Tuning the Extended Kalman Filter for Rice Cultivation Date Estimation in Tropical Area using MODIS NDVI data

Accurate crop cultivation date estimation is crucial for crop yield forecast, parasite attack trend forecast and for suitable government aid calculation in case of a natural disaster. The 8-day composite of the Normalized Difference Vegetation Index (NDVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS) data is a good candidate to be used thanks to its global coverage and its free availability. Its spatial resolution is at 250 meter. However, in a tropical country as Thailand, the NDVI data is very noisy due to high cloud density. An appropriate estimation technique and tuning method must therefore be used to cope with the high measurement noises. An Extended Kalman Filter (EKF) was proposed in the literature for the rice cultivation date estimation in Thailand. The NDVI was modelled by a triply modulated cosine function with the mean, the amplitude and the initial phase as state variables. The estimated cultivation date was counted as the date where the seasonal part of the estimated NDVI, i.e. the estimated NDVI minus its mean, is higher than a certain threshold. However, the brute force algorithm in which the large numbers of tuning parameters were tested until the smallest mean error was obtained was used. The achieved tuning parameters were not only not consistent with the nature of the NDVI but might also be not appropriate for the study areas where the ground truths are not available. In this paper, we propose a method for tuning the EKF which is consistent with the nature of the NDVI data. The mean error of the estimated cultivation dates is at 21 days compared to 25 days of the existing method in the literature. 1