［Objective］ The precision of using coupled prediction models in quantifying microtopographic soil loss was evaluated for varying slopes， rainfall intensities， temporal scales， and surface treatments in order to provide scientific referrences for microtopographic soil loss prediction. ［Methods］ The microtopography of bare loess slopes was studied. The soil conservation service curve number （SCS-CN） model was modified according to measured surface runoff （QT） and surface roughness （SR） values for predicting runoff volume. The modified model was then coupled to the revised universal soil loss equation （MUSLE） for predicting soil erosion. ［Results］ ① The modified models， SCS-Q （using CN values back-calculated from QT） and SCS-SR （modified via SR）， predicted runoff （QCN and QSR， respectively） with R² values of 0.933 8 and 0.769 1， respectively. The developed models were more accurate by 32% and 9%， respectively， than the original SCS-CN model， which predicted runoff， QO， with an R² of 0.705 6. ② The coupled models produced more accurate microtopographic erosion predictions than the RUSLE models. The MUSLE-SR model （combining SCS-SR with MUSLE） was more accurate （NSE ∈ ［0.50， 0.94］） than the MUSLE-Q model （combining SCS-Q with MUSLE） （NSE ∈ ［0.23， 0.94］）. ③ During the microtopographic soil loss prediction， the surface measures more strongly affected the microtopographic erosion prediction accuracy of the coupled model （ΔNSE=63%） than the rainfall intensity （ΔNSE=52%） or slope （ΔNSE=40%）. ［Conclusion］ The microtopographic soil erosion prediction accuracy of the coupled models considerably increased with rainfall duration： accuracy was low in the first 20 min of rainfall （R²<0.5） but markedly increased after 20 min （R²>0.8）.