We demonstrate in this work online in situ characterization of potash fertilizer, a powder material, at its final production stage in factory on the production conveyer belt for quality assessment, with a specifically developed laser-induced breakdown spectroscopy (LIBS) instrument and dedicated data treatment software based on machine learning. Besides the usual difficulties encountered in online LIBS analysis, the specific challenge resides in moisture variation in the product, which results in a complex sample of powder of particle size ∼100 μm mixed with water (H₂O). The influence on the LIBS spectrum was clearly observed, while no detailed physical model is available to describe such an influence. In addition, the emission line intensity from hydrogen (Hα line) observed in the spectrum did not show a clear relationship to the H₂O concentration. The approach of analysis by correlation of the whole spectrum to the concentration was used to first determine the H₂O concentration, which was further used as an additional parameter to concatenate with a LIBS spectrum in the formation of a generalized spectrum. The last was used as the input vector to train a potassium chloride (KCl) concentration calibration model. More specifically, LIBS spectra were first transformed into 2-D images with continuous wavelet transform (CWT). A convolutional neural network (CNN) then allowed mapping of the spectrum-images to the H₂O concentrations of the corresponding samples, while a back-propagation neural network (BPNN) mapped generalized spectra to the KCl concentrations of the samples. The tests with online LIBS spectra and the corresponding offline analysis data of 119 samples taken during the period of LIBS measurements demonstrate advanced analytical performances of the trained models for H₂O and KCl. Comparison between the model-predicted concentrations and the data from the offline analysis shows determination biases which fulfil the requirements of the concerned national standards (bias ≦0.20% for H₂O and ≦0.598% for KCl) for the quasi totality of the tested samples.