An End‐to‐End U‐Net Based Framework for Monitoring Páramo Ecosystems in Colombia

Abstract The páramo, referred to as a variety of tundra ecosystems, are considered one of the most essential biomes in South America, as they are primary sources of fresh water for many countries. However, páramos are increasingly threatened by agricultural expansion, livestock grazing, mining activities that cause deforestation and pollution, urbanization and infrastructure development that fragment and destroy habitats, and climate change, disrupting the delicate balance of these ecosystems. While many approaches have been proposed to monitor forest, grassland, mountain, and other ecosystems, páramo monitoring remains overlooked in the literature. In this work, we propose an end‐to‐end framework for monitoring páramo ecosystems in Colombia by leveraging deep neural networks and cloud‐based technologies. We develop a U‐Net convolutional neural network architecture trained on data set specifically tailored to páramo ecosystems, created through manual annotation of high‐resolution satellite imagery from Google Earth Pro. The model captures spatial features of páramo vegetation to enable pixel‐level semantic segmentation across different time periods. We demonstrate that the model achieves 90% accuracy in páramo versus non‐páramo classification on a manually annotated test data set. Additionally, we deploy the model in a cloud‐based service, enabling scalable and reproducible páramo monitoring that can be replicated on any platform. This comprehensive framework—encompassing data set creation, model development, and production deployment—provides a generalizable approach for developing and deploying ecosystem surveillance models in operational monitoring systems.