Federated learning-based semantic segmentation framework for sustainable development
Abstract
More than a third (38% to be exact) of all inhabited land is covered by forests, which serve many purposes including nutrient cycling, water, climate management, water purification, primary production, fuel wood, etc. They play a vital role in sequestering carbon and providing a home for a wide range of plant and animal life. Agriculture relies on the services provided by forests. Changes in land cover can be easily detected by using satellite imagery, which provides a wealth of useful information. Sustainable development and human well-being rely on effective forest utilization and management, which is the subject of this effort. Federated Learning protects user privacy by processing data locally on client devices rather than storing it centrally on a server. Instead of sending the same model to all clients at once, as is done in traditional training paradigms, we suggest a new paradigm called FedStv, in which the model trained on the active client in each round is used to train the next active client, as chosen by the server, in the following round. All of the clients use the derived server average once more for subsequent training. Finally, the uncertainty map estimate standard deviation for the projected segmentations has been calculated. The experimental results demonstrate that the suggested model can produce higher Dice Scores and Intersection over Union (IoU) values when applied to the dataset of Forest aerial pictures for segmentation.