Crop Classification with SAR Data and Deep Learning Algorithms: A Method for Areas with Limited Labels and High Cloud Cover

The rural insurance market faces significant challenges in matters such as risk analysis, insurance pricing and claims verification, among others. In this context, obtaining additional information to support the decision-making process is a crucial step for insurers in addressing these issues. Deep learning algorithms, in conjunction with Earth Observation (EO) data, have proven to be highly effective in various tasks within this sector. However, the vast scale of EO data makes the creation of large, pixel-level expert-annotated datasets costly and time-intensive. Additionally, physical limitations, such as high cloud density over extended periods of the year, reduce the usability of optical satellites - a potentially rich source of visual information. Motivated by these challenges, this project aims to develop a method capable of performing crop classification at the field plot level for a municipality with limited labels, independently of cloud cover. To achieve this objective, segmentation models will be used to extract field plot geometries, along with clustering models to address the scarcity of labeled data. In addition, the use of Synthetic Aperture Radar (SAR) images, which are unaffected by cloud presence, is also introduced. It is expected that the proposed method will improve accuracy and efficiency in gathering information for the rural insurance market, leading to significant advancements in risk analysis and management in the sector.