Nowadays, near infrared (NIR) and mid-infrared (mid-IR) reflectance spectroscopy are recognised useful approaches for quantifying soil properties, cost and time effectively. The aim of this work was to compare predictions of soil carbon (C) and nitrogen (N) content, C/N ratio, substrate-induced respiration (SIR) and denitrifying enzyme activity (DEA) using NIR and mid-IR spectroscopy over a diverse set of 360 Malagasy topsoils. Partial least squares regression was used for fitting NIR and mid-IR spectra to conventional data through procedures of calibration, either global (one prediction model for all samples) or LOCAL (one prediction model per sample). Prediction accuracy was assessed according to validation r², standard error of prediction (SEP) in proportion to the mean, and ratio of standard deviation to SEP (RPD). Using both NIR and mid-IR spectroscopy, global calibration over the whole sample set yielded predictions that were excellent for C and N ( r² > 0.9, SEP < 20%, RPD ⩾ 3), good for C/N, acceptable for SIR, but poor for DEA. LOCAL calibration improved C/N and SIR predictions with both NIR and mid-IR spectroscopy, while DEA prediction became acceptable with NIR spectroscopy only. Additional improvement was achieved when LOCAL calibration was carried out over the fine-textured sub-set, especially for SIR ( r² > 0.9, SEP < 20%, RPD > 3). In contrast, LOCAL calibration over the coarse-textured sub-set was clearly not useful for improving prediction accuracy. NIR outperformed mid-IR spectroscopy whatever the variable, the calibration procedure and the sample set (except for SIR over the coarse-textured sub-set, where both performed similarly), suggesting its possible superiority for tropical soils.