There is a large body of legacy scientific code written in languages like Fortran that is not optimised to get the best performance out of heterogeneous acceleration devices like GPUs and FPGAs, and manually porting such code into parallel languages frameworks like OpenCL requires considerable effort. We are working towards developing a turn-key, self-optimising compiler for accelerating scientific applications, that can automatically transform legacy code into a solution for heterogeneous targets. In this paper we focus on FPGAs as the acceleration devices, and carry out our discussion in the context of the OpenCL programming framework. We show a route to automatic creation of kernels which are optimised for execution in a "streaming" fashion, which gives optimal performance on FPGAs. We use a 2D shallow-water model as an illustration; specifically we show how the use of \emph{channels} to communicate directly between peer kernels and the use of on-chip memory to create stencil buffers can lead to significant performance improvements. Our results show better FPGA performance against a baseline CPU implementation, and better energy-efficiency against both CPU and GPU implementations.