Abstract:
Abstract: With the development of hyperspectral technology, it is of great significance to establish a specific compression and reconstruction method that can be used for conducting quantitative remote sensing of vegetation. Such a method is expected not only to improve the efficiency of data storage and transmission, but also to maintain the main spectral characteristics in interpreting some physiological and biochemical parameters. In this study, the compressive sensing technique was introduced to the compression and reconstruction of plant spectrum. Three critical physiological and biochemical parameters of plant, i.e. water content, carotenoid content and chlorophyll content, were chosen to test the retrieving efficacy of the proposed method. To facilitate such an analysis, a spectral dataset consisting of 2 500 spectra and corresponding plant physiological and biochemical parameters with multivariate normal distribution was generated by a classic plant leaf radiation model PROSEPCT. Based on the data, the compression and reconstruction method that was specific for vegetation spectral processing was proposed, described and evaluated. To process the spectral dataset using the compressed sensing method, the spectral dataset was firstly sampled by means of constructing random matrix. Then, the sampled data were reconstructed by a classic orthogonal matching pursuit algorithm, and the normalized root mean square error between the reconstructed data and the original data was analyzed. The performance of the method was thoroughly evaluated on 3 different levels: the spectral level, the feature level and the model level. At the spectral level, an error analysis was performed by directly calculating the original spectra and reconstructed spectra of corresponding samples. At the spectral index level, the spectral index was calculated based on both the original spectra and the reconstructed spectra. Then the error of the spectral index was analyzed. At the model level, the retrieving models for water content, carotenoid content and chlorophyll content were calibrated and validated based on the spectral indices from both original data and reconstructed data. For each modeling process, the entire dataset was divided into calibration data and validation data with a ratio of 7:3. All the retrieving models were constructed using a partial least squares regression method. The normalized root mean square error was used to indicate the accuracy of the retrieving models. Over all the analysis, the compressed sensing method was processed with different sampling rates. And the influence of sampling rate was also investigated. The results showed that the required sampling rate was associated with the correlation among the bands. It was found that for the spectral data with more than 65% bands moderately correlated to each other (absolute value of correlation coefficient was greater than 0.8), the reconstruction error conformed a certain pattern in all 3 evaluation levels. For the spectral level, the reconstruction error achieved lower than 2% when the sampling rate was higher than 0.25. For the feature level, the spectral index exhibited different sensitivity to the sampling rate. To ensure that the reconstruction error was no more than 10%, the sampling rate should be higher than 0.25, 0.15 and 0.1 for water content, chlorophyll content and carotenoid content, respectively. In the validation of the retrieving model, when the sampling rate was less than 0.2, the retrieving error of water content and carotenoid content was significantly correlated with the sampling rate. When the sampling rate was greater than 0.2, the normalized root mean square error of the carotenoid and chlorophyll content maintained at 15.4% and 8.2%, whereas that of the water content dropped below 16.5%. Overall, based on the theory of compressed sensing, this work proposed a compression and reconstruction method that is specific for vegetation spectral data. The features in reducing the data volume of plant spectra and the capability in maintaining plant critical spectral characteristics make the method have great potential in supporting vegetation remote sensing.