Table 2 LAGOS-US LANDSAT workflow decisions and fit statistics for the Full-data model for each water quality predictive model.
From: LAGOS-US LANDSAT: Remotely sensed water quality estimates for U.S. lakes over 4 ha from 1984 to 2020
Variable | Workflow components tested and selected | Final model fit statistics | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Image and pixel processing by lake | Modeling steps | Matchup training data | |||||||||||
Pixel retrievals | Band ratio calculation | Scene quality: Negative reflectances | Scene quality: Scene clouds > 50% | In situ data transformed | # of bands and ratios | Training model type | Matchup window (±days) | Final matchups (N) | Variance explained (%) | RMSE | MAPE (%) | MAPE (%) Improvement of log10 model | |
CHL | Whole-lake | Pixel-specific | Removed | Removed | Log10 | All bands & ratios | Random forest | 1 | 43,755 | 44.1 | 0.43 | 23.3 | (+9.8) |
Secchi | Whole-lake | Pixel-specific | Removed | Removed | none | All bands & ratios | Random forest | 7 | 586,368 | 63.7 | 1.23 | 22.8 | — |
True color | Whole-lake | Pixel-specific | Removed | Removed | Log10 | All bands & ratios | Random forest | 7 | 33,194 | 48.6 | 0.26 | 7.5 | (+12.7) |
DOC | Whole-lake | Pixel-specific | Retained | Retained | Log10 | All bands & ratios | Random forest | 7 | 7,466 | 42.0 | 0.24 | 10.2 | (+6.1) |
TSS | Whole-lake | Pixel-specific | Removed | Removed | Log10 | All bands & ratios | Random forest | 1 | 10,845 | 20.7 | 0.62 | 27.0 | (+16.3) |
Turbidity | Whole-lake | Pixel-specific | Removed | Removed | Log10 | All bands & ratios | Random forest | 1 | 58,999 | 52.1 | 0.37 | 8.8 | (+20.9) |
