Segmentacija primerkov v daljinskem zaznavanju

Instance segmentation in remote sensing imagery plays a crucial role in appli- cations such as urban planning, agriculture monitoring, and environmental management by enabling precise object-level analysis. Across all the chal- lenges in developing robust segmentation models for satellite imagery, the lack of well defined parcel borders is especially important, as it makes it dif- ficult to distinguish between individual fields or land plots accurately. Fur- thermore, this issue gets compounded by factors such as spectral diversity, limited labeled data, and variable imaging conditions, which further compli- cate model training and generalization. This thesis investigates instance seg- mentation on satellite imagery using the PASTIS benchmark dataset, which provides multi-temporal observations of agricultural fields. We borrow the idea from the field of microscopy analysis to predict pixel-wise gradient di- rections to guide Cellpose for instance mask reconstruction, complemented by parallel semantic segmentation to assign class labels at the instance level. Existing architectures such as U-Net and Set Transformer are adapted to effectively incorporate temporal sequences and the rich spectral information of the dataset.