Spectral-Temporal–Spatial Feature Optimization for Dioscorea Polystachya Turczaninow Classification Using Time Series Sentinel-2 Data

Dioscorea polystachya Turczaninow is one of the most famous traditional Chinese Materia Medica. However, there is a lack of large-scale classification method, which is crucial for its growth status monitoring and yield estimation. This study proposed a reliable D. polystachya Turczaninow classification model based on spectral-temporal–spatial feature optimization using time series Sentinel-2 data. First, 16 monotemporal classification models were developed using five vegetation indices (VIs) and random forest (RF) algorithm. Then, temporal feature optimization was conducted by identifying the most effective time phase combinations based on Sentinel-2 time series normalized difference vegetation index (NDVI) data, Gaussian mixture modeling algorithm, and $F1$ score of D. polystachya Turczaninow in each monotemporal model. Next, spectral features were optimized by replacing NDVI with the optimal VI corresponding to each time phase, thus constructing a multi-VI-based time series dataset. Finally, the spatial feature optimization was conducted using the 3-D convolutional neural network (3-D CNN) algorithm and the multi-VI-based time series Sentinel-2 data. Following the comprehensive feature optimization, the final D. polystachya Turczaninow classification model was determined. The results found that Sentinel-2 data acquired during the rhizome enlargement stage played a crucial role in classifying the D. polystachya Turczaninow. By using the optimized features, the classification model achieved the D. polystachya Turczaninow $F1$ score of 95.00%, which improved by 11.49% compared to only using the time series NDVI data. This spectral, temporal, and spatial feature optimization method also has the potential to the development of large-scale, dynamic, and accurate mapping for other crops.