Hyperspectral images are images acquired with narrow, continuous bands. This allows a more continuous spectrum to be collected than sensors that collect only collect few distinct bands. The continuous data of hyperspectral imagery allows for precise spectral measurements of the objects being analyzed.
Hyperspectral imagery can detect minute differences in color across landscapes. This creates the ability to take very accurate measurements and classifications.
We have collected a time series of hyperspectral imagery in years; 1999, 2003, 2006, 2009, 2010, and 2014. This is an unprecedented time series of the Lamar Valley in Yellowstone National Park. We are performing a series of classifications on these images
and after carefully aligning each image we can see how the landscape has changed. Even more interesting, these images were collected before, during and after a major drought, so we can also quantify the effect this drough had on an important wildlife corridor.
The classification technique we are using is called a "random forest" classification algorithm. This algorithm takes input training data (a.k.a. field measurements) and analyzes the associated locations spectrum collected fromt he sensor. It uses thousands of decision trees to find the best classification. This technique is useful for this project becuase it works well with sparse data. With such a long time series, our field data is sparse and inconsistent, so using a flexible classification technique is important.
Only 2010 and 2014 are currently available