Forecast sensitivity to observations in the Met Office Global numerical weather prediction system

An adjoint‐based method for calculating the impacts of observations in the Met Office's global four‐dimensional variational assimilation (4D‐Var) system is documented. Our approach is novel, as we seek from the outset a linearized approximation to partition the finite impact from a batch of observations, but our method and results are very similar to those of other systems. The large beneficial impacts measured for satellite radiances and radiosonde soundings in our system are interesting results. We also identify areas for potential improvement, such as the assimilation of Northern Hemisphere stratospheric satellite radio occultation (GPSRO) observations, polar atmospheric motion vectors and tropical cyclone (TC) bogus observations synthesized automatically from tropical cyclone warning centre advisory messages. We use a toy model to explain why only just over 50% of observations improve the forecast. Growing modes, observational errors and errors in the verifying forecast contribute approximately equally and we suggest that our necessarily imperfect knowledge of background‐error statistics also contributes.