Assessment of Bayesian structure of hidden Markov model for real time prediction of maize phenology

The Crop Progress Percentage (CPP) in a given phenology stage reflects growth status in life cycle. Generally, routine field measurements of this parameter are lacking, hence various alternative approaches have been proposed for its estimation. The statistical methods such as Bayesian approaches and hidden Markov models (HMMs) have appropriate structural skills for solving problems with variety of continuous or discrete data and can be combined with remotely sensed data also. The aim of this study is evaluation of hidden Markov models' skill in real time prediction of maize progress percentage in research field of university of Tehran located in Karaj. The HMMs follow the Bayesian structure in which, there are usually two layers; hidden and observable. Different phenological stages including Emergence to Milky were considered as the hidden layer and both Normalized Difference Vegetation Index (NDVI) and Accumulated Growth Degree-Day (AGDD) features, extracted from air temperature and LANDSAT7 ETM+ images, as a vector variable of observable layer. Calibration and evaluation of the model was performed using a 9 years (2002-2010) data set of the field phenology observations and meteorological data. According to the results, in general, for all phenological stages, the HMM was able to estimate the CPPs with average RMSE of 14%, which confirms the applicability of this approach as a suitable tool.  Further studies in other climatic regions of the country are recommended for more scrutiny of phenological prediction models using remotes sensing and statistical approaches.