Integrating Smoothing Techniques with Convolutional Neural Networks for Rice Cropping Systems Classification in Suphan Buri, Thailand

Abstract

This study compares smoothing methods for classifying rice cropping systems in Suphan Buri, Thailand, using enhanced vegetation index (EVI) time series from Sentinel-2 imagery between 2023 and 2025. Three smoothing techniques: Savitzky–Golay (SG), locally estimated scatterplot smoothing (LOESS), and Gaussian smoothing are evaluated. Using continuous wavelet transform (CWT), the smoothed EVI time series are converted into a two-dimensional (2D) time-frequency representation, or scalograms. Results demonstrate that Gaussian smoothing provides the most stable and reliable representation of crop growth dynamics, achieving an overall accuracy (OA) of 0.908 and a kappa coefficient of 0.877. The classification effectively maps single crop (SC), double crop (DC), two-and-a-half crop (HC), and triple crop (TC) systems, consistent with local irrigation conditions and agricultural practices. This framework enhances the reliability of rice cropping system mapping and facilitates operational rice monitoring. It also informs crop insurance assessment and irrigation management in Thailand and other climate-constrained regions.