Determination of patterns of rainfall history creating situations for accurate classification of rain-fed paddy fields with SAR backscatter coefficients

It is worthwhile to classify paddy fields based on the long-term frequency of ponding, which strongly determines the yield of rice. However, the direct measurement of the frequency requires a large number of backscatter images from synthetic aperture radar (SAR). Another approach is the indirect estimation of the relative frequency of ponding from differences in ponding frequency with the same rainfall history. Any combination of backscatter coefficients with such rainfall history should result in a high accuracy of classification of paddy fields based on the relative frequency of ponding. This study aims to establish methodologies for the classification of rain-fed paddy fields into two categories: 'wet' and 'dry' fields, based on ponding frequency using a small number of time-series SAR backscatter coefficients that depend on the flood situation and soil water content of the land surface. By analyzing 21 images acquired by the ALOS-PALSAR during 2007-2011, patterns of rainfall history with obvious large differences between the two categories of paddy fields were determined. Through supervised classifications using a random forest classifier, it was evident that the backscatter coefficients onset of rainy seasons for years with an above average precipitation, would be suitable as variables for accurate classification. In years where the precipitation (1379 mm/year) neared the average (1330 mm/year), the supplementary use of a quartile deviation of the backscatter coefficient within each plot, that indicated a degree of partial ponding, increased the classification accuracy. It was also evident that the effective rainfall index, which registers a rough estimate of rainwater acculturation in the soil, was useful in the determination of periods of significantly large variations in the backscatter coefficient between categories of rice fields.