Super-resolution enhancement of Sentinel-2 image for retrieving LAI and chlorophyll content of summer corn

Sentinel-2 satellite is a new generation of multi-spectral remote sensing technique with high spatial, temporal and spectral resolution. Especially, Sentinel-2 incorporates three red-edge bands with central wavelength at 705, 740 and 783 nm, which are very sensitive to vegetation changing, heath and variations. Unfortunately, their spatial resolution is only 20 m, which is lower comparably. This spatial resolution brings difficulties for mining the potential of Sentinel-2 image in vegetation monitoring. Therefore, we focus on enhancing the spatial resolution of Sentinel-2 red edge band images to 10m using the SupReME algorithm. Furthermore, the summer corn canopy leaf area index (LAI), leaves chlorophyll content (LCC) and canopy chlorophyll content (CCC) were retrieved by the linear and physical models for the corn growth monitoring purpose. The results showed that the spatial resolution of Sentinel-2 images had been enhanced to 10m from original 20m, and the estimation accuracy (EA) was over 97% for pixels planted by summer corn. Moreover, the accuracy of summer corn canopy LAI, LCC and CCC was improved respectively using enhanced Sentinel-2 images by SupReME method. During these three parameters retrieval, the red-edge bands or SWIR bands were introduced into optimal cost function and vegetation index which the accuracy of these models was high. The SupReME algorithm provides a valuable way for Sentinel-2 images enhancement, which is of great potential to mining Sentinel-2 images and multitude


33
Accurate estimation of vegetation biophysical variables with high spatial and temporal

55
However, the availability of 10m images provides an opportunity to improve the resolution of other 56 20m images, leading to an effective means to maximize output without increasing input costs.

101
Therefore, this study is focusing on enhancing Sentinel-2 multi-spectral images with different 102 spatial resolutions to generate high spatial resolution multi-spectral images using the SupReMe 103 algorithm, aiming at improving the retrieving accuracy of summer corn canopy LAI, LCC and CCC.

104
For this purpose, we have formulated following study contents:    154 was used to measure the chlorophyll content of corn leaves nondestructively (Campbell et al., 1990) 155 on the position A and B. When measuring chlorophyll content, two or three corn plants were

167
The measured LAI, ALIA and LCC for 113 quadrats were analyzed in this sutdy. It was found 168 that these growth parameters of summer corn in study area were approximately Gaussian 169 distributed (as shown in Fig.3), which also provided a priori knowledge for parameter input and 170 sensitivity analysis of the later physical model.

184
In the continuous spectral curve, the two adjacent bands have strong correlation. Therefore,

191
(ii) Blur each band, such that the blur of all the bands is equivalent to the strongest blur.

192
(iii) Perform singular value decomposition on the blurred data. Retain the singular vectors of 193 largest singular values as the columns of .

194
Then image fusion can be formulated as a convex optimization problem: where is a quadratic regularisation term, based on weights and is the regularization

225
In view of the limitation of spatial and spectral resolution of early satellite, the vegetation index The pixel size of enhanced image is 10m×10m, and the reflectance of a single pixel is a mixed 235 spectrum of the whole vegetation canopy. We inverted LAI and CCC directly from canopy spectra, 236 and then calculated LCC by the ratio of CCC to LAI.

281
Generally, we take the 10% fluctuation of the results from the optimal linear model as its limit.

282
Furthermore, the band combination used for cost function is an important factor affecting the 283 retrieval accuracy. The number of combinations M can be expressed as:

284
(16)  In order to evaluate the enhanced reflectance results from the developed SupReME algorithm,

315
From the analyzed spectral correlation in Table 3, we found that the reflectance of enhanced 316 image was highly correlated with the original image. And correlation coefficient (R 2 ) between the 317 original bands and enhanced bands are all greater than 0.87. The band with the highest correlation 318 value is Band12 with R 2 =0.99, and the band with the lowest correlation is Band11 with R 2 = 0.87. The

319
EA of the whole image is higher than 79%, including corn planted area, roads, water etc.

320
Fortunately, the EA in corn planted area is higher than 97% with the significant value at the 0.01 321 level. The correlation results in Table 3 revealed that the enhanced image using SupReME algorithm 322 improves the spatial details of the image while maintaining the invariance of the spectrum. We 323 found that the EA in the vegetation pixels of SupReME algorithm is higher than that of buildings, 324 especially for Band11 and Band12.

327
We developed RI and NDI by combing every two bands of the Sentinel-2 images and found the 328 optimal vegetation index (Table 4)

344
The other three VIs used at least three bands in the process of construction, in which at least 345 one band was red edge band with an original spatial resolution of 20 m. In conclusion, the 346 estimated results using MCARI, MTCI and AIVI showed that the improvement of Sentinel-2 spatial 347 resolution using the SupReME algorithm can improve the estimation accuracy of LAI and CCC.

402
showed that the SupReME algorithm can be used to enhance the spatial resolution of remote sensing 403 image, in the meantime of maintaining the spatial details and spectral variation of original image 404 (Fig. 5).

405
In the comparison of super-resolution enhanced images (10m) and original low spatial 406 resolution images (20m), we found that the R 2 of the two kind images were higher than 0.87 of six 407 bands, and the EA were higher than 79%. Note that the EA of vegetation pixels with reflectivity less 408 than 0.6 were higher than 97% (Table 3). On the one hand, the results showed that the reflectivity of 409 the enhanced image kept high consistency with the original image, and the spatial details were finer.

410
On the other hand, the accuracy of the algorithm on vegetation pixels was higher than that on 411 non-vegetation pixels. Results showed that the pixels with poor EA were generally existed in cities, 412 especially those buildings with glass curtain walls or other high-reflective materials (such as 413 shopping malls, high-speed railway stations or airports). This kind of ground object had undergone 414 mirror reflection, which made the reflectivity of the original image abnormal (more than 100%), so 415 the reflectivity accuracy of the enhanced image was low. However, the pixels with abnormal 416 reflectance only accounted for about 1% of all pixels, and our research target was vegetation, and 417 consequently, the error was negligible.

418
The comparation of retrieved summer corn canopy LAI, CCC and LCC using enhanced image 419 and original image was done to validate the improvement of enhanced image to the original image.

423
When calculated vegetation index (MCARI, MTCI and AIVI), the enhanced image used the same 424 band as the original image with different spatial resolution. The results (Table 4) showed that the 425 parameter accuracy of the image enhanced by the SupReME algorithm was higher than that of the 426 original image. Our work confirmed that the SupReME algorithm not only enhanced the spatial 427 details of red-edge bands of Sentinel-2 image, but also improved the retrieving accuracy of summer 428 corn canopy LAI, CCC and LCC.

438
There were two kinds of band combinations were used in cost function for vegetation 439 parameters retrieval in this study: one was the vegetation index, the other was the original bands.

440
The red-edge bands of Sentinel-2 were incorporated especially into vegetation indexes for LAI and 441 CCC retrieval. For the combination of two bands, Fig. 6

465
In addition, we found that the retrieval accuracy of CCC, LAI and LCC using whether

490
Currently, Sentinel-2 is the only satellite available free of charge which include multiple red-edge 491 bands with medium spatial resolution (20m). However, it is possible to improve the resolution of 492 the red-edge band because of other 10m spatial resolution images. The SupReME algorithm 493 produced a series of image products with a spatial resolution of 10m without increasing any cost.

494
Because these images come from the same sensor, the uncertainty of these super-resolution 495 enhanced images is less affected by the observation angle, imaging time and spectral response 496 function and other reasons. It can increase spatial resolution while maintaining spectral consistency 497 (Table 3), and the product accuracy is higher (Lanaras et al. 2017). 498 Due to the management style of household in China's rural areas, the area of most planting field 499 is small. In addition, the difference of time, variety and management of crops planted by farmers, the 500 difference between adjacent fields will be more obvious. In the application of agricultural 501 quantitative remote sensing, the smallest unit of study is the pixel, which is used as the benchmark 502 for the field measurement and all models. The purity or variability of the pixels has an important 503 influence on the final results, especially when the field is fragmented (Huang et al., 2019). The 504 spatial heterogeneity of vegetation in one pixel of 20m spatial resolution is greater than that of 10m 505 spatial resolution (Zheng et al., 2017). Therefore, the 10m spatial resolution image could reduce the 506 uncertainty caused by these reasons, which could be used to improve the accuracy of the crop 507 canopy parameters retrieval, crop growth monitoring and estimation of crop yield.

524
We enhanced the Sentinel-2 images using the SupReME algorithm for generating 10m 525 multi-spectral images, retrieving LAI, CCC and LCC of corn canopy using linear model and physical 526 model on large-scale regions. We draw the following conclusions: (1) The SupReME algorithm had 527 high accuracy in fusing Sentinel-2 image, and its spectrum remained basically consistency as the

533
We demonstrated that the enhanced image had high application value by quantitatively 534 analyzing the results. These results indicated that SupReME algorithm was an effective and 535 promising approach to enhance the potential and value of Sentinel-2 image at a regional scale in the 536 North China Plains. In future research, we will apply the SupReME algorithm to the Sentinel-2 537 images with long time series and also consider the uncertainty of the physical model, and deep 538 learning will be explored instead of cost function for monitoring vegetation growth and estimation 539 of crop yield at regional scale.   1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60  61  62  63  64  65