With the deepening of oil and gas exploration in recent years, exploration targets have gradually shifted from structural reservoirs to lithologic reservoirs dominated by thin interbed reservoirs. The tunable effect of thin layer inhibits reflection energy of seismic wave at relatively high frequency and decreases the resolution of seismic data. The thinner the stratum is, the weaker the amplitude of the received seismic wave is. meanwhile, seismic waveforms will merge with each other, which further increases the difficulty of thin layer prediction. In this paper, a multi-scale iterative inversion technique is introduced to construct initial inversion model by introducing logging information from outside. Continuous wavelet transform is used to divide the frequency of logging and seismic data, and corresponding large, medium and small-scale signal components are obtained respectively. The constraint model of each scale is constructed by inverse distance weighting method. In the inversion process, Bayesian theory is introduced to modify the regularization parameters, and relationship between resolution and stability is adjusted adaptively to achieve the best balance of inversion results. The multi-scale iterative inversion technique improves initial model accuracy of the relative thin layer step by step on the premise of accuracy of thick layer inversion, and finally meets the purpose of weakening tuning effect, improving inversion resolution and effectively identifying thin layer. The technique has achieved good results in thin reservoir prediction in the slope zone of the Oriente Basin, Ecuador, with an average prediction accuracy of 87% for 10 ft thick reservoirs.