Rice cultivation and harvest date identification based on a hidden Markov model

Rice cultivation and harvest dates are very useful information since they are the key factors in rice monitoring, yield estimation and damage assessment. This paper proposes a new approach to estimate rice cultivation and harvest dates by using the 8-day composite normalized difference vegetation index (NDVI) derived from Moderate Resolution Imaging Spectroradiometer (MODIS) data. However, the NDVI time-series data suffered from cloud contamination. Using the filter to reconstruct to the cloud-free NDVI data can introduce the artifact that may result in incorrect estimation of cultivation harvest dates. As a result, we employ the hidden Markov models to characterize the rice growth states and atmospheric conditions. Here, we divide the rice growth states into 4 states, nothing, growing, mature, and harvest in which two atmospheric conditions, namely, the clear and cloudy skies can occur. The optimum growth states and atmospheric conditions are determined using the Viterbi algorithm. In the experiment, we compared with the ground truth data with the estimated cultivation and harvest dates, and found the average errors for cultivation dates and harvest dates of the rain-fed rice 16.128 days and 8.734 days, respectively. For the irrigated rice, the errors are 17.524 days and 12.516 days for cultivation and harvest dates, respectively.