[en] Manual validation of all predicted items recently became a standard practice in automatic recognition of plankton. This semi-automated identification, although faster than fully manual processing of the samples, remains time-consuming. We present here an original method where only the most suspect, automatically identified, particles are manually validated. Statistical analysis of this partial validation, combined with a mathematical modeling of the error, allows to optimize the predication of the abundance per group. As an example, this technique is applied to an artificial mixture of known composition containing 5 different groups digitized with the FlowCAM and classified by Zoo/PhytoImage with 18% of global error (from 32% to 10% per group). Manual validation of the 20% most suspect particles decreases the global error down to 9% (from 19% to 1% per group). Statistical correction of the remaining error further reduces the error down to 3% (from 12% to 0.2% per group). Moreover, ecological or biological information can be used to better detect suspect particles. As far as we know, it is the first time that such metadata are used to correct prediction error.