Abstract: At the present, the crop classification method based on time series feature extraction needs much prior knowledge and manual intervention. It is difficult to automate the classification and it is easy to reduce the accuracy due to neglecting some effective features. We propose a convolutional neural network crop classification method based on time series spectral reconstruction. The method constructs a time-series image for every ground pixel, with the time dimension as the vertical axis and the spectral dimension as the horizontal axis, and then uses CNN, optimized by using the Adam gradient descent method and 40% connection rate dropout, to classify the time-series images. The experimental results show that the method reduces the salt-and-pepper noise, and the boundary of the plot is clear. The overall classification accuracy reaches 95.12%, which is higher than those of the time series multispectral + random forest method(88.58%), time-series NDVI + random forest method(90.25%), and time-series NDVI + convolutional neural network method(91.79%). For spring corn and tomato with close spectral similarity, the F₁-scores of our method reach 95.9% and 89.9% respectively, which are significantly improved compared with the control groups. The results of this study provide useful information for the automatic and fine mapping of croplands.

Key words: crop classification, convolutional neural network, remote sensing, Sentinel-2A, time series, feature extraction