Detection and characterization of oil palm plantations through MODIS EVI time series

Oil palm is a perennial tree that well fits the humid tropical climate; fresh fruit bunchesare the palm raw fruit for oil mills. Palm oil is the world highest yielding oil crop determining that palms are extensively planted in South-East Asia, especially in Malaysia, Thailand, and Indonesia where plantations have been spreading in response to the increasing market demand. Cultivation of oil palm in tropical countries is an important economic factor, but it has already proved of endangering biodiversity and degrading environment with a global impact related to forest loss. Remote sensing well fits requirements of precision farming that many stakeholders involved in palm oil production are currently approaching to decrease or monitor environmental impacts. In this work, an enhanced vegetation index (EVI) time series of 415 images was obtained from the MODIS Vegetation Index 16 days composite product (MOD13Q1-v5) to explore tropical vegetation changes. The EVI time series covers a period of 18 years; it was processed aiming at mapping new oil palm plantations in the reference period, giving an estimate of their age, production, and economic value. In this work, a new methodology for oil palm detection and characterization was presented based on local EVI temporal profile analysis. Pixel EVI temporal profile proved to be effective in describing both vegetation macro-phenology and forest loss at that position. Consequently, the proposed algorithm looks for abrupt changes along the local EVI time series (sudden decreasing). The minimum EVI value recorded in the detected changing period is assumed as a predictor of the starting date of new plantations, being the latter reasonably related to forest loss and preliminary soil preparation. Starting date is then used by the algorithm to estimate oil palm age and, consequently, the present local (potential) production. Accuracy assessment showed an overall accuracy in new palm oil plantation detection of about 94%. Starting age estimation proved to be accurate enough: 76% of the estimates, in fact, were placed in a range of uncertainty of 1 year.