Hyperspectral Characterization and Chlorophyll Content Inversion Study of Reclaimed Vegetation in Rare Earth Mines


 Taking the Lingbei rare earth mining area in Dingnan county of Jiangxi Province as the research object of the reclaimed vegetation, the original spectrum, derivative spectrum and the continuum removed spectrum of the reclaimed vegetation were detected. The spectral characteristics and variation regularity of the typical reclaimed vegetation were analyzed, the correlation between chlorophyll content and spectral characteristic index of reclaimed vegetation was analyzed, and the sensitive spectral parameters were extracted. Partial Least Squares Algorithm, Back Propagation Neural Network Algorithm and Sparse Autoencoder Network Algorithm were selected to construct the estimation model of chlorophyll content, and compare the accuracy. The results show that; The vegetation spectrum of rare earth mine reclamation has the spectral characteristics of higher reflectance in visible region, red shift of green peak and red valley, blue shift of “red edge”, with less spectral variation in bamboo willow; Variability in the sensitive spectral parameters extracted from different vegetation; Sparse Autoencoder network algorithm is the optimal estimation model (R2 value of three vegetation is 0.9117,0.7418 and 0.9815 respectively). In the case of the small sample, it has higher estimation precision and universality for different reclaimed vegetation.


29
Rare earth has played an irreplaceable role in many cutting-edge military and high-tech 30 fields, including precision-guided weapons and aerospace. However, for a long time, the 31 disorderly and crude mining pattern of rare earth has left large areas of abandoned mining 32 sites (Chen et al. 2007). The initial solution leaching method of acid solution injected into ore 33 body is adopted in the rare earth mine, which leads to the soil degradation in the mine area, Hyperspectral remote sensing has the advantages of high spectral resolution and 44 real-time, rapid and non-destructive monitoring of chlorophyll content. Therefore, many 45 scholars use bloom spectrum to study vegetation growth monitoring, physiological and 46 biochemical parameters inversion. Under the stress of external factors, the normal growth 47 state of the vegetation is inhibited, and its physiological and biochemical parameters will 48 change, resulting in the variation of its spectral characteristics, which in turn affects the 49 accuracy of the inversion of chlorophyll content (Gong et al. 2014). Zhu found that the 50 spectral differences between copper-stressed leaves and healthy leaves were irregular. Not 51 only related to the growth period, but also related to plant species (Hengkai L et al. 2020). Lu 52 proposed that vegetation under heavy metal stress exhibits changes in the internal structure 53 and spectral characteristics of its leaves (Huang et al.2004). Wang analyzed the relationship 54 between chlorophyll values and spectral characteristics in different stress regions of typical 55 plants, and the variation regularity was different (Hao X et al. 2015). In the process of 56 chlorophyll content retrieval, it was found that there were significant differences in retrieval 57 methods for different stresses, which was attributed to the differences in the study subjects. 58 For example, Zheng used the continuum elimination method combined with partial least 59 squares as the optimal model for chlorophyll estimation in rapeseed leaves (Hengkai L et al.  The growth environment of ion-type rare earth deposits is special, to explore the spectral 69 characteristics of the reclaimed vegetation under stress, and according to the spectral 70 characteristics of the reclaimed vegetation, targeted selection of estimation models to achieve 71 the optimal prediction of vegetation physiological parameters (Lu et al. 2020). Based on the 72 hyperspectral data of typical reclaimed vegetation and the chlorophyll content obtained from 73 the ion-type rare earth mineral. In order to find out the best estimation parameters and 74 methods, a model of chlorophyll content estimation was established by using partial least 75 squares, BP neural network and SAE Neural Network Algorithm, it provides theoretical basis 76 and technical support for monitoring the growth status of reclaimed vegetation. The Lingbei rare earth mining area, covering an area of about 213km 2 , is located in the 80 northern part of Dingnan County, Ganzhou province. Long time the rare earth mining, 81 especially the use of early pool leaching/heap leaching process, accumulated a large amount 82 of waste rock and tailings. The leaching liquid produced by "in-situ leaching" destroyed the 83 original soil structure and nutrient distribution, resulting in vegetation degradation, soil 84 erosion and serious damage to the ecological environment (Li et al. 2020). In this study, the 85 Aobeitang ore spot in the mining area is taken as the research area. The ore spot is located in 86 the south of Lingbei rare earth mine. The specific coordinates are 115°04'37″~115°05'41″ E, 87 24°54'10″~24°56'42″ N, and the area is about 1.52 km 2 . The ore site has undergone three 88 mining processes: pool leaching, heap leaching and in-situ leaching, leaving a large area of 89 exposed tailings. Since water pollution, vegetation growth is not good. In order to study the 90 change of spectral characteristics of reclaimed vegetation under the influence of 91 environmental stress factors, Tung oil tree, bamboo willow and photinia glabra were selected 92 as the research objects.

Data acquisition 96
The hyperspectral data in this study were collected from the field. The spectral 97 reflectance of the leaves was recorded by ASD Spectral Devices (350-2500 nm). The spectral 98 resolution was 3 nm at 350-1000 nm, the sampling interval was 1.4 nm; 1000-2500 nm, the 99 spectral resolution was 10 nm, and the sampling interval was 2 nm. The data was collected in 100 clear, cloudless and windless weather and at noon (11:00~14:00) instrument optimization and 101 Whiteboard calibration measurements are performed prior to recording actual observations. 102 In order to ensure the accuracy of the determination results as much as possible, the whole 103 determination process should meet the requirements of the specification (Mahajan et al. 104

2017). 105
During the collection, the survey team selected tung oil tree, bamboo willow and 106 Photinia glabra as the acquisition objects. In addition, in order to obtain the difference 107 between the reclaimed vegetation spectrum and the normal vegetation spectrum, the normal 108 tung oil tree, bamboo willow and photinia glabra were collected within 1km from the mine 109 site. According to the principle of uniform distribution, 110 groups of samples were randomly 110 measured for each vegetation in the study area; To ensure the accuracy of the measurement 111 data, each group of samples was measured 10-15 times, and the average value was calculated 112 as the final reflection spectrum of the sample point. 113 The SPAD-502 chlorophyll meter was used to collect the SPAD values immediately 114 after the single group of sample points were measured. In order to improve the accuracy of 115 measurement data, the position of leaf vein should be avoided. The average value was used as 116 the chlorophyll content of the sample. 117

Data processing 118
Hyperspectral data collected in the field are easily affected by soil environment and 119 human operation, so it is necessary to remove invalid data with large deviation or abnormal 120 fluctuation (Yan et al. 2016). In addition, the spectral data processing software ViewSpec Pro 121 is used to de-noise and smooth the spectral curve. In the process of spectral analysis, the noise 122 (1) 128 In the formula, is the wavelength of each band; ( ) is the original spectral 129 reflectivity of waveband i; ′ ( ) is the first derivative spectral value of wavelength ； 130 is the interval from wavelength +1 to , determined by the spectral sampling interval 131 (Hengkai L et al. 2020). 132

= ⁄
(2) 133 In the formula, and correspond to the envelope and shell values at band i; 136 , and correspond to the reflectivity at the beginning, the end and the i of absorption; 137

And
, and correspond to the wavelengths at the beginning, the end and the i of 138

absorption. 139
The existing vegetation index cannot get excellent performance when it is directly 140 applied to the reclaimed vegetation in rare earth mining area. In this study, vegetation index 141 NDVI and RVI were used as the prototype, and the optimal band combination of vegetation In the formula, , , and are reflectivity at any band, where i≠ j, m > n. New 148 spectral parameters extraction based on first derivative and the continuum removed transform 149 spectrum. The parameters are as follows: 150

Chlorophyll estimation method 153
In this study, hyperspectral data of leaves were used as independent variables, and The decision Coefficient R 2 , root mean square error (RMSE) and relative error RE were 184 used to evaluate the prediction ability of the model. The larger the decision Coefficient R 2 (no 185 more than 1), the smaller the root mean square error (RMSE) and relative error RE, the 186 stronger the ability of model fitting and prediction. The formula is as follows: 187 In the formula, and ′ is the measured value and the predicted value respectively; 191 and n is the number of samples.

Spectral characteristic analysis 194
Vegetation growth is usually affected by the environment, and the degree of impact can 195 be shown by spectral changes. Generally speaking, when the growth of vegetation is under 196 some kind of stress, the content of chlorophyll will decrease, which will eventually lead to the 197 decline of photosynthetic capacity, the decrease of spectral reflectance, the "red shift" of 198 wave peak and trough, and the "blue shift" of the "red edge" position of the first derivative 199 spectrum. Because of the ecological stress, the vegetation growth in the reclaimed land of rare 200 earth mine is consistent with that in other mining areas.

247
It can be seen from Figure.  In order to further analyze the difference of derivative spectra between the same type of 257 reclaimed vegetation and normal vegetation and different types of vegetation, the statistical 258 results of "trilateral parameters" in Table 3 were used for comparison. The blue shift of "red 259 edge" was found in all the red edge parameters of the three vegetation types, but the changes 260 of the three vegetation types were not consistent. The "red edge" of the reclaimed tung oil 261 tree, bamboo willow and photinia glabra were 697 nm, 719 nm and 696 nm, respectively. and at the λ r site decreases by 22 nm. Among them, the difference of " Trilateral parameters" of bamboo willow is smaller than that of tung oil tree and photinia glabra. And 272 the adaptability of reclaimed bamboo willow in the mining area is stronger. 273

Analysis of spectral characteristics based on the continuum removed spectra 274
In this study, the original spectra of tung oil tree, bamboo willow and photinia glabra 275 were continuum removed ( Figure. 4). At the same time, the vegetation spectral information is 276 analyzed with the help of the new absorption characteristic parameters included in the 277 continuum removed spectrum, as shown in Table 4.   vegetation is located in the red valley λ P will shift to the long wave direction, and the 289 absorption capacity will decrease, resulting in R P reflectivity increases. In addition, the "red 290 shift" phenomenon causes the red valley wavelength to become larger, and the value of S for 291 reclaimed vegetation is smaller than that of normal vegetation, indicating that the absorption 292 in rare earth mining areas, separate spectral parameter extraction is required. In this study, 304 based on the original spectrum, derivative spectrum, the continuum removed spectrum, the 305 spectrum curve, characteristic parameters and vegetation index, the sensitive spectrum 306 parameters for reclaimed vegetation were extracted. 307 Based on the original spectra of the three reclaimed vegetations, the correlation between 308 the entire band range of the first derivative spectra and the chlorophyll content is calculated. Compared with the original spectrum, the correlation between the first derivative 317 spectrum and the chlorophyll content is improved obviously, and the range and correlation of 318 the band are better than that of the original spectrum. In conclusion, the correlation between 319 the first derivative spectrum and chlorophyll content is higher than that of the original 320 spectrum in the sensitive band, which shows that the first derivative can eliminate the 321 influence of soil background to a certain extent. However, the stability of the first derivative 322 Spectra at some wavebands is obviously worse than that of the original spectra.

329
In this study, the highly significant correlation Coefficient (p < 0.01) was used as the 330 sensitive spectral parameter extraction standard, and the relatively large band of absolute 331 correlation Coefficient was selected as the candidate spectral parameter, the Derivative 332 Spectra have more spectral parameters than the original spectra. 333

337
Based on the "trilateral" parameter of the derivative spectrum and the absorption 338 parameter of the continuum removed, the correlation analysis with the chlorophyll content is 339 carried out, and the parameters that reach the extremely significant correlation level of P<0.01 340 are selected as candidate spectral parameters for the chlorophyll content estimation model 341 construct. 342 Table 6 343 Extraction statistics of the "trilateral" parameters and absorption parameters of tung oil tree, bamboo 344 willow and photinia glabra.

346
In the "trilateral" parameters, the absolute value of the correlation coefficient of Tung oil 347 above 0.9 is the blue edge position λ b , the red edge position λ r , the red edge area to blue edge 348 area ratio SD r / SD b , the normalized red edge area and blue edge area. The edge area ratio( 349 SD r -SD b )/ (SD r +SD b ) indicates that it is easier to extract the sensitive spectral parameters 350 from the red and blue spectral angles. Among the four "trilateral" parameters extracted from   Table 7 369 Extraction of vegetation index of tung oil tree, bamboo willow and photinia glabra.

428
In order to further compare the estimation models of the reclamation vegetation 429 chlorophyll content established by the three algorithms, statistics of the determination 430 coefficient R 2 , root mean square error RM, and relative error RE of various vegetation 431 validation set model evaluation indicators are calculated. The results are shown in Table 6. 432 On the whole, in the three chlorophyll content estimation models established by various 433 reclaimed vegetation, the determination coefficient R 2 is gradually increased by the PLS 434 algorithm, BP neural network algorithm, and SAE neural network algorithm. Although the 435 SAE neural network algorithm of Photinia glabra is slightly smaller than the BP neural 436 network algorithm, the R 2 values of both are above 0.98; and the root mean square error RM 437 and the relative error RE are determined by the PLS algorithm, BP neural network algorithm, 438 and the SAE neural network algorithm gradually decreases, which shows that the estimation 439 effect based on SAE neural network is the best, which is higher than the other two estimation 440 methods. The reason is that the relationship between vegetation chlorophyll and spectral 441 characteristics is not fixed, but a nonlinear relationship exists. The non-linear 442 problem-solving ability of the SAE algorithm makes it superior to the linear analysis method 443 PLS in chlorophyll estimation; SAE method effectively solves the local optimization problem 444 in the network through the layer-by-layer pre-training and fine-tuning process, and 445 accelerates the network convergence speed. In contrast, the BP neural network method is 446 better than the PLS method in estimation accuracy because of its non-linear problem-solving 447 ability, but its own existence is easy to fall into local problems, making its estimation 448 accuracy inferior to the SAE neural network. On the whole, in the estimation methods of the 449 rare earth mines reclamation vegetation, the SAE neural network estimation method is the 450 best and can be applied to the reclamation vegetation; the BP neural network estimation 451 method of photinia glabra is better than the SAE, the difference is very small, only 0.003 452 difference. The precision of bamboo willow is the lowest among all the estimation methods.

499
In this study, the spectrum analysis of three types of reclaimed vegetation in rare earth 500 mining areas, Tung Oil Tree, Bamboo Willow, and Photinia Glabra, combined with a variety 501 of spectral processing methods, found that there are serious spectral variations between the 502 reclaimed vegetation and the normal vegetation. Spectral characteristics such as increased 503 reflectance of the area, red shift of the green peak and red valley position, and blue shift of 504 the "red edge" position. The degree of variation of different vegetations is inconsistent, 505 indicating that there are certain differences in adaptability to the environment, among which 506 bamboo willow are better adapted to the mining environment. After calculating the 507 correlation between vegetation spectrum and chlorophyll content and extracting sensitive 508 spectrum parameters, it is found that the sensitive parameters extracted by different 509 vegetations have similarities and differences. The sensitive parameters belong to the same 510 range, but the specific parameters are not consistent. Its unique sensitive spectral parameters 511 are attributed to the spectral difference caused by rare earth minerals. 512 Finally, an estimation model of the optimal chlorophyll content of each reclaimed 513 vegetation was constructed. The SAE neural network estimation model can be applied to the 514 reclaimed vegetation of rare earth mines, but Photinia glabra can also achieve the same 515 accuracy through the BP neural network. Studies have shown that traditional linear statistical 516 models are not effective when compared with nonlinear models such as neural networks. 517 Inversion models based on neural networks and deep learning will become the trend of future 518 hyperspectral development, which can provide rapid growth of vegetation in large-scale 519 mining areas. Dynamic monitoring, UAV monitoring and hyperspectral remote sensing 520 applications lay a solid research foundation. 521 522 Declarration 523 Fundings 524 There are currently no Funding Sources in the list 525 Author information 526 School of Civil and Surveying&Mapping Engineering., Jiangxi University of Science and 527 Technology, No.86 Hongqi Road, Ganzhou, Jiangxi, 341000, China 528 Hengkai Li, Feng Xu, Beibei Zhou and Zhian Wei 529 Corresponding author 530 Correspondence to Hengkai Li. 531 Authors' contributions 532 Hengkai Li conceived the research and wrote most of the manuscript. 533 Feng Xu performed the statistical analysis of population and territory, performed the GIS analysis, 534 and produced the graphs and tables. 535 Beibei Zhou produced the maps and analysis and wrote the pertaining parts. 536 Zhian Wei checked and perfect the whole paper. All authors have read and approved the 537 manuscript. 538

Ethics declarations 539
Competing interests 540 The authors declare that they have no competing interests. 541 Ethics approval and consent to participate 542 Not applicable. 543 Consent for publication 544 Not applicable. 545 Availability of data and materials 546 The datasets used or analysed during the current study are available from the corresponding author 547 on reasonable request. 548 549