Interpreting transition and emission probabilities from a Hidden Markov Model of remotely sensed snow cover in a Himalayan Basin

Remote sensing is often used to monitor snow cover in areas without sufficient in-situ observations. However, a common problem with remotely sensed snow cover products is the mis-classification of snow cover due to obscuration of the land surface by cloud, along with the similar spectral characteristics of snow and cloud. In a two part preliminary investigation we assessed the ability of a Hidden Markov Model (HMM) to reduce classification errors in optical snow cover mapping along with the transition and emission probabilities output from the model, to our knowledge for the first time documented. This research focuses on the latter and was conducted within the Sustainable Development Investment Portfolio (SDIP) at CSIRO. A Hidden Markov model can utilise a series of input observations and calculate the probability that they represent the ground state. These probabilities are then used to model the most likely series of states, this effectively provides a dynamic filter that can mitigate the problems faced when remotely sensing snow in mountainous areas. As part of a larger project, we applied this approach to snow cover mapping over a single sub-basin in the Himalayas in Eastern Nepal based on imagery from the (MODIS) instruments on the Terra and Aqua satellites. This study analysed spatially mapped transition and emission probabilities extracted from a two state Hidden Markov Model. The ability of a Hidden Markov model to employ a dynamic filter that can utilise entire sequences of observations will likely offer improved accuracy when compared to other time-series filtering methods. This is because it is able to utilise a much larger number of observations without compounding losses in accuracy or causing reductions in temporal resolution. Our probability analysis shows the potential for a HMM approach to provide a robust and flexible method for processing ‘noisy’ data such as remotely sensed snow cover measurements. The improvement of spatio-temporal snow cover measurements has broader implications for hydrological modelling, particularly in countries dependent on snow melt for subsistence agriculture and hydroelectric facilities such as Nepal. Improvements also benefit the long-term analysis of snow cover trends which are an important proxy for assessing the impacts of climate change in sensitive mountain areas. Further studies should apply this method to multiple study sites and quantitatively compare it to other cloud-cover reduction techniques for snow cover imagery.