Abstract: 【Objective】 To fully explore the temporal and spatial information of remote sensing images, accurately identified crop types on land parcels, which could provide reliable data support for crop type spatial distribution mapping,yield estimation and agricultural production decision-making, etc. 【Method】 Using Google Earth images as a reference,obtained the complete boundary of the study area of Kings County, California, United States, and used multi-temporal Sentinel-2 images to construct NDVI time series and time-texture two-dimensional representation maps as classification features. The NDVI time series captured the phenological changes of crop growth, while the time-texture two-dimensional representation maps captured the dynamic changes of spatial features over time. Then, convolutional neural network（CNN） + long short-term memory（LSTM） dual-stream architecture was used to combine temporal and spatial features to achieve accurate recognition of crops. 【Result】 The experimental results showed that compared with traditional methods that only used NDVI time series, the method incorporating texture time series greatly improved classification accuracy. The classification accuracy of the random forest increased from 0.89 to 0.93, and the classification accuracy of the support vector machine increased from 0.88 to 0.93. This indicated that the texture time series with spatial features effectively improved crop classification ability. The overall accuracy of using the CNN+LSTM dual-stream architecture classification model for land parcel crop classification reached 0.95, in particular, the classification accuracy of grape and winter wheat improved greatly, F₁ increased to 0.90 and 0.92 respectively. This demonstrated that, compared to traditional classifiers, the CNN + LSTM dual-stream architecture achieved more accurate land parcel crop recognition.【Suggestion】When conducting remote sensing classification of land parcels crops in areas with complex planting structures and similar crop growth habits, it is considered to incorporate texture time series features into the classification system and use a CNN + LSTM dual stream architecture to capture the temporal and spatial characteristics of crop growth separately. This method of integrating temporal and spatial information can improve the accuracy of crop classification on land parcels.