A weakly-supervised data-driven approach, improved similarity-based modeling (SBM) approach, was proposed aiming at the problem of dust deposition monitoring of photovoltaic modules. The great error of theoretical formulas approaches, the high cost of computer vision approaches and the large requirements for training data of classical machine learning approaches were considered. Improvements in the steps of parameters selection, state matrix construction and update and similarity operator were made based on the original SBM in order to adapt the approach to the PV systems monitoring and increase the estimated accuracy and response rate. The improved SBM and five other approaches, theoretical formulas, feedforward neural network (FNN), support vector regression (SVR), random forest (RF), and the original SBM were employed for the dust estimation, and their performances were evaluated based on the real data from a dust deposition experiment. Results show that the improved SBM has the best accuracy at the cost of acceptable response time increase.