There are two common approaches for severity estimation, first disease region localization and second severity prediction with the help of a pre-learning model. The disease localization is a major challenge in the case of disease containing complex patterns and also annotation is difficult. Considering both challenges, a semi-supervised method is explored and proposed where feature learning is on healthy leaves images and based on the learning, severity is estimated for disease leaves. The proposed method has advantages; healthy images are easily available and annotation is not required. Using the proposed method, a linear relation is observed between expected (ground-truth annotated) severity and learning-based severity for different diseases such as wheat tan, maize rust etc., where disease contains multiple irregular shapes, patches, and dots of different sizes.