Advanced AI for Earth Observation

Scene Classification: The Art of Labeling the Big Picture

If semantic segmentation is micromanagement and object detection is middle management, then scene classification is the laid-back CEO. It doesn’t bother with details - it just glances at the whole image and says: “Ah yes, that’s an airport.” No bounding boxes, no pixel-by-pixel drama - just a confident, big-picture decision.

In this section, you’ll learn how machines decide whether an image is a meadow, a dense residential area, or the occasional suspicious swimming pool in the desert. We’ll explore:

• How pre-trained CNNs (like AlexNet and ResNet) can be re-purposed as feature extractors (because why reinvent the wheel when ImageNet already trained for years?),

• How ensemble models (aka a committee of CNNs arguing until they vote) boost accuracy,

• How multi-scale CNNs handle the fact that airplanes look different whether they’re parked or in the air,

• And how GANs can generate synthetic training images when you don’t have enough labeled data (cue the AI imagination).

By the end, you’ll master the art of teaching AI to look at a remote sensing image and say something more intelligent than: “This looks… green?

Object Detection: When Bounding Boxes Become Life

Scene classification is like saying: “This image has airplanes.” Semantic segmentation micromanages every pixel. But object detection? That’s the practical middle manager: “Here’s the airplane… inside this neat rectangle. And here’s another one… also in a rectangle. Everything important in boxes, just the way we like it.”

In this section, we’ll unpack the toolbox of modern detectors:

• From the classics (R-CNN family - basically slow but careful detectives),

• To YOLO (“You Only Look Once,” aka the speedster that finds airplanes faster than you can blink),

• To oriented detectors (because in aerial images, airplanes don’t politely sit horizontal - they tilt, twist, and overlap, so we need rotated boxes to catch them).

We’ll discuss how anchor boxes, feature pyramids, and non-maximum suppression keep everything in order - so you don’t end up detecting the same cow 47 times.

By the end, you’ll be able to spot, localize, and label objects in remote sensing imagery like a pro - turning “There might be an airplane here” into “Yep, there’s definitely an airplane, 57.3 meters long, right at these coordinates.”

Semantic Segmentation: Giving Every Pixel a Job

Imagine you’re the boss of a huge company with millions of employees (pixels). Scene classification just says: “This company works in aviation.” Object detection says: “Here’s where the airplanes are.” But semantic segmentation? That’s the micromanaging boss who points at each pixel and says: “You - runway! You - airplane! You - meadow! And you - suspiciously green building pretending to be a meadow!”

In this section, we’ll dive into the algorithms that make this pixel-level democracy possible:

• From old-school patch-based CNNs (aka the slow interns cutting images into tiny squares)

• To fully convolutional networks (FCNs), U-Net (the encoder–decoder power couple with skip connections),

• To DeepLab (because sometimes you need dilated convolutions and CRFs to keep those pixels in line).

You’ll also see how weakly- and semi-supervised approaches come in when labeling every single pixel by hand feels like a cruel punishment (spoiler: it does).

By the end, you won’t just see the whole picture — you’ll understand it pixel by pixel.

⏳ Change Detection: Spot the Difference, But for Satellites

Remember those “spot the difference” puzzles in magazines? Change detection is basically that - except instead of cartoon ducks, we’re comparing multi-temporal satellite images. And the differences aren’t “hat vs. no hat,” but “forest vs. shopping mall” or “coastline vs. coastline-that’s-now-underwater.”

In this section, we’ll see how AI figures out:

• Pixel-based methods (the brute force approach: subtract image A from image B and hope for the best),

• Object-based methods (because sometimes it’s smarter to track buildings than every single pixel of sidewalk),

• Deep learning approaches like Siamese networks (twins that look at two images and gossip about what’s changed),

• And end-to-end CD networks that fuse, compare, and classify changes with minimal human babysitting.

We’ll also cover real-world applications: wildfire damage, urban growth, deforestation, and post-disaster building assessment (aka “is this house still standing?”).

By the end, you’ll be able to teach your model to not just look at images, but remember the past, compare it to the present, and point out what really changed — without confusing seasonal snow for a new shopping mall.

?️ Anomaly Detection: Finding the Weird Stuff from Space

If scene classification tells us what’s normal and change detection tells us what’s different over time, then anomaly detection is the suspicious neighbor who always notices when something just doesn’t look right. It doesn’t need labels, it doesn’t need to know if it’s a building, a car, or a cow - it just screams: “Hey! That pixel doesn’t belong here!”

In this section, we’ll dive into how machines identify the rare, unexpected, or downright weird:

• Statistical methods (the “math police” that flag anything too far from the mean),

• Spectral–spatial techniques (because sometimes a “tree” with metal reflectance is probably not a tree),

• Deep autoencoders (compress–rebuild–compare: if the model can’t reconstruct it, it’s probably suspicious),

• And GAN-based approaches that learn what “normal” looks like - so they can side-eye anything unusual.

From illegal mining and oil spills to ships where there shouldn’t be ships, anomaly detection turns satellites into cosmic detectives hunting outliers.

By the end, you’ll have the knowledge to train your AI to ask the all-important question: “Is this supposed to be here?”

So, you’ve learned how to build the perfect AI model. It can spot an airplane from 30,000 feet, segment a city block in seconds, and tell a forest from a freeway. You've prepped the data, trained the model, and deployed it. Time to relax, right?

Not so fast.

What if I told you that your genius model could be tricked by an "adversarial T-shirt" into thinking a person isn't there? Or that adding a tiny bit of "invisible" static - a perturbation - could make it believe a panda is a gibbon with 99.3% confidence?

This isn't just a party trick. In remote sensing, this means an attacker can add a nearly invisible layer of noise and make your model classify an "Airplane" as a "Runway," or turn a 93% accurate land cover map into 19% garbage.

Welcome to the world of AI Security, where our models are ridiculously powerful, incredibly smart, and... catastrophically gullible. In this section, we’re pulling back the curtain on the dark arts of AI hacking.

Part 1: The "Stranger Danger" (Adversarial Attacks)

This is the most common threat: an attack that happens after your model is trained and deployed. It's all about feeding the model a "poisoned sample" that looks perfectly fine to you and me, but makes the AI go haywire.

We'll look at the hacker's toolkit:

• FGSM (Fast Gradient Sign Method): The classic, "one-shot" attack. It’s the brute-force way to find the fastest path to confusing the model.

• I-FGSM (Iterative FGSM): The sneakier version. Why push the model once when you can nudge it ten times until it breaks?

• Black-Box vs. White-Box: What if the attacker doesn't have your model's blueprints? No problem! We'll see how "transferability" means an attack designed to fool one model has a scary-good chance of fooling your model, too.

  
  

Part 2: The "Inside Job" (Backdoor Attacks)

This one is scarier. This attack happens during training. The attacker doesn't trick your model; they build the trick into it.

It’s a sleeper agent. The model works perfectly... until it sees the secret "trigger" - a tiny, harmless-looking patch of pixels. When it sees that trigger, it activates the hidden backdoor, like forcing every image with that patch to be classified as "Harbor," no questions asked.

Part 3: The Boot Camp (How We Fight Back)

So, are we doomed? Not at all. It just means we need to send our models to a security boot camp. We'll explore the AI security arsenal:

• Adversarial Training: The "vaccination" approach. We show the model thousands of these tricky examples during training until it learns to ignore them.

• Adversarial Purification: The "bouncer" at the door. This technique "denoises" or "purifies" the input, scrubbing away the perturbation before the model ever sees it.

• Model Reconstruction: The "exorcism" for backdoors. We'll learn how to "prune" or retrain a suspicious model to remove the hidden threat.

  
  

By the end, you'll be able to see remote sensing images not just for what they are, but for what a hacker wants your AI to think they are.