Land cover and land use (LCLU) monitoring plays an important role for the monitoring of the landscape evolution due to anthropogenic or natural transformations and serves in many fields of applications such as urban planning, environmental monitoring, or even tourism. High accuracy land cover maps are usually produced by a team of trained photo interpreters with high resolution image data. However, the process of annotations is a time-consuming and demanding process and must be repeated in regular cycles to continue representing the reality of the field.Recently, deep learning and artificial intelligence have been put to the purpose of land cover and land use mapping with promising results. Machine learning algorithms can be trained on large datasets of satellite images to automatically identify and classify different land cover types with high accuracy. Despite these promising results, using AI for land cover mapping still faces several challenges.The prediction of the land cover categories constitutes itself an important difficulty: the delimitation of each class can complex, depending on fixed set of rules or additional information not directly visible in the images.
The photo-interpreters rely on a variety of data sources to produce maps such as textual information or ground-level images, that complement the aerial imagery. Recent AI models have demonstrated the ability to effectively incorporate multi-modal data, which can further improve the accuracy and reliability of the results.
Additionally, the interpretability and transparency of AI models can be an issue. Understanding how and why an AI model makes certain predictions is essential for building trust and ensuring the reliability of the result.
Therefore, this project aims at addressing these challenges and developing new AI models that are robust and effective in real-world scenarios for land cover mapping.
Partners and funding