基于多时序特征和卷积神经网络的农作物分类

近年来，以卷积神经网络为主的深度学习模型在各种遥感应用中都显示出巨大的潜力。以加州帝国郡为研究区，以Landsat 8 OLI年内时序遥感影像计算时序植被指数NDVI、EVI、RVI以及TVI，组合后输入到构建的一维卷积神经网络 模型，以实现作物的高精度精细分类。为了验证卷积模型的优越性，另搭建了基于递归神经网络及其变体的深度学习模型。结果表明：①引入其他时序特征后，能够有效地提高卷积神经网络的分类精度。NDVI+EVI+TVI+RVI组合特征总体精度和Kappa系数最高，分别是89.667 4%和0.856 0，对比NDVI时序特征总体精度和Kappa系数提高了近4%和0.6。②在与其他深度学习模型的对比中，一维卷积神经网络分类精度最高，能够从时序数据中较为准确捕捉作物时序特征信息，尽管递归神经网络被广泛应用于序列数据的研究，但分类结果要略差于卷积神经网络。实验表明在NDVI的基础上引入其他植被指数辅助，能够有效地提高分类精度。基于一维卷积神经网络的深度学习框架为长时间序列分类任务提供了一种有效且高效的方法。

Crop Classification based on Multi-temporal Features and Convolutional Neural Network

Recently， Convolutional Neural Network （CNN） shows great potential in various remote sensing applications Taking Imperial County of California as the study area， and calculating vegetation index NDVI， EVI， RVI and TVI form landsat-8 OLI time series remote sensing images. Then， input it into the constructed CNN model to achieve crop classification. In order to verify the superiority of the convolution model， a deep learning model based on recurrent neural networks and its variants was built. The results show that： ①Adding other time series features can effectively improve the classification accuracy of CNN. The overall accuracy and Kappa coefficient of NDVI+EVI+TVI+RVI combination features are best， respectively 89.6674% and 0.8560， which is nearly 4% and 0.6 higher than the single time series features. ②Convolutional neural networks have the highest classification accuracy in comparison with other deep learning models. It can capture crop timing feature information more accurately from time series data. Although RNN is widely used for sequential data representation， but the classification results are slightly worse than the convolutional neural network. Experiments show that the introduction of other vegetation index assistance on the basis of NDVI can effectively improve the classification accuracy. A deep learning framework based on 1D convolutional neural networks provides an effective and efficient method for Multi-Temporal classification tasks.