Abstract: 【Objective】Timely and accurate crop classification mapping is an important basis for agricultural situation monitoring. With the significant increase in spatial and temporal resolution of satellite remote sensing data acquisition, crop classification and early identification based on intensive time series remote sensing data have become possible. Meanwhile, the rapid development of machine learning methods with deep learning has provided new technical support for crop remote sensing classification.【Method】In this paper, we take the Yellow River Delta region as an example and construct a time-series NDVI dataset using Sentinel-2 year-round available satellite images as the data source. Based on a recurrent neural network architecture, a Bidirectional Long Short-Term Memory (Bi-LSTM) network model is built to explore its potential application in crop classification using its feature learning capability for structured time series data. 【Result】Winter-wheat-summer-maize is the main crop cultivation pattern in the region, and the overall accuracy of the Bi-LSTM model reached 90.9% with a Kappa coefficient of 0.892. Crop accuracy improves with the increase of time series data length. The accuracy of crops such as winter-wheat and rice will improve significantly after the emergence of unique characteristics. Crops such as cotton and spring-maize require complete growth sequences to ensure classification accuracy. 【Conclusion】Structured feature information embedded in satellite image time series can effectively reduce crop spectral confusion at specific time periods. The Bi-LSTM model is able to consider both forward and backward temporal state information and can learn the spectral change characteristics of crops, which is excellent in the identification of confusing crops such as rice, cotton and spring corn. By testing the effect of different time series length on crop classification, the earliest identifiable time of typical crops was obtained for early mapping of crops under certain mapping accuracy. In addition, the deep learning model can effectively capture the variation trend on the sample in general, and shows better generalization ability and robustness in the crop multi-classification task. This study provides a feasible idea for regional crop mapping with high accuracy by integrating deep learning and remote sensing time series.

Key words: crop classification, early identification, time-series remote sensing, Bi-LSTM model, model generalization