## **Problem statement**

The objective of this task is to classify different landcover types in a satellite image. This problem is approached as a machine learning task known as semantic segmentation, where the goal is to predict the class label for each individual pixel in the image.

## **Dataset**
The [dataset](https://www.kaggle.com/datasets/balraj98/deepglobe-land-cover-classification-dataset) used for this project is from the 2018 DeepGlobe Landcover Classification Challenge. It consists of a total of 803 satellite images, each with dimensions of 2448x2448 pixels. Each image in the dataset is accompanied by a segmentation mask that assigns class labels to the pixels.

| Landcover Name            | Color                    | Explanation / Function                                         |
| -------------------- | ------------------------ | ------------------------------------------------------------ |
| Urban land           | <span style="color:cyan">Cyan</span>      | Man-made, built-up areas with human artifacts                |
| Agriculture land     | <span style="color:yellow">Yellow</span>  | Farms, planned plantations, cropland, orchards                |
| Rangeland            | <span style="color:magenta">Magenta</span> | Non-forest, non-farm, green land, grass                      |
| Forest land          | <span style="color:green">Green</span>     | Land with at least 20% tree crown density and clear cuts     |
| Water                | <span style="color:blue">Blue</span>       | Rivers, oceans, lakes, wetlands, ponds                       |
| Barren land          | <span style="color:gray">White</span>      | Mountains, rocks, deserts, beaches, vegetation-free land     |
| Unknown              | <span style="color:black">Black</span>     | Clouds and others                                             |

## **Model**
For this task, we utilized a pre-trained UNet model with weights pretrained on the ImageNet dataset. We then fine-tuned the UNet using the DeepGlobe Landcover Classification dataset. The training process took approximately 2 hours using a single NVIDIA T4 GPU.
## **Team members**
David Mora

Eduard's Mendez


## **Aditional information**

If you are interested in contributing to the project or just getting more information about the details you can head over to our GitHub [repository](https://github.com/DavidFM43/landcover-segmentation).