Particle filter approach for crop phenological stage estimation using time series of NDVI images

The Normalised Difference Vegetation Index (NDVI) is well suited to provide the phenological state of crops, but time-series of NDVI data gathered by satellites are usually too noisy to be used directly. Common approaches apply signal processing steps based on fitting analytical functions and then a few stages can be estimated by defining thresholds along the fitted curves. Dynamic tools constitute an alternative way to exploit these NDVI data with the advantages of providing real-time estimates which are defined in a continuous range of values. In this work, we propose a particle filter method to retrieve phenological states of rice crops using time-series of NDVI satellite observations. This methodology achieves a high determination factor for the complete phenological range. The end of season dates are estimated and compared with different fitting-based methods, providing much better results.