Fig. 7
From: LAGOS-US LANDSAT: Remotely sensed water quality estimates for U.S. lakes over 4 ha from 1984 to 2020
Fit statistic plots for the Holdout-data model of each of the six water quality variables in original units. The top six plots show predicted versus in situ quantile lines for the median quantile (black) and the 5th, 25th, 75th, and 95th quantiles (shaded colored lines). The maximum range of the X and Y axes is trimmed to the 90th percentile of in situ data (CHL = 53 ug/L; Secchi depth = 5.3 m; true color = 63 PCU; DOC = 13 mg/L; Turbidity = 19 NTU; TSS = 133 mg/L). The dashed red line is the 1:1 line. The internal random forest OOB variance explained from the model training data as well as the r-squared value of the holdout data are shown. The bottom 6 plots show the absolute percent error between predicted and in situ values in the model training dataset by intervals of predicted values for the training dataset. The red dashed line is the threshold identified as above which predictions are thought to be less reliable. Outliers are not plotted.
