|
Continuous wavelet transform based CNN modeling for rice crop system mapping in Suphan Buri, Thailand |
|---|---|
| รหัสดีโอไอ | |
| Title | Continuous wavelet transform based CNN modeling for rice crop system mapping in Suphan Buri, Thailand |
| Creator | Woraman Jangsawang |
| Contributor | Teerayut Horanont, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Continuous wavelet transform, Convolutional neural network, Rice cropping systems, Time series reconstruction, Sentinel-2, Remote sensing, Classification |
| Abstract | Accurate mapping of rice cropping systems is essential for agricultural planning in Thailand, particularly in regions affected by frequent cloud contamination that distorts vegetation index (VI) time series. This study presents a method for classifying rice cropping systems in Suphan Buri Province by integrating reconstructed Enhanced Vegetation Index (EVI) time series data with the Continuous Wavelet Transform (CWT) and a Convolutional Neural Network (CNN). The EVI time series was reconstructed using an adaptive-weight spline smoothing pipeline to produce stable and interpretable phenological profiles. The reconstructed profiles were transformed into CWT scalograms, which captured the characteristic time–frequency signatures of each cropping system and served as structured two-dimensional (2D) inputs for CNN-based learning.In comparison, the baseline Random Forest (RF) model was performed to evaluate its performance against the proposed CNN with CWT approach. The RF, trained on raw EVI sequences, achieved an overall accuracy of 0.74 and a Cohen’s Kappa of 0.649, with noticeably lower F1-scores for the two-and-a-half and triple cropping systems (HCR and TCR). In contrast, the CNN model attained an accuracy of 0.88 and a Kappa of 0.837 and produced more spatially coherent and irrigation-consistent patterns. These findings indicate that the use of CWT scalograms enables the extraction of distinctive temporal–frequency features, allowing convolution-based learning to better discriminate cropping-cycle patterns. This combination provides a more reliable framework for rice crop system mapping in cloud-prone environments. |
