关键词: 叶绿素含量, 高光谱, 自动编码器, 卷积神经网络

Abstract: Chlorophyll content is an important indicator for evaluating plant growth condition and photosynthesis capacity. The study used a chlorophyll meter to measure the SPAD (soil plant analysis development) value that characterizes the chlorophyll content of lilac leaves and used hyperspectral image technology and the machine learning algorithm inversely calculated the chlorophyll content of lilac leaves. Aiming at the problems of noise information interference, strong correlation between adjacent bands, and redundant information during data collection, this paper took advantage of the AtrousConvolutional Denoising AutoEncoder (Atrous-CDAE) to decrease the original hyperspectral data from 204 dimensions to 51 dimensions, and lowered noise interference. Finally, a prediction model of the chlorophyll content of lilac leaves was established based on 1DCNN, and compared with the original data and four other data processing methods. The results indicated that compared with the original hyperspectral data and other data processing methods, the data after Atrous-CDAE showed the best prediction results:with a prediction concentration coefficient (R2) of 0.972 3 and a root mean square error (RMSE) of 1.326 6. The combination of the data processed by Atrous-CDAE and other classical prediction models had achieved better prediction results, indicating that Atrous-CDAE can effectively extract the potential representation of the data. Combining other data with the 1DCNN model proposed in this article, the R2 was above 0.94, and the RMSE was below 2, indicating that the prediction model had a certain degree of adaptability.

Key words: chlorophyll content, hyperspectral, auto encoder, convolutional neural network