The Lost Meadows Model uses random forest machine learning to find places in the Sierra Nevada that share the hydrogeomorphic signature of existing riparian (stream-associated) meadows: low-gradient settings where surface water slows and groundwater collects. Many predicted areas are "lost meadows" that likely held meadows in the past and filled in with forest as channels incised and water tables dropped. Others share meadow-like geomorphology but may never have been meadow. Click a watershed to download its predictions or view them on the map.
Each pixel gets a score from 0 to 1 for how closely it resembles meadow conditions, and we draw polygons at two thresholds. The high-confidence set uses a stricter cutoff: fewer polygons, with more certainty that each marks meadow-like ground. The medium-confidence set uses a looser cutoff and picks up more candidate areas. Both cutoffs are deliberately conservative, so the model tends to miss real meadows rather than over-predict them. These polygons under-represent likely meadow habitat rather than over-represent it.
The model was trained within a 60-watershed area of the Sierra Nevada, outlined in purple, and validated there against an independent meadow dataset and high-resolution imagery. Read predictions outside that area with more caution. Meadow formation depends heavily on elevation and geology, and the model has not been validated beyond the Sierra Nevada, so where a watershed's elevation or bedrock differs from the training area its predictions are exploratory. Hosting these layers for download does not mean the authors endorse any particular polygon. Treat them as a planning aid, not ground truth.
Click a watershed to download its prediction files.
Lost Meadows random-forest model (Sierra Nevada model)
About this map · Files on Google Drive