This paper presents a data-driven approach for the flattening of atomic force microscopy images. This acquisition process is widely used for the characterization of surfaces at the nanoscale. Unfortunately the technique is subject to several defects such as linear deviation and jumps. We present a data-driven methodology, fast to implement and which accurately corrects these deviations. The proposed methodology makes no assumption on the object locations and can therefore be used as an efficient preprocessing routine. Experimental results show the efficiency of the methodology compared to existing softwares.