Climate change is one of humanity’s biggest challenges. This year we have seen new record-breaking heatwaves across the globe, unprecedented floods, water has become a scarce resource in many areas, and an alarming number of species have become extinct. To stop the advance of climate change, we must preserve our existing wetlands. Wetlands are a vital part of our ecosystem. They sequestrate carbon dioxide, act as natural barriers to prevent floods, purify the water humans consume, and are home to many flora and fauna1. Wetlands are composed of both open water and vegetated water areas, and their area size changes constantly depending on the season of the year or weather conditions. Sadly, 70% of the wetlands worldwide have disappeared since the 20th century and continue to do so, despite conservation efforts such as the Ramsar convention for wetland protection. If we want to preserve our wetlands, first, we must know (i) Where are they located? (ii) What is the extent and dynamics of surface water coverage? (iii) How has surface water extent changed over time? This is key to identifying which of our wetlands are rapidly degrading and taking countermeasures before they disappear. Calculating the lateral water extension area of a wetland is a challenging problem. One difficulty is that the water hidden under the vegetation cannot be detected using optical sensors, and fieldwork is costly and often logistically challenging. Moreover, there is a lack of field data and obtaining additional data is expensive. The DeepWetlands project has three primary research questions: 1. How can we quantify the water extension of Swedish wetlands at a specific point in time? 2. How can we accurately detect the quantity of water hidden under the vegetation? 3. How has the Swedish wetlands' water extension changed over the period 2015-2025? In this project, we will be working on developing novel machine learning algorithms to help us measure the water extension of wetlands. We will have a particular focus on self-supervised models to overcome one of the biggest bottlenecks in deep learning: the need for manually annotated data. In our project we will work on algorithms that automatically generate training data. Preliminary results in our work show that we are able to generate such data with positive results. However, Convolutional Neural Networks (CNNs) have shown limited applicability since they have to be re-trained for each region. In this new project we want to use Visition Transformers (ViT) to create models that can generalize well across regions and detect water regardless of the topography of the terrain. To achieve our goals, we need GPU time to train these large neural network models. I am the principal investigator for this project and my project is funded by Digital Futures. I am working jointly at Stockholm University and at KTH. In Stockholm University I collaborate directly with associate professor Fernando Jaramillo and at KTH I collaborate with associate professors Amir Payberah and associate professor Zahra Kalantari.