Model‐resolution impacts on all‐sky infrared radiance simulation and cloud‐dependent observation errors over Indonesia

Indonesia experiences rapidly evolving convection modulated by strong land–sea interactions across the Maritime Continent, requiring frequent observations for accurate numerical weather prediction (NWP). Sparse in situ observations, particularly over oceans, make geostationary infrared satellite observations an important source for assimilation in Indonesia's NWP system. However, operational use remains restricted to clear‐sky conditions, because all‐sky infrared assimilation is challenging: observation‐minus‐background (OB) departures are often large and strongly non‐Gaussian due to measurement, model, and representativity errors. Since model and representativity errors vary with model resolution, understanding cloud‐dependent observation‐error characteristics across scales is essential. This study investigates how model resolution affects (nearly) all‐sky infrared radiance statistics and cloud‐dependent observation‐error modelling over Indonesia. Cloud‐dependent OB standard‐deviation diagnostics are applied to the 6.9‐m water‐vapour channel using cloud‐impact formulations based on the 10.4‐m window channel. These diagnostics are used to propose cloud‐dependent observation‐error models using directly binned and piecewise‐linear forms. These are evaluated using Himawari‐9 observations and simulated brightness temperatures generated through fast radiative transfer simulations applied to 9‐ and 3‐km WRF forecasts within the Badan Meteorologi, Klimatologi, dan Geofisika (BMKG) Indonesian‐NWP (INA–NWP) system. Compared with the 9‐km configuration, the 3‐km model reduces OB bias, spread, and non‐Gaussianity, particularly under cold and high‐cloud conditions, and improves cloud detection across all cloud phases. However, it also produces larger false‐alarm rates in clear and liquid‐cloud regimes, indicating increased cloud mislocation and greater small‐scale variability. Proposed cloud‐dependent observation errors are generally smaller at 3‐km resolution, consistent with reduced model and representativity errors. Normalising OB statistics using cloud‐dependent observation‐error models improves Gaussianity and supports all‐sky assimilation. Directly binned models reduce OB departures but retain heavy tails at high cloud impact, while piecewise‐linear models suppress these non‐Gaussian tails further, particularly at 9‐km resolution. These findings highlight the importance of scale‐aware observation‐error modelling for future all‐sky infrared radiance assimilation.