On July 6, 2020, the Rangeland Analysis Platform will switch to cover 2.0 as the default product for rangeland cover. Cover 1.0 will continue to be available at https://rangelands.app/coverV1/
Cover 1.0 was released in summer 2018. Since then there have been advances in data handling and modeling, creating new opportunities for an improved product. More on-the-ground data have also been collected and released by the federal monitoring programs. Cover 2.0 was developed with the intent to create a new product that uses the best available data, modeling, and resources.
The primary change is the use of a neural network to predict rangeland cover. This approach better handles the continuous time series measurements from Landsat satellites and also produces a multivariate response of cover. That is, the interactions among functional groups are learned and predicted together, resulting in an improved model that reflects real world interactions among functional groups, e.g., an increase in tree cover results in a decrease in grass and forb cover. These interactions are defined by the data, eliminating the need for predetermined conditions or rules.
Other changes include increased on-the-ground plot data and Landsat measurements, the removal of non Landsat predictors (e.g., weather), and the production of spatially explicit, pixel-level estimates of model uncertainty. The table below summarizes the major changes in cover 2.0.
|Cover 1.0||Cover 2.0|
|Model||Random Forest||Convolutional Neural Network|
|Data||31,300 plots||57,792 plots|
|Landsat measurements||3; seasonal averages||12; 64 day averages|
|Predictors||40; variable per group||10; Landsat based|
|Prediction pipeline||Google Earth Engine||Google Earth Engine
Google AI Platform
Cover 2.0 error metrics decreased, indicating greater predictive accuracy. Futher discussion of error metrics can be found in the preprint.
|Vegetation cover||Annual forbs
|Mean absolute error (%) cover 2.0||7.0||10.3||5.8||2.8||6.7|
|Root mean square error (%) cover 2.0||11.0||14.0||8.3||6.8||9.8|
|Mean absolute error (%) cover 1.0||7.8||11.1||6.9||4.7||7.3|
|Root mean square error (%) cover 1.0||11.8||14.9||9.9||8.5||10.6|
A preprint describing the methods and evaluation results of cover 2.0 is available on bioRxiv.
Cover 1.0 will be produced and maintained for a period of at least two years after this update. It can be accessed at https://rangelands.app/coverV1/
Major version updates (e.g., 1.0 to 2.0) will be assessed on a two year cycle, and will occur only if there are significant changes, i.e., model changes, additional features, etc. The entire time period will be produced to ensure consistency. Previous major versions (e.g., 1.0) will be produced and maintained for a period of at least two years after updating. A major update will supersede a minor update.
Minor version updates (e.g., 2.1) will be carried out on a two year cycle, and will primarily include training data updates or additions, new Landsat sensors, etc. The entire time period will be produced to ensure consistency. One year cycles may be considered if determined valuable. Previous minor versions will be produced and maintained for a period of at least two years after updating.