Exponentially Weighted Moving Average Change Detection Around the Country (and the World)

Friday, 19 December 2014: 8:00 AM
Evan Brooks1, Randolph Hamilton Wynne1, Valerie A Thomas1, Christine E Blinn1 and John Coulston2, (1)Virginia Polytechnic Institute and State University, Blacksburg, VA, United States, (2)USDA Forest Service Southern Research Station Forest Inventory and Analysis Unit, Knoxville, TN, United States
With continuous, freely available moderate-resolution imagery of the Earth’s surface available, and with the promise of more imagery to come, change detection based on continuous process models continues to be a major area of research. One such method, exponentially weighted moving average change detection (EWMACD), is based on a mixture of harmonic regression (HR) and statistical quality control, a branch of statistics commonly used to detect aberrations in industrial and medical processes. By using HR to approximate per-pixel seasonal curves, the resulting residuals characterize information about the pixels which stands outside of the periodic structure imposed by HR. Under stable pixels, these residuals behave as might be expected, but in the presence of changes (growth, stress, removal), the residuals clearly show these changes when they are used as inputs into an EWMA chart. In prior work in Alabama, USA, EWMACD yielded an overall accuracy of 85% on a random sample of known thinned stands, in some cases detecting thinnings as sparse as 25% removal. It was also shown to correctly identify the timing of the thinning activity, typically within a single image date of the change. The net result of the algorithm was to produce date-by-date maps of afforestation and deforestation on a variable scale of severity. In other research, EWMACD has also been applied to detect land use and land cover changes in central Java, Indonesia, despite the heavy incidence of clouds and a monsoonal climate. Preliminary results show that EWMACD accurately identifies land use conversion (agricultural to residential, for example) and also identifies neighborhoods where the building density has increased, removing neighborhood vegetation. In both cases, initial results indicate the potential utility of EWMACD to detect both gross and subtle ecosystem disturbance, but further testing across a range of ecosystems and disturbances is clearly warranted.