基于Sentinel-1A影像和一维CNN的中国南方生长季早期作物种类识别

作物的早期识别对粮食安全至关重要。在以往的研究中,中国南方作物早期识别面临的主要挑战包括:1)云层覆盖时间长、地块尺寸小且作物类型丰富;2)缺少高时空分辨率合成孔径雷达(synthetic aperture radar,SAR)数据。欧洲航天局Sentinel-1A(S1A)卫星提供的SAR图像具有12 d的重访周期,空间分辨率达10 m,为中国南方作物早期识别提供了新的机遇。为在作物早期识别中充分利用S1A影像的时间特征,本研究提出一维卷积神经网络(one-dimensional convolutional neural network,1D CNN)的增量训练方法:首先利用生长季内全时间序列数据来训练1D CNN的超参数,称为分类器;然后从生长季内第一次S1A影像获取开始,在每个数据获取时间点输入该点之前(包括该点)生长季内所有数据训练分类器在该点的其他参数。以中国湛江地区2017年生长季为研究实例,分别基于VV、VH和VH+VV,评估不同极化数据在该地区的作物分类效果。为验证该方法的有效性,本研究同时应用经典的随机森林(random forest,RF)模型对研究区进行试验。结果表明:1)基于VH+VV、VH和VV极化数据的分类精度依次降低,其中,基于VH+VV后向散射系数时间序列1D CNN和RF测试结果的Kappa系数最大值分别为0.924和0.916,说明S1A时间序列数据在该地区作物分类任务中有效;2)在研究区域内2017年生长季早期,基于1D CNN和RF的5种作物的F-measure均达到0.85及以上,说明本文所构建的1D CNN在该地区主要作物早期分类任务中有效。研究结果证明,针对中国南方作物早期分类,本研究提出的1D CNN训练方案可行。研究结果可为深度学习在作物早期分类任务中的应用提供参考。

Early growing stage crop species identification in southern China based on sentinel-1A time series imagery and one-dimensional CNN

Timely and accurate estimation of the area and distribution of crops is vital for food security. In previous research, the main challenges of early crop identification in southern China were: 1) cloudy days are frequent; 2) most parcels are small and crop types are variant; 3) high spatial and temporal resolution synthetic aperture radar (SAR) data is lack. The European Space Agency Sentinel-1A (S1A) satellite, which recently became operational, is a satellite system providing global coverage of SAR with 12-days revisit period at high spatial resolution at 10 m. This provides a new opportunity for early crop identification in southern China. Compared with classical machine learning methods, deep learning has many advantages, such as end-to-end training, mobility, it provides a new chance for using high spatio-temporal data efficiently. One-dimensional convolutional neural network (1D CNN) and recurrent neural network (RNN) have been shown to be effective deep learning methods for extracting temporal features for classification tasks. However, the parameters (but not hyper-parameters) of RNN are determined by the length of the time series. Compared with RNN, 1D CNN has the advantages of simple structure and few parameters, but very few studies have applied 1D CNN to time series data for early stage crop''s identification at present. In this study, we proposed to combine 1D CNN hyper-parameters training with an incremental time series training method to attain a classification model for each time point in the growth season. Firstly, we trained the 1D CNN hyper-parameters using the full time series data during the growth season, and refer to the 1D CNN with hyper-parameters as a classifier. Then, starting at the first time point, the incremental training process was carried out, at the acquisition time point of each S1A image, input all data in the growing season before (including) the point to train other parameters of the classifier at that point, and then obtained a classification model with all parameter values (including the previous hyper-parameters) at each time point. Finally, test accuracies of each time point were assessed for each crop type to determine the optimal time series length. A case study was conducted in Zhanjiang City, China. To verify the effectiveness of this method, a comparative experiment with the classical random forest (RF) method was carried out. In order to evaluate different polarizations mode (VV, VH, VH+VV) of S1A data for crop classification in the study area, we performed the above training process for each polarization time series data. The results were as follows: 1) the classification accuracies of VH+VV, VH, and VV decrease in order, based on the VH+VV backscatter coefficient time series, the maximum Kappa coefficient values of 1D CNN and RF model were 0.924 and 0.916 respectively, illustrating that S1A time series data was valid for crop classification task in the study area; 2) in the early growing season of 2017 of the study area, F-measures of 1D CNN and RF model were above 0.85, which indicated that 1D CNN in this work was effective for early crop classification. All results indicated that the proposed training method of 1D CNN was valid for early stage crop''s classification. At the same time, 33 optical images of sentinel-2 in the study area of 2017 growth season were downloaded, of which only 6 were not hindered by clouds. Therefore, S1A SAR with 12 days revisit period can not only obtain data under any weather conditions, but also track crop growth accurately. This method provides a new perspective for the application of deep learning in early stage crop''s classification tasks. In addition, all parameters of 1D CNN can be trained by using time series data from different years. Although the performances of 1D CNN almost similar to those of the RF, deep learning models have advantages that other methods do not have. Therefore, we believe that deep learning methods will play an important role in early crop identification in the near future.