A Neural-Network Method to Evaluate Data Quality

Quality control (QC) of observational datasets has been a challenge for many years. The chief obstacle is the need to have a reference dataset and/or empirical rules to compare with the observations. In the case of our system -- a portable surface flux network -- rules are difficult to create since conditions and even sensors change significantly from deployment to deployment.

To address this problem, we have developed a neural-network methodology. The network is trained with other sensor data from the initial stages of the project and allowed to predict the sensor output for the remainder of the study. Clearly, a lot of physics is masked (and assumed!) by this process. Thus, we are only using the neural network to flag suspect data. We are not using it to "fill in" data gaps.

We will show example datasets comparing manual and neural-network QC processing.