Abstract: Multi-temporal remote sensing data in large-scale crop phenology identification and classification has become increasingly utilized, particularly for precision management in arid oasis agricultural regions with complex cropping systems. In this study, we developed a deep learning framework integrating Sentinel-2 multi-temporal imagery and normalized difference vegetation index (NDVI) time series for mapping cotton, winter jujube and tiger nut crops in Tumushuke City, Xinjiang, China. We employed the minimum redundancy maximum relevance (mRMR) algorithm for spectral and vegetation index feature selection, followed by Savitzky-Golay (S-G) filtering and double logistic function fitting, to automatically extract the key phenological parameters: start of season (SOS), peak of season (POS), and end of season (EOS), significantly improving phenological feature extraction accuracy. By incorporating multi-temporal Sentinel-2 data and a multi-scale feature fusion approach, we could systematically compare five classification models (MLP, ResNet-18, ConvLSTM, Transformer, and RFC), demonstrating that high-resolution spatial details substantially enhance crop boundary delineation and classification consistency in complex environments. Further optimization of Transformer's spatial representation through multi-scale window analysis revealed that 1×1+3×3+5×5 convolutional windows achieves optimal balance between accuracy and computational efficiency. Independent validation on unseen areas confirmed robust model transferability, with F1-scores of 94.37%, 87.75% and 86.35% for the three crops, respectively. This study validates the high-precision identification potential of Sentinel-2 temporal data and deep neural networks for multi-crop environments, enabling the precise spatial mapping of crop distributions and providing methodological support for smart agricultural decision-making in arid oasis regions.