Hidden Markov Models for Annual Land Cover Mapping—Increasing Temporal Consistency and Completeness

Abstract

This article aims at investigating the hidden Markov model (HMM) approach for the automated processing of classified satellite images for land cover and land-use change (LCLUC). HMM’s account for transitions between classes at the same location, but that cannot be directly observed due to classification errors. Using a set of transition and emission probabilities, HMM’s allow filtering out errors and recovering the actual sequence of LCLUC, which are typically overestimated when directly estimated from the classified images. After presenting the HMM framework, the methodology is illustrated on three 300-m annual time series of classified images from 2003 to 2019 over <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$756\times756$ </tex-math></inline-formula> km <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> areas in Brazil, People’s Republic of China, and Mali. It is shown how the emission and transition probabilities can be estimated from these time series using a simple Viterbi training, alleviating computationally demanding algorithms. Special attention is paid to the processing of missing observations caused by clouds. Combining these three datasets with a simulation study, it is concluded that the HMM emission and transition probabilities can be estimated with low biases and variances thanks to the vast number (hundreds of thousands) of pixels at hand. The speed of the Viterbi training and decoding steps makes it possible to consider large-scale land cover mapping at moderate or even high spatial resolution as long as the legend of the LCLUC involves a reasonable number of classes like the six main Intergovernmental Panel on Climate Change (IPCC) land categories.