Spectra processing is crucial in metabolomics approaches, especially for proton NMR metabolomic profiling, since each processing step may impact the following steps. Among the different processing steps, data reduction (binning or bucketing) strongly impacts subsequent statistical data analysis and potential biomarker discovery. Based on a recently published work, we propose an improved method of data reduction, called ERVA which stands for Extraction of Relevant Variables for Analysis. This new method, by providing buckets centred on resonance peaks and rid of any non-significant signal, helps to recover the chemical fingerprints of metabolites. Moreover, we take advantage of the concentration variability of each compound from a series of samples of a complex mixture, to highlight chemical information. This is performed by linking the buckets into clusters based on significant correlations, thus bringing a helpful support for compound identification. As a proof of concept, this new method has been applied to a tomato 1H-NMR dataset to test its ability to recover fruit extract composition.