随着遥感技术的发展，遥感物候监测能够为农业生产者提供及时、准确的作物信息。Sentinel-2提供了全球范围的多光谱影像数据，并具有多个对植被敏感的波段，在植被监测中表现出巨大的潜力。本文以辽宁省平原区为研究区，以Sentinel-2为主要数据源，构建样本数据集。基于循环神经网络构建深度学习模型，综合利用多光谱时间序列提取玉米、水稻两种作物的两个关键物候期。将循环神经网络引入遥感物候监测，提出一个新的遥感物候监测方法。探究了数据时间密度和多光谱选择对模型结果的影响。本研究的主要工作和成果如下：

1、构建了研究区中玉米和水稻作物的时间序列样本数据集。共选取了130个样本点（玉米62个，水稻68个），采用曲线拐点法计算样本点的两个关键物候期SOS（D1）和EOS（D4）并标注生成样本的物候类型时间序列。经对样本点的9个波段反射率时间序列曲线进行重构、去除异常值、填充缺失值、筛选等处理，共构建了111个样本136个时间序列的样本集。玉米样本点SOS主要在第160天至180天，EOS集中在第260天至280天，大部分水稻样本的SOS集中在160天至170天，EOS集中在280天至300天。

2、分别基于LSTM、CNN-LSTM构建深度学习神经网络提取农作物物候期，主要工作包括：（1）构造模型的网络结构：基于LSTM构建的模型为三层结构，第一层为LSTM层，第二层为全连接层，最后经过SoftMax层输出模型结果。基于CNN-LSTM构建的模型在基于LSTM构建的模型的前端添加1-D CNN结构；（2）开展了对不同超参数组合方案的模型比较及模型在各个时间点的物候类型分类准确度评价；（3）运用最优参数模型对研究区Ⅱ进行物候监测。（4）可视化LSTM的输出层跟踪模型监测物候的过程。实验结果表明，两个模型均能够综合利用Sentinel-2的9个波段反射率进行物候类型识别，对物候类型分类的准确度随时间而增加，在物候转变期间模型的分类准确度下降。基于LSTM构建的模型在测试数据集的RMSE为12.66天，在训练数据集的RMSE为10.9天。基于CNN-LSTM构建的模型在测试数据集中的RMSE为10.26天，在训练数据集中的RMSE为10.8天。添加1-D CNN结构的模型物候结果偏差更小。LSTM对多光谱提取的特征点在时间上具有连续性，同种物候类型的特征点具有聚集性。

3、研究进一步测试了两个模型在不同时间密度下的模型表现，探究时间密度对模型结果的影响。对原有测试数据集的反射率时间序列进行等距抽取，分别保留为原数据集时间长度的1/2、1/4、1/6、1/8，使用两个模型提取新数据集的物候日期，并使用均方根误差评价模型表现。结果表明：LSTM模型在时间密度减少的情况下对物候类型分类准确度降低，但仍能对玉米、水稻三种不同物候类型进行区分。在时间密度减少为原来的1/2、1/4、1/6、1/8时，模型提取的物候日期延后，RMSE分别增加0.99天、1.34天、1.78天和4.73天。CNN-LSTM模型在减少时间密度的情况下，难以区分不同物候类型，对时间密度的包容性较差。

本文还测试模型在不同波段数量下的表现，探究光谱对模型结果的影响。使用2个波段数据训练基于LSTM深度学习模型，并使用2个波段的时间序列提取物候。实验结果表明： LSTM模型使用9个波段提取物候日期的RMSE为12.66天，使用2个波段提取物候日期的偏差为15.49天。使用多波段能够更好反映植被生长，提高对物候提取结果的准确度。

With the advancement of remote sensing technology, crop monitoring by remote sensing now provides timely and accurate crops information services to agricultural producers. Sentinel-2 provides multispectral image data on a global scale, with several vegetation-sensitive bands, and has great potential for monitoring vegetation. In this paper, the plain region of Liaoning Province serves as the study area, and Sentinel-2 is used as the primary data source to compile an example dataset. Using a combination of multispectral time series, a deep learning model based on a recurrent neural network is developed to extract two important phenological periods of two crops, maize and rice. Adopting recurrent neural networks for remote sensing phenology monitoring enables the development of a novel method for remote sensing phenology monitoring. The effects of data time density and multispectral selection on model outcomes were examined. The primary contributions and conclusions from this study are as follows:

1.Construction of time-series sample datasets for maize and rice crops in the study area. A total of 130 sample points were selected, including 62 for maize and 68 for rice. Using the curve fitting method, the two key phenological periods SOS (D1) and EOS (D4) of the sample points were calculated and labeled to construct the phenological type time series of the samples. After reconstructing, removing outliers, filling in missing values, and filtering the 9-band reflectance time series curves of the sample points, a total of 136 time series consisting of 111 samples were constructed. The SOS of maize sample locations was primarily at day 160 to 180 and the EOS was primarily at day 260 to 280, whereas the SOS of rice sample locations was primarily at day 160 to 170 and the EOS was primarily at day 280 to 300.

2.To extract crop phenological periods, respectively, deep learning neural networks based on LSTM and CNN-LSTM are constructed, with the primary work consisting of: 1) Constructing the network structure of the model: the model constructed based on LSTM has a three-layer structure, with the LSTM layer as the first layer, the fully connected layer as the second layer, and the SoftMax layer as the final layer through which the model results are output. The CNN-LSTM model augments the front side of the LSTM model with a 1-D CNN structure.2) Model comparisons of various hyperparameter combinations and classification accuracy assessments at each time point were conducted. 3) Monitoring the phenology of study area II using the optimal parameter model. 4)Visualization of the LSTM output layer to monitor the model's phenology monitoring process. The experimental results demonstrate that both models are capable of synthesizing the 9-band reflectance of Sentinel-2 for the identification of phenology types, and that the classification accuracy of phenology types increases with time, while the classification accuracy of the model decreases during the phenological transition. The RMSE of the LSTM-based model in the test dataset is 12.66 days and in the training dataset it is 10.9 days. The RMSE of the CNN-LSTM-based model in the test dataset was 10.26 days and in the training dataset it was 10.8 days. With the addition of 1-D CNN structure, the model's phenological results are less biased. The LSTM has temporal continuity in the extracted feature points from multiple spectra and aggregates feature points with the same phenological type.

3.To investigate the effect of time density on model results, the study compared the performance of the two models at various time densities. The reflectance time series of the original test dataset were extracted at equal intervals and kept as 1/2, 1/4, 1/6, and 1/8 of the original dataset's time length, respectively. The phenological dates of the new dataset were extracted using both models, and the performance of the models was evaluated using root mean square error. The results demonstrated that the LSTM model was less accurate at classifying phenological types with reduced temporal density, but it was still capable of differentiating between three distinct phenological types of maize and rice. The RMSE increased by 0.99, 1.34, 1.73, and 4.73 days, respectively, when the time density was reduced to 1/2, 1/4, 1/6, and 1/8 of the original value. The CNN-LSTM model struggled to differentiate between phenology types and was less inclusive of time density as time density decreased.

This paper also evaluates the performance of the model in different band numbers and investigates the effect of spectrum on model outcomes. Two bands of data are used to train the LSTM-based deep learning model, and time series extracts from two bands are used for waiting. The experimental findings demonstrate: The RMSE of the LSTM model using nine bands to extract events was 12.66 days, while the bias of using two bands to extract events was 15.49 days. Multiple bands can more accurately reflect the growth of vegetation and increase the precision of phenological extraction results.