To automatically classify biological images, machine learning techniques have been widely used to train the classifiers from labeled images. For a new category of biological object, a tedious and expensive labeling process is needed from a human expert. With the growing amount of biological data and the increasing number of categories to recognize, a more efficient method to train the classification system is required. The aim of this dissertation research is to effectively reduce the labeling effort of human experts in training the image classification methods. The contributions of this research consist of the following key components: First, the size differential regularization is employed to refine the ranking of classification rules to alleviate the risk of over-fitting in the case of a small number of training samples. Second, the spatiotemporal connectivity among the unlabeled samples is utilized to determine the weighting scheme of the existing classifiers from multiple sources. Third, the target directed sampling is proposed to focus the search for additional samples which are most likely to belong to the new class. The approaches are demonstrated to be effective in biological experiments including cell detection, insect detection, and pollen classification. The experimental results indicate that the proposed methods can achieve comparable performance to the current machine learning approaches while significantly reduce the amount of training data.