“…The best predictive models for each fraction compared favourably with other predictions of soil texture based on IR spectra; for example, for predicting clay content, R 2 of 0.56 (Ben‐Dor & Banin, ), 0.67 (Viscarra Rossel et al ., ), 0.88 (Viscarra Rossel & Behrens, ) to up to 0.93 (Lacerda et al ., ); for predicting sand content, R 2 of 0.75 (Viscarra Rossel et al ., ) or 0.96 (Lacerda et al ., ); for predicting silt content, R 2 of 0.52 (Viscarra Rossel et al ., ). The good prediction performance was further emphasized by the large RPDs; the best models had RPDs well above the cut‐off value of 2.0 used to distinguish unreliable from reliable models (Chang et al ., ), and well above the values of 2.35, 1.63 and 2.06 reported for clay, silt and sand content in Australia (Viscarra Rossel & Webster, ) or 3.17 for clay content in Brazil (Lacerda et al ., ). Similarly, the RMSEs of the best performing models were around 4–5% for clay and silt, indicating excellent prediction accuracy (Stenberg et al ., ; Viscarra Rossel & Behrens, ).…”