On a Data-Driven Approach for Detecting Disturbance in the Brazilian Savannas Using Time Series of Vegetation Indices

Abstract

Remote sensing of disturbance in the savannas from Brazil is challenging, especially due to confounding effects of the vegetation phenology and natural soil exposure on the detection of clearing and fire events. In this study, we investigated the detection of disturbance over this global hotspot of biodiversity using seven vegetation indices (VIs) calculated from the Landsat time series (2017–2019) and the Continuous Change Detection and Classification (CCDC) algorithm. The selected VIs represented distinct biophysical characteristics of the savannas. We evaluated the effects of disturbance on these VIs and assessed the accuracy of CCDC-detection in 2019, considering individual VIs, ensemble VIs, and the type of disturbance (savanna clearing and fire). Finally, we analyzed the possible existence of seasonal patterns of disturbance in a study area located at the new agricultural frontier of the Cerrado biome. The results showed that the overall accuracy of CCDC detection of total disturbance ranged from 51.2% for the Green-Red Normalized Difference (GRND) to 65.9% for the Normalized Burn Ratio (NBR2). It increased to 71.2% for ensemble VIs, whose multivariate approach reduced the omission errors in the analysis when compared to the use of single VIs. For detecting events of savanna clearing and fire, the most important VIs used near-infrared and shortwave infrared reflectance bands on their formulations (NBR2, NBR, and Moisture Stress Index—MSI). The CCDC accuracy was generally higher for detecting clearing than for mapping burned areas. In contrast, the recorded date of disturbance occurrence was less precise for detecting clearing than for recording events caused by fire, especially due to the existence of some gradual processes of vegetation degradation until complete clearing. Our findings showed also the existence of a seasonal pattern of disturbance occurrence. Savanna clearing predominated in the transition from the rainy to the dry season (April to July) to open new areas for agriculture. It preceded most events of fire disturbance between August and October that occurred near the consolidated areas of agriculture and extended into the native vegetation areas. Results reinforce the importance of data-driven approaches for generating early warning alerts of disturbance in the Cerrado to be further checked in the field.