Potential Statistical Features of Sentinel-1 Time-Series Data for Rice Field Mapping in Central of Thailand

In this study, we present an approach to select statistical features from time-series Sentinel-1 SAR backscattering data, where we consider five statistical features: maximum, minimum, difference, variance, and mean. By combining three of these features, we create a false-color composite image that results in the highest accuracy for land-use classification. We define four land-use classes: paddy rice, unchanged area, water body, and high backscattering water surface. We include the last class because it exhibits activity similar to paddy rice, which can lead to confusion in classification. We experimented with three different approaches to combine statistical features, namely Approach I that combines maximum, difference, and minimum features, Approach II that combines maximum, variance, and minimum features, and our approach, Approach III, which combines mean, variance, and minimum features. We found that all three approaches produce high classification accuracies for paddy rice (>97%), unchanged area, and water body (98-99%). However, Approach I performs poorly for the high backscattering water surface class with an accuracy of only 2.3%, while our approach (Approach III) achieves the highest accuracy of 73%. Our study highlights the effectiveness of our approach in accurately classifying paddy rice and reducing confusion between paddy rice and high backscattering water surface. Further research on backscattering profiles of time-series data may provide additional insights.