Crop mapping is an important research area in agriculture monitoring, since an accurate map of agricultural land provides a comprehensive understanding of crop distribution, yield estimation, and the factors influencing crop production.

Over the past years, the launch of Earth Observation satellites with diverse sensors has produced abundant satellite image time series containing valuable agricultural information. However, effectively leveraging the data for accurate crop mapping still faces the following problems: a) The optical satellite image time series suffer from data gaps caused by haze or cloud shadow, severely affecting the performance of subsequent classification or segmentation task; b) Current crop mapping methods attempt to make more comprehensive use of information from different sensors by fusing them trivially, but they ignore the data corruption existed in part of modalities. c) Current deep neural networks for semantic segmentation often exhibit over-confidence in their predictions due to overfitting, resulting in unreliable decision making, such as improper agricultural subsidy allocation.

To address these problems, the corresponding solutions are proposed: a) Explore a diffusion model for cloud removal in optical satellite time series. The model leverages both spatial and temporal information for cleaner, high-quality images and improved crop mapping performance; b) Develop an adaptive attention-based fusion method with a quality-aware module, which dynamically highlights the most relevant high-quality features from different modalities to obtain more accurate and robust predictions; c) Design a probability calibration method to adaptively calibrate the prediction of each pixel, which is a crucial step towards achieving trustworthy predictions in practical applications.

Through these solutions, the research aims to enhance the utilization of satellite image time series and develop more effective and robust algorithms to improve the performance of crop mapping.