To what extent can we accurately predict water depth from overhead images using linear and non-linear methods, in a scaled laboratory estuary environment

Abstract

Physical scale experiments improve our comprehension of fluvial, tidal, and coastal processes. However, acquiring accurate, precise, and continuous data on water depth has been difficult due to the limitations inherent in the measuring equipment. In this paper, we try to predict water depth in a two-fold process using different models. First we address the issue of zero-inflated data. As over 50% of the training data cells were calculated to contain no water. We aim to develop a binary classification model that predicts pixels with no water and with water, then use regression to predict the depth of cells with water.

Binary classification models such as Logistic Regression, Random Forests, and SVM were compared. These models are trained using overhead imagery from water depth calibration experiments. The overhead imagery was overlaid on water depth maps. The water depth maps were calculated from the calibration experiments using the DEM and increasing weir heights.

In the second step, regression models were developed and compared. Linear regression, Random Forests, and SVM models were trained on the pixels containing water in the water depth maps. The two-step Random Forest model was selected and applied to different overhead images under similar experimental conditions, the resulting predicted water depth maps produced promising results and closely resembled the overhead imagery.

The implication of the experimental data–model integration is that future experiments can derive water depth from overhead imagery in a simple, affordable, and labour-efficient manner.