Deep learning-based phenology extraction and crop classification in arid oasis using Sentinel-2 time series

Area of Science:

• Agricultural Remote Sensing
• Machine Learning for Agriculture
• Precision Agriculture

Background:

• Arid oasis regions present complex challenges for agricultural monitoring due to intricate cropping systems.
• Multi-temporal remote sensing, particularly Sentinel-2 imagery and NDVI time series, is crucial for precision management.
• Accurate crop identification and phenology monitoring are vital for optimizing agricultural practices in these areas.

Purpose of the Study:

• To develop and validate a deep learning framework for precise crop mapping in arid oasis regions.
• To extract key phenological parameters (SOS, POS, EOS) using advanced signal processing techniques.
• To compare the performance of various deep learning models for crop classification using multi-temporal Sentinel-2 data.

Main Methods:

• Utilized Sentinel-2 multi-temporal imagery and NDVI time series data.
• Applied minimum redundancy maximum relevance (mRMR) for feature selection, followed by Savitzky-Golay filtering and double logistic fitting for phenological parameter extraction.
• Integrated multi-scale feature fusion and compared five classification models: MLP, ResNet-18, ConvLSTM, Transformer, and RFC.
• Optimized Transformer model using multi-scale convolutional windows (1×1+3×3+5×5).

Main Results:

• The deep learning framework achieved high accuracy in identifying and mapping winter jujube, cotton, and tiger nut crops.
• Phenological feature extraction accuracy was significantly improved by the integrated signal processing techniques.
• The Transformer model, with optimized spatial representation, demonstrated superior performance and computational efficiency.
• Independent validation confirmed robust model transferability with high F1 scores (94.37% for winter jujube, 87.75% for cotton, 86.35% for tiger nut).

Conclusions:

• Sentinel-2 temporal data combined with deep neural networks offer high-precision crop identification capabilities in complex multi-crop environments.
• The developed framework enables precise spatial mapping of crop distributions, supporting smart agricultural decision-making.
• This study provides a robust methodological foundation for agricultural monitoring and management in arid oasis regions.