In this study, we leverage published machine learning models to predict soil moisture in agricultural areas from satellite data. We test how well these models perform when trained on one crop or season and then applied to another. This approach mimics real-world shifts that often happen in farming. We evaluate the diversity of training data needed for reliable results by comparing model performance across diverse agricultural conditions using the International Soil Moisture Network (ISMN). These findings can inform future field campaigns for developing agriculture-focused soil moisture datasets in new regions.