Deep learning has become indispensable in designing solutions for several real-world scientific problems. We review some state-of-the-art deep learning approaches that form the basis of these solution frameworks. These include strategies that incorporate the underlying physical constraints into the design process, as well as situations that can benefit from a judicious fusion of multimodal data. Finally, we provide three case studies of deep learning applications in agriculture to highlight the potential of such strategies in agricultural sciences and other allied fields. In particular, we discuss models for predicting crop yield, applications in plant phenotyping, and multi-omics data integration for agriculture.