions
• Various elevations
• Coastal vs inland

Atmospheric Diversity:

• Clear vs hazy days
• Different solar angles
• Seasonal lighting

Class Diversity:

• Variations within classes
• Edge cases
• Transitional zones

For robust models, ensure your training data covers the range of conditions the model will encounter in operational use.

Example: Urban Classification

Poor Diversity

• Only Metro Manila samples
• Only concrete roofs
• Only high-density areas
• Fails in other cities

Good Diversity

• Large cities, small towns
• Various roof materials (concrete, metal, nipa)
• Different architectural styles
• Different densities
• Works across Philippines

If training only on Metro Manila, the model might not recognize small towns or rural settlements with different characteristics.

Label Quality

Label Quality is Critical

Common Label Issues:

• Mislabeled samples
• Positional errors (GPS drift)
• Temporal mismatch (old labels, new image)
• Ambiguous classes
• Inconsistent definitions
• Mixed pixels

Impact:

• Model learns wrong patterns
• Contradictory signals
• Poor generalization
• Low confidence predictions
• Wasted compute

One bad label can corrupt model learning!

Ground truth labels might have errors - GPS inaccuracy, outdated information, or human mistakes. These errors propagate to model predictions.

Label Quality Best Practices

1. Clear Class Definitions

• Write explicit criteria
• Provide examples
• Define edge cases
• Document ambiguities

2. Consistent Protocols

• Standard operating procedures
• Same interpretation rules
• Calibration sessions
• Regular training for labelers

3. Multiple Annotators

• Independent labeling
• Compare for consistency
• Resolve disagreements
• Build consensus labels

4. Expert Validation

• Domain experts review samples
• Random quality checks
• Iterative improvement

Invest time in defining classes clearly and training labelers. Consistency matters more than speed.

Label Quality Example

Poor Labels

• “Forest” defined inconsistently
• Mixed with shrubland
• Temporal mismatch
• Positional errors

High-Quality Labels

• Clear forest definition
• Careful boundary delineation
• Image-label temporal match
• Validated position

High-quality labels are worth the effort. A smaller dataset with accurate labels often outperforms a larger dataset with noisy labels.

ALaM Project: Addressing Labels

DOST-ASTI’s Automated Labeling Machine

• Automates labeling process
• Crowdsourcing capabilities
• Expert validation workflow
• Addresses EO’s biggest bottleneck

Remember from Session 1: DOST-ASTI’s ALaM project specifically addresses the label quality and quantity challenge through automation and crowdsourcing.

Practical Data-Centric Tips

Data-Centric Workflow

Before Training:

1. Audit your data: Visualize samples, check distributions
2. Clean aggressively: Remove clouds, fix labels, filter outliers
3. Balance classes: Address imbalances through sampling or augmentation
4. Document everything: Track data sources, preprocessing, versions

During Training:

5. Analyze errors: Which samples does model get wrong?
6. Identify patterns: Are errors systematic? (e.g., all in one region)
7. Fix data: Add more diverse samples, improve labels
8. Iterate: Retrain with better data

Data-centric approach means continuously improving data quality based on model feedback. Look at errors to understand what data you’re missing.

Data Quality Checklist

• Atmospherically corrected (Level-2A)?
• Clouds and shadows masked?
• Geometric alignment verified?
• Temporal consistency checked?
• Label accuracy validated?
• Classes clearly defined?
• Training data balanced?
• Geographic diversity ensured?
• Seasonal coverage adequate?
• Edge cases included?
• Quality flags documented?

Use this checklist before training any model. Addressing data issues upfront saves time and improves results.

Case Study: Better Data = Better Results

Scenario:

Coral reef mapping project

Initial Results:

• 70% accuracy
• Fails in turbid water
• Confuses reef with sand

Problem Identified:

All training data from clear water

Data-Centric Solution:

1. Add turbid water samples
2. Include reef-sand transition zones
3. More diverse depths
4. Improve label precision

New Results:

• 90% accuracy
• Works in turbid water
• Better boundary detection

10x improvement from better data, same model!

Real example of how data improvements had bigger impact than model tuning. The data was the key, not the algorithm.

2025 Developments

Foundation Models for EO

–>{fig-align=“center” width=“75%”}

What are Foundation Models?

• Large models pre-trained on massive EO datasets
• Learn general representations
• Fine-tune for specific tasks
• Dramatically reduce labeled data needs

Examples (2025):

• Google AlphaEarth Foundations (DeepMind, 2025) - 1.4 trillion embeddings/year in GEE
• NASA-IBM Geospatial Foundation Model (open-source, Aug 2024)
• Prithvi (IBM/NASA/ESA collaboration)
• Clay Foundation Model (open-source)
• Planet Labs + Anthropic Claude integration

Timing: 4 minutes

Key Points: - 2025 Update: Foundation models are THE major innovation in EO AI - Google AlphaEarth Foundations: Virtual satellite model, 10x10m resolution, integrates Sentinel-1/2 + Landsat + radar, available in Earth Engine, 16x less storage than other AI systems - NASA-IBM model released August 2024 as open-source - Trained on massive Sentinel-2 datasets (1 billion parameters) - Can be fine-tuned with just hundreds of labeled samples (vs thousands before) - Philippine Application: Use foundation models to jumpstart projects with limited labeled data - AlphaEarth embeddings already in GEE!

Example: “Instead of manually labeling 10,000 images for rice mapping, use AlphaEarth embeddings in GEE or fine-tune Prithvi with just 500 samples and achieve similar accuracy”

On-Board AI Processing

ESA Φsat-2 (Launched 2024)

• 22×10×33 cm CubeSat
• Onboard AI computer (Intel Myriad X VPU)
• Real-time cloud detection
• Process before downlink
• Saves bandwidth

Satellogic Edge Computing

• “AI First” satellites
• Onboard GPUs
• Real-time processing
• Immediate insights
• Ship/object detection

Timing: 3 minutes

Key Points: - 2025 Update: On-board AI is operational on multiple satellites - ESA’s Φsat-2 launched 2024 with Intel AI chip - Processes images on-orbit before transmitting - Use case: Only download cloud-free portions, save 90% bandwidth - Future: Real-time disaster detection from space

Philippine Relevance: “Imagine typhoon damage detected and reported automatically from orbit within minutes, not hours”

Self-Supervised Learning

Concept:

• Learn from unlabeled data
• Define pretext tasks (e.g., predict missing patches)
• Model learns useful representations
• Fine-tune with small labeled dataset

Why Important for EO:

• Abundance of unlabeled satellite imagery
• High cost of labeling
• Improves transferability

–>{width=“100%”}

Self-supervised learning is particularly relevant for EO due to abundance of unlabeled imagery. Models learn useful features without expensive labeling.

Explainable AI (XAI)

–>{fig-align=“center” width=“75%”}

Why XAI Matters:

• Understand model decisions
• Build trust in AI systems
• Debug and improve models
• Regulatory compliance

Methods:

• SHAP: Feature importance
• LIME: Local explanations
• Grad-CAM: Visual attention maps
• Saliency Maps: What pixels matter?

As AI systems make important decisions (disaster response, resource allocation), understanding why they make those decisions becomes crucial.

Summary & Key Takeaways

What We Covered

1. AI/ML Basics: What it is and why it’s powerful for EO
2. ML Workflow: 7-step process from problem to deployment
3. Supervised Learning: Classification and regression with labeled data
4. Unsupervised Learning: Clustering and pattern discovery
5. Deep Learning: Neural networks and CNNs for images
6. Data-Centric AI: Quality, quantity, diversity, labels
7. 2025 Trends: Foundation models, on-board AI, XAI

We’ve covered the fundamental concepts you need to understand before diving into hands-on implementation.

Key Takeaways

1. Focus on Data First

• Quality beats quantity
• Diversity enables generalization
• Good labels are gold

2. Start Simple

• Try traditional ML before deep learning
• Random Forest is often enough
• Add complexity only when needed

3. Iterate Continuously

• Analyze errors
• Improve data
• Retrain models
• Deployment is not the end

These principles will serve you well throughout your AI/ML journey. Data quality and iterative improvement are more important than fancy algorithms.

Practical Advice

For Your Projects:

• Define the problem clearly before collecting data
• Invest in high-quality training data - it’s worth it
• Validate rigorously on independent data
• Document everything (data sources, preprocessing, model versions)
• Start with baselines (simple models, existing methods)
• Iterate based on errors - let failures guide improvements
• Consider operational constraints early

These practical tips come from real-world experience. Following them will save you time and frustration.

The Data-Centric Mindset

When model performs poorly, ask:

1. Is my data clean?
2. Are labels accurate?
3. Is training data representative?
4. Do I have enough diversity?
5. Are there systematic biases?

Before trying:

• More complex model
• More epochs
• Different hyperparameters
• New architecture

Check your data first!

“Better data beats a cleverer model” - Andrew Ng

Adopt a data-centric mindset. When models underperform, investigate data issues before blaming the algorithm.

Connection to Sessions 3 & 4

Session 3: Python Basics

• Load and explore data
• GeoPandas (vector)
• Rasterio (raster)
• Foundation for all ML work

Session 4: Google Earth Engine

• Access Sentinel data at scale
• Cloud masking (data quality!)
• Temporal compositing
• Export for ML workflows

Everything builds on these concepts!

The hands-on sessions this afternoon put these concepts into practice. You’ll actually work with data and see these principles in action.

Looking Ahead: Days 2-4

Day 2:

• Random Forest classification
• Land cover mapping
• CNN basics
• TensorFlow/Keras intro

Days 3-4:

• U-Net for segmentation
• Flood mapping (DRR focus)
• Time series with LSTMs
• Foundation models
• Explainable AI

Over the next three days, we’ll implement these concepts in real EO applications for DRR, CCA, and NRM.

Resources for Continued Learning

Online Courses:

• NASA ARSET: ML for Earth Science
• EO College: Introduction to ML for EO
• Coursera: Machine Learning (Andrew Ng)
• Fast.ai: Practical Deep Learning

Papers & Tutorials:

• “Data-Centric ML for Earth Observation” (ArXiv 2025)
• Google Earth Engine tutorials
• TensorFlow Earth Observation tutorials

Communities:

• SkAI-Pinas network
• Digital Space Campus (CoPhil)
• DIMER model repository

These resources will support your continued learning after the training. The Digital Space Campus will have all our materials for reference.

Session Summary

What We Covered:

✅ AI/ML/DL definitions and relationships
✅ End-to-end ML workflow for EO
✅ Supervised learning (classification, regression)
✅ Unsupervised learning (clustering)
✅ Deep learning & CNNs for satellite imagery
✅ Data-centric AI philosophy
✅ 2025 innovations: Foundation models, on-board AI

Timing: 2 minutes

You now have conceptual foundation for all ML work in this course. Sessions 3-4 today and Days 2-4 will implement these concepts.

Q&A

AI/ML Concepts

• Supervised vs unsupervised?
• When to use deep learning?
• Foundation models for my use case?

Practical Questions

• Data quality challenges?
• Label collection strategies?
• Computing requirements?

Timing: 5-8 minutes for Q&A

Common Questions: - “Do I need a GPU?” → Not for Random Forest, yes for deep learning - “How many labels do I need?” → Depends: 100s with foundation models, 1000s for CNN from scratch - “Which algorithm should I use?” → Start simple (RF), then deep learning if needed - “Can I use foundation models for Philippines?” → Yes! They’re global and open-source

Next: Session 3

Hands-on Python for Geospatial Data

Coming up after 15-minute break:

• Google Colab environment setup
• GeoPandas for vector data
• Rasterio for raster data
• Work with Philippine boundaries
• Load and visualize Sentinel-2 imagery
• Calculate NDVI

Get ready to code! 💻

Thank You!

Resources

Foundation Models:
NASA-IBM Geospatial: https://huggingface.co/ibm-nasa-geospatial
Prithvi: https://github.com/NASA-IMPACT/Prithvi
Clay: https://clay-foundation.github.io

Learning:
NASA ARSET: https://appliedsciences.nasa.gov/arset
EO College: https://eo-college.org
SkAI-Pinas: https://asti.dost.gov.ph/skai-pinas

Session 2 Complete!

15-minute break before Session 3. Make sure participants have: - Google Colab access working - GEE account registration started (will finalize in Session 4)