5  Machine Learning Tasks for Satellite Data

Before we dive into a few machine learning algorithms, I want to focus on the higher level. Because whenever we want to use machine learning, we have to translate our question or problem into a machine learning task. For example, if we want to know how many new buildings there are in an area, we can translate this into an object detection task and solve it with machine learning. To know what kind of questions you can answer with machine learning, you have to know which tasks there are.

The following is a non-exhaustive list of machine learning tasks related to remote sensing:

Task	Example	Input	Output	Type
Image-wise classification	Land cover classification	Tensor	Probability vector	Supervised
Segmantic segmentation (=pixel-wise classification)	Land cover mapping	Tensor	Binary or multi-class image mask	Supervised
Instance Segmentation	Counting trees or buildings	Tensor	Masks + labels per mask	Supervised
Image-wise regression	Population density estimation	Tensor	Scalar or continuous value	Supervised
Pixel-wise regression	Soil moisture estimation	Tensor	Tensor	Supervised
Object detection	Detecting ships	Tensor	Bounding boxes + class labels	Supervised
Change Detection	Detecting deforestation	Multi-temporal tensors	Binary or multi-class change mask	Supervised / Unsupervised
Inpainting	Filling data gaps	Masked tensor	Completed tensor	Supervised / Self-supervised
Anomaly detection	Detecting constructions in protected areas	Tensor	Anomaly score or mask	Unsupervised / Semi-supervised
Denoising	Cloud removal	Tensor (noisy one)	Tensor	Supervised / Self-supervised
Super-resolution	Enhancing satellite images.	Tensor	Higher dimensional tensor	Self-supervised
Autoencoding	Pretraining for other tasks	Tensor	Embedding or reconstruction	Self-supervised

Be creative when translating your project into a machine learning task and sometimes it can be useful to try something unusual.