Knowing what crops are grown in farmers’ fields, where croplands are located and what extent croplands cover is of great importance for better monitoring of global croplands, especially given the urgent need to double global food production by 2050 to feed a roughly 9.7 billion population. Compared to the survey- or census-based approaches, Earth observation (or satellite) provides consistent and rapidly monitoring of global croplands. With the increasingly accessible satellite data and advances in algorithms, the bottleneck in identifying crop extents and types has shifted to the lack of ground truth. Ground truth (or labels) refers to information collected on the location that is used to train and validate satellite-based models. Ground truth requires great investments in time and labor to collect and is with low reusability once collected. As such, in most cases, satellite-based crop extent and type identification can only rely on either outdated historical ground truth, limited within-year ground truth, or even no ground truth at all. The scarcity of ground truth prominently hinders the implementation of interventions to promote sustainable agriculture. For example, an accurate crop type map in the early season is beneficial to numerous pre-harvest applications but is barely accessible because it is difficult to collect ground truth during the early time. The goal of my dissertation is to develop new theories and approaches to address the constraint of ground truth and illustrate their utility in different crop extent and type identification applications that promote sustainable agriculture from different angles.