基于Bi-LSTM模型的时间序列遥感作物分类研究

【目的】及时、准确地作物分类制图是农情监测的重要依据。本研究基于双向长短期记忆网络模型探究深度学习技术在时间序列遥感作物分类与早期识别中的应用潜力。【方法】本文以黄河三角洲地区为例，以哨兵2号全年可用卫星影像为数据源，构建年时间序列NDVI数据集;采用循环神经网络构架，搭建针对结构化时序数据的双向长短期记忆网络模型（bidirectional long short-term memory，Bi-LSTM），开展遥感作物分类，并评估模型的泛化能力;通过输入不同长度时间序列遥感数据，探究满足一定制图精度条件下的作物最早可识别时间。【结果】作物年生长时序特征对于大多数作物遥感分类识别都具有较好的区分能力，基于年时间序列NDVI数据的Bi-LSTM模型作物分类总体准确率达90.9%，Kappa系数达到0.892。通过测试不同时间序列长度对作物分类的影响发现，对大多数作物来说，其分类精度随着数据时间序列长度增加而不断提高，冬小麦、水稻等作物在生长季早期即具有较为独特的分类特征，因而利用生长季早期的时间序列影像即可获得较高的制图精度，而棉花、春玉米等作物需要完整生长序列影像才能更好地保证分类精度。【结论】卫星影像时间序列蕴含的结构化特征信息可以有效地降低特定时段的作物光谱混淆;双向循环神经网络模型能够同时考虑前向和后向的时间状态信息，可以学习作物不同阶段的光谱变化特征，在水稻、棉花、春玉米等易混淆作物的识别上表现优异;模型能够有效地把握样本总体上的变化趋势，在农作物多分类任务中表现出较好的泛化能力和鲁棒性。本研究通过集成深度学习和遥感时间序列，为及时、快速的区域作物高精度制图提供了可行的思路。

Crop Classification with Time Series Remote Sensing Based on Bi-LSTM Model

【Objective】Timely and accurate crop classification mapping is an important basis for agricultural situation monitoring. This study explores the potential of deep learning in time series remote sensing crop classification and early identification based on a bidirectional long short-term memory network model.【Method】In this paper, Yellow River Delta region was chosen as an example and a time-series NDVI dataset were constructed by using Sentinel-2 year-round available satellite images as the data source. A recurrent neural network architecture is used to build a bidirectional long short-term memory (Bi-LSTM) model for structured time-series remote sensing data to carry out crop classification, then the generalization ability of the model is evaluated. Through adjusting the length of time series, we explore the earliest identifiable time of different crops under the condition of satisfying certain mapping accuracy. 【Result】 Growth characteristics represented by time series remote sensing images have great potential to discriminate different crops. The overall accuracy of the Bi-LSTM model reached 90.9% with a Kappa coefficient of 0.892. By testing the effects of different time series lengths on crop classification, the earliest identifiable time of typical crops was obtained. The accuracy of crops such as winter-wheat and rice could improve significantly after the emergence of unique characteristics. Crops such as cotton and spring maize required complete growth sequences to ensure classification accuracy.【Conclusion】The structured feature information embedded in satellite image time series could effectively reduce crop spectral confusion at specific time periods. The Bi-LSTM model was able to consider both forward and backward temporal state information and could learn the spectral change characteristics of crops, which was excellent in the identification of confusing crops such as rice, cotton and spring maize. In addition, the deep learning model could effectively capture the variation trend on the sample in general, and showed better generalization ability and robustness in the crop multi-classification task. This study provided a feasible idea for regional crop mapping with high accuracy by integrating deep learning and remote sensing time series.