Video 1: Machine Learning Foundations for Earth Observation (Session 2)

Description:
This video introduces the fundamentals of machine learning in the context of Earth Observation. You’ll explore supervised and unsupervised learning methods, understand how to prepare and evaluate models, and see how algorithms like logistic regression, SVMs, k-nearest neighbors, and k-means clustering are applied to remote sensing tasks such as damage assessment, object detection, and quantitative analysis.

Learning Goals:

1. Explain the basic machine learning pipeline and its relevance to remote sensing workflows.

2. Differentiate between supervised and unsupervised learning and select suitable approaches for EO tasks.

3. Apply logistic regression, SVMs, and k-NN for classification of geospatial data.

4. Use k-means clustering to discover hidden patterns in unlabeled Earth Observation data.

5. Understand cost functions, gradient descent, and evaluation metrics to optimize model performance.

We have made the following files available for download: Session 2_Machine Learning Foundation.ipynb (769.3 kB, sample code), Session 2_Machine Learning Foundation.pdf (2.2 MB, session slides), and UCM_toy.zip (sample dataset).

Supervised learning is like teaching a computer by example. It’s where the magic of machine learning truly begins. In Earth Observation, this means showing an algorithm satellite images (inputs) along with the correct labels, like “forest,” “urban,” or “water”, so it can learn the patterns that distinguish them. Once trained, the model can classify entirely new, unseen data with impressive accuracy.

From damage assessment after natural disasters to mapping agricultural crops or monitoring urban growth, supervised learning empowers us to extract meaningful insights from vast amounts of remote sensing data. By leveraging techniques such as logistic regression, support vector machines, and k-nearest neighbors, you’ll see how algorithms learn to recognize complex features, build decision boundaries, and turn pixels into actionable geospatial intelligence. It’s the cornerstone of many AI4EO breakthroughs and a powerful skill you’ll master in this course.

Unsupervised learning is like exploring a new world without a map. It discovers hidden patterns and structures in data without needing labeled examples. In Earth Observation, this means feeding a model vast amounts of satellite or hyperspectral imagery and letting it uncover natural groupings, such as distinguishing forests, cities, and water bodies, without any prior guidance.

This approach is especially powerful when labeled data is scarce or expensive to obtain, which is common in remote sensing. Techniques like k-means clustering can reveal meaningful land cover classes, detect unusual changes, or highlight emerging features in time-series imagery. By mastering unsupervised learning, you’ll gain the ability to reveal insights that even experts might overlook, making it a critical skill for tackling complex, real-world AI4EO challenges.

Video 2: Deep Learning Foundations for Earth Observation (Session 3)

Description:
This video explores deep learning as a powerful subset of machine learning for Earth Observation. You’ll learn about feed-forward networks, autoencoders, convolutional and recurrent neural networks, GANs, and their applications to hyperspectral classification, damage mapping, SAR–optical translation, and more. Practical training tips—like normalization, dropout, data augmentation, and transfer learning—are also covered.

Learning Goals:

1. Describe the key deep learning architectures (feed-forward, CNN, RNN/LSTM, autoencoders, GANs) and their roles in EO.

2. Explain how CNNs, RNNs, and GANs are applied to hyperspectral, SAR, and multispectral imagery.

3. Recognize challenges like gradient vanishing and overfitting and apply strategies such as residual connections, early stopping, and dropout.

4. Use data augmentation, transfer learning, and optimizers to improve deep learning model performance.

5. Relate deep learning concepts to practical EO tasks such as scene classification, semantic segmentation, object detection, and change detection.

Machine Learning (ML) and Deep Learning (DL) are closely related, but they differ in how they learn and the complexity of the tasks they can handle:

• Machine Learning uses algorithms (e.g., logistic regression, SVMs, k-nearest neighbors) that often rely on handcrafted features or simpler models. In Earth Observation, ML might involve extracting vegetation indices or texture features from satellite images and then training a classifier. ML works well with smaller datasets and is easier to interpret, but its performance depends heavily on feature engineering and expert knowledge.

• Deep Learning is a subset of ML that uses multi-layered neural networks to learn features automatically from raw data. Instead of manually designing feature extractors, DL models—like CNNs, RNNs, and GANs—directly discover patterns in imagery, enabling breakthroughs in tasks like scene classification, semantic segmentation, and change detection. DL excels with large datasets and computational power (GPUs), achieving state-of-the-art performance in complex EO applications.

In short: ML = simpler models + handcrafted features, while DL = neural networks that learn features automatically, unlocking higher accuracy and advanced capabilities for AI4EO.

A Feed-Forward Neural Network (FNN) is the simplest yet foundational type of deep learning model—think of it as information flowing in one direction, from input to output, without looping back. Data enters through the input layer, is transformed by hidden layers of neurons that learn weighted connections and biases, and finally produces a prediction at the output layer.

In Earth Observation, FNNs can classify satellite pixels, estimate environmental parameters, or map land cover types. Although they’re less powerful for complex patterns compared to CNNs or RNNs, feed-forward networks are easy to understand, fast to train on smaller datasets, and provide an essential starting point for anyone learning deep learning.

Convolutional Neural Networks (CNNs) are the workhorses of modern computer vision, designed to recognize patterns directly from raw images without the need for handcrafted features. Instead of treating every pixel independently, CNNs slide small filters (kernels) across an image to detect edges, textures, shapes, and higher-level features. Pooling layers then reduce complexity while preserving essential details, and fully connected layers translate these features into meaningful predictions.

In Earth Observation, CNNs excel at tasks like scene classification, semantic segmentation, object detection, and change detection. They can distinguish urban areas from forests, map flood damage, or identify ships in SAR imagery with remarkable accuracy. Architectures like LeNet, AlexNet, VGG, GoogLeNet, and ResNet have set milestones, each improving how networks learn deeper, richer representations. CNNs are not just powerful—they’re transformative, enabling AI4EO to analyze massive satellite datasets and extract actionable insights about our planet with unprecedented precision.

Recurrent Neural Networks (RNNs) are like storytellers for your data—they don’t just look at individual frames or snapshots but remember what came before, weaving past information into future predictions. In Earth Observation, this means they can track the narrative hidden in satellite image sequences: the slow greening of crops, shifting coastlines, or the unfolding impact of a natural disaster.

Yet, classic RNNs sometimes “forget the plot,” struggling with very long stories because of vanishing or exploding gradients. That’s where their smarter relatives step in: Long Short-Term Memory (LSTM) networks use gates and memory cells to preserve important details for hundreds of steps, while Gated Recurrent Units (GRUs) offer a leaner, faster alternative without sacrificing performance. Bidirectional RNNs even read the story both forwards and backwards, giving them richer context for complex sequences.

These innovations make RNNs and their modifications indispensable for time-series remote sensing—unlocking powerful insights for crop monitoring, land-cover change detection, damage mapping, and beyond. They don’t just process data; they reveal the evolving stories our planet tells over time.

Generative Adversarial Networks (GANs) are one of the most exciting breakthroughs in deep learning. A GAN consists of two competing neural networks—a generator and a discriminator—locked in a creative “game.” The generator tries to create data (like satellite images) that look real, while the discriminator tries to tell whether an image is real or fake. Over time, both networks improve, and the generator learns to produce highly realistic outputs.

In Earth Observation, GANs are revolutionizing tasks such as data augmentation, image translation (e.g., converting SAR images to optical imagery), hyperspectral data synthesis, and super-resolution for sharpening low-quality satellite images. By learning how to mimic complex geospatial patterns, GANs help overcome data scarcity, improve model performance, and open new possibilities for simulating and analyzing the Earth’s surface.

Effective deep learning training for Earth Observation relies on several best practices: normalize input data to stabilize learning, use learning rate schedules to refine convergence, and choose optimizers like Adam or RMSProp for efficient updates. Monitor both training and validation loss to detect overfitting, and apply regularization methods such as dropout alongside data augmentation techniques like flips, rotations, and color jittering to improve generalization. Leverage transfer learning with pre-trained models to save time and resources, and employ early stopping to prevent unnecessary overfitting. Finally, make use of GPUs and mini-batch training for faster, more stable computations, ensuring your models perform robustly on complex tasks like scene classification, change detection, and semantic segmentation.