Rice is one of the world’s most important staple crops, sustaining food security and rural livelihoods across many tropical countries. Accurate mapping of rice crop systems is therefore essential for monitoring agricultural production, supporting irrigation management, and assessing seasonal cultivation patterns. Reliable discrimination between crop systems depends not only on the quality of satellite observations but also on the time-series representation’s ability to preserve crop phenological characteristics. This study compares representative image encodings for rice crop system classification using Sentinel-2 Enhanced Vegetation Index (EVI) time series over Suphan Buri province, Thailand. The EVI profiles were reconstructed into daily observations using a Savitzky–Golay filter and encoded as images by the Continuous Wavelet Transform (CWT), Synchrosqueezing Transform (SST), Gramian Angular Field (GAF), Markov Transition Field (MTF), Recurrence Plot (RP), and Short-Time Fourier Transform (STFT), with the unencoded one-dimensional (1D) series as the baseline. To ensure a fair comparison, all representations were classified using the same convolutional neural network architecture and identical training samples, and evaluated on 200 independent ground truth points (with 50 samples per class). Overall accuracy ranged from 0.788 to 0.899 across the seven representations. STFT was the most accurate at 0.899, followed by CWT at 0.871, with a statistically significant difference between them. Confusion between the two most intensive systems accounted for about seven-tenths of the difference in accuracy between the best encoding and the baseline. Encoding alone was insufficient, since CWT did not significantly exceed the baseline, and SST, RP, GAF, and MTF fell below it. Encodings that preserve the interval between growth cycles are therefore preferable when classes are defined by cycle count rather than by signal amplitude.
Landslides present significant threats, inducing infrastructure and economic losses. Accurately predicting landslide susceptibility is essential for effective risk management. Traditional methods often struggle to address the complex and non-linear relationships among geological, hydrological, and meteorological factors. This paper uses geospatial data to propose an enhanced ResNet152 model that integrates multi-head self-attention mechanisms to improve landslide susceptibility mapping in Uttaradit Province, Thailand. The self-attention integration enables the model to capture complex dependencies within the input data, focusing on critical features that traditional convolutional layers might miss. This enhancement leads to a performance boost, with the ResNet152+Att model achieving a precision of 91.44%, recall of 91.74%, F1-score of 91.59%, and ROC-AUC of 0.9678, outperforming the baseline ResNet152. Using self-attention in deep residual networks significantly improves model accuracy and interpretability, providing a more robust framework for landslide susceptibility assessment. The implementation script is available at https://kaopanboonyuen.github.io/geoai-landslides.