A HMM-based approach for historic and up-to-date land cover mapping through Landsat time-series in the state of Sao Paulo, Brazil

Abstract

The paper debates a novel approach for land cover (LC) mapping based on the Hidden Markov Model. The proposed methodology is aimed to address both the urgent demand of off-line (or historic) LC information retrieval and of near-real time LC monitoring. The discrete-time model employs short steps of 16 days, that conveniently fits the Landsat revisit time while providing a continuous and temporally dense representation of the land cover dynamics. Two temporal pattern typologies were identified and modeled within the proposed Markov chain architecture: a seasonal and synchrounous behavior which can be associated to the observables of LC classes such as forest and grasses, and a highly asynchronous behaviour, which characterizes the crop observables. The first typology is addressed by introducing time-dependency in state output probabilities, whereas the latter is rendered through a sequence of (sub-class) states interlinked by means of a 'left-right' based model. Such model inherently incorporates crop growth tracking functionalities as an added value. In this paper the methodology has been tailored to Sao Paulo state (Brazil) scenario, showing overall accuracy above 80% on the test sample. A particular emphasis is attributed to the identification of sugarcane plantations, that are indeed responsible for major land use changes.