In the process of synchronous ground observation for quantitative remote sensing inversion, measurement uncertainty factors like human subjective factor, environmental change and condition restriction will induce data noise inevitably, which degrades the retrieval accuracy of the suspended matter concentration. Therefore, a measurement uncertainty-aware retrieval method named as the adaptive sample consensus extreme learning machine (ASAC-ELM) is proposed. ASAC-ELM integrates the merits of extreme learning machine (ELM), random sample consensus (RANSAC) and N adjacent points sample consensus (NAPSAC). The algorithm adaptively selects RANSAC or NAPSAC to estimate model parameters with the guidance of the parameter dimension, which avoids the problem that the ELM algorithm is sensitive to the non-zero normal distributed data noise. The ASAC-ELM algorithm selects inlying points (non-noise points) for model construction, thus can remove the interference from noise, and enhance the accuracy and flexibility of the model. In order to investigate the effectiveness of the proposed method under different noise conditions, a series of additive noise with non-zero mean normal distribution is introduced in the training data. The comparison among ASAC-ELM, ELM and traditional back propagation (BP) neural network algorithms is also conducted. The results show that for the retrieval of inland water suspended matter concentration under various noise conditions, the inversion accuracy and stability of ASAC-ELM is higher than those of ELM and the traditional BP neural network.