Basically, what we have here is a way for computers to look at satellite images and figure out what they’re showing us like if it’s a forest or a parking lot or something in between.

Now, before we dive into the details of how this works, let me explain why vision transformers are so great for this task. Unlike traditional computer vision methods that rely on convolutional neural networks (CNNs), which process images by breaking them down into smaller and smaller pieces, vision transformers use a different approach called self-attention.

Self-attention allows the model to focus on specific parts of an image without having to explicitly identify those parts first kind of like how our brains can instantly recognize faces in a crowd without having to scan every single person’s face individually. This makes vision transformers much more efficient and accurate than CNNs, especially when dealing with large-scale datasets or complex scenes.

So, let me walk you through the process step by step:

1. First, we load our satellite images into memory (or “memory” in this case is actually a fancy computer chip called a GPU). These images are usually pretty big like 256×256 pixels or more so it’s important to make sure they fit within the available memory space.

2. Next, we preprocess the images by converting them into a format that our model can understand (i.e., turning each pixel value into a number between 0 and 1). This is called normalization or standardization, depending on who you ask.

3. Then, we feed these preprocessed images through our vision transformer model, which consists of several layers of self-attention blocks (also known as “transformer encoders”). Each block takes in an input sequence and outputs a new sequence that has been transformed by applying the attention mechanism to it.

4. The output sequences from each block are then concatenated together to form a final feature vector, which is used for classification purposes. This process is called “pooling” or “max pooling”, depending on who you ask.

5. Finally, we pass this feature vector through a fully connected layer (also known as a “dense layer”) that converts it into a set of probabilities one probability for each possible class label in our dataset. These probabilities are then used to make predictions about which class the input image belongs to.

That’s how vision transformers work in a nutshell. Of course, there are many more details and nuances that go into making this process work well for satellite images specifically but hopefully this gives you a good idea of what’s going on under the hood.