Estimating chlorophyll content of apple leaves based on preprocessing of reflectance spectra
Deng Xiaolei,Li Minzan,Zheng Lihua,Zhang Yao,Sun Hong
DOI: https://doi.org/10.3969/j.issn.1002-6819.2014.14.018
2014-01-01
Abstract:Great progress has been achieved in the prediction of vegetation biological parameters based on spectroscopy, and most studies were focused on building models by the mathematical combination of the reflectance, the red edge and the blue edge, and different modeling methods, to improve the accuracy of the models. Different combinations of preprocessing methods can get different accuracies. However, it is not inevitable that all preprocessing methods can help to improve the accuracy, and the same combination of the methods for the same data with different steps may get different accuracies. Thus, in this paper, the impacts of the preprocessing methods and steps on the spectral feature extraction and the models are discussed. Derivative spectra can eliminate the effect of baseline drift, reduce background interference, and provide higher spectral resolution than the original spectra. Wavelet packet transform can decompose the low-frequency and high-frequency parts of the signal and thus, show obvious advantages in signal denoising. Therefore, these two preprocessing methods and the combinations with different steps were studied. Taking apple leaf chlorophyll content as the research object, spectral autocorrelation coefficients, correlation coefficients between spectral data and the chlorophyll content, and stepwise regression modeling were calculated for the reflectance spectra, including the original reflectance spectra, wavelet packet denoising reflectance spectra, first-order differential reflectance spectra, the first-order differential of the wavelet packet denoising reflectance spectra, and wavelet packet denoising of the first-order differential reflectance spectra. The 60 apple leaf samples were collected from the top, middle, and bottom positions of sunny main branches from 20 apple trees, and the reflectance and the chlorophyll content were then measured. The spectral data of the 60 apple leaf samples, ranging from 300 to 850 nm by different preprocessing methods, were formed into matrices (60×551), and the spectral autocorrelation coefficients were then calculated. The effects of the denoising methods were evaluated by peak signal-to-noise ratio (PSNR), lower mean square error (MSE) and maximum squared error (MAXERR). At the same time, the accuracies of the predicted models were evaluated by the r and root mean square error (RMSE). The spectral curve can be smoothed by the 3-layer sym8 wavelet packet de-noising, but the modeling accuracy was not improved. Therefore, it was not reliable in evaluating the effect of the denoising methods only by the naked eye. It was important to choose the proper parameters for wavelet packet denoising. Although the noise was amplified by the first-order differential, the baseline drift was removed and thus, the accuracy of the model was improved. The wavelet packet denoising of the first-order differential reflectance spectra had higher PSNR, lower MSE and MAXERR than the first-order differential of the wavelet packet denoising reflectance spectra. The r and RMSE of the regression models for these two methods were 0.746, 5.01 and 0.683, 5.44, respectively. The former method had higher r and lower RMSE. Therefore, the denoising of the first differential reflectance spectra had a better denoising effect and linear regression model accuracy than the first differential of the denoising reflectance spectra. Thus, wavelet packet denoising of the first-order differential reflectance spectra could be considered as an effective preprocessing method to improve modeling accuracy. The study can satisfy the demands of evaluating the nutritional status of apple tree and precision fertilization.