Learning representations that capture meaningful underlying information of data is a promising solution to reduce the reliance on labeled data for downstream applications. With the advent of the big remote sensing data era, self-supervised deep learning methods have become a valuable tool to extract high-level, complex abstractions and representations from large volumes of data. However, classical methodologies such as Variational Autoencoders are focused on imposing statistical constraints on the latent space and they do not learn generic and interpretable representations of the data. To address such limitation, this work presents a generic physics-guided representation learning methodology to discover semantic representations. To address it,the proposed approach constrains the learning process with the incorporation of prior physical knowledge. This study shows through an example how the methodology can be used to solve remote sensing inverse problems. Specifically, the inversion of a crop phenology model derived from NDVI time series is proposed. As a result, the probability distributions of the intrinsic physical model parameters are inferred. The feasibility of the method is evaluated on both simulated and real Sentinel-2 data and compared with different standard algorithms.