Remote Sensing of Lake Water Clarity: Performance and Transferability of Both Historical Algorithms and Machine Learning

There has been little rigorous investigation of the transferability existing empirical water clarity models developed at one location or time to other lakes and dates imagery with differing conditions. Machine learning methods have not widely adopted for analysis lake optical properties such as clarity, despite their successful use in many applications environmental remote sensing. This study compares model performance a random forest (RF) machine algorithm simple 4-band linear 13 previously published non-machine algorithms. We Landsat surface reflectance product data aligned spatially temporally co-located situ Secchi depth observations from northeastern USA over 34-year period this analysis. To evaluate across space time, we compare fit using complete dataset (all images samples) single-date approach, which separate are each date more than 75 field samples. On average, all algorithms had lower mean absolute errors (MAE) root squared (RMSE) dataset. The RF highest pseudo-R2 approach well dataset, suggesting that an outperforms traditional regression-based when modeling satellite imagery.