Standardized Time-Series and Interannual Phenological Deviation: New Techniques for Burned-Area Detection Using Long-Term MODIS-NBR Dataset

Abstract

Typically, digital image processing for burned-areas detection combines the use of a spectral index and the seasonal differencing method. However, the seasonal differencing has many errors when applied to a long-term time series. This article aims to develop and test two methods as an alternative to the traditional seasonal difference. The study area is the Chapada dos Veadeiros National Park (Central Brazil) that comprises different vegetation of the Cerrado biome. We used the MODIS/Terra Surface Reflectance 8-Day composite data, considering a 12-year period. The normalized burn ratio was calculated from the band 2 (250-meter resolution) and the band 7 (500-meter resolution reasampled to 250-meter). In this context, the normalization methods aim to eliminate all possible sources of spectral variation and highlight the burned-area features. The proposed normalization methods were the standardized time-series and the interannual phenological deviation. The standardized time-series calculate for each pixel the z-scores of its temporal curve, obtaining a mean of 0 and a standard deviation of 1. The second method establishes a reference curve for each pixel from the average interannual phenology that is subtracted for every year of its respective time series. Optimal threshold value between burned and unburned area for each method was determined from accuracy assessment curves, which compare different threshold values and its accuracy indices with a reference classification using Landsat TM. The different methods have similar accuracy for the burning event, where the standardized method has slightly better results. However, the seasonal difference method has a very false positive error, especially in the period between the rainy and dry seasons. The interannual phenological deviation method minimizes false positive errors, but some remain. In contrast, the standardized time series shows excellent results not containing this type of error. This precision is due to the design method that does not perform a subtraction with a baseline (prior year or average phenological curve). Thus, this method allows a high stability and can be implemented for the automatic detection of burned areas using long-term time series.