Research On Image Super-Resolution And Red Tide Detection Methods Of GF-1 WFV Satellite Images

The processing and recognition of satellite remote sensing images is an important application field of computer technology such as image processing and computer vision.With the development of aerospace technology,satellite remote sensing technology has been increasingly applied to the monitoring of marine disasters,such as red tide and green tide.Red tide is one of the major ecological disasters in China,which endangers marine ecological system,marine aquaculture industry,and seashore tourism.For their high spectral resolution,the ocean color satellites have been widely used in red tide monitoring.However,their data usually has a low spatial resolution and cannot be applied to monitor the fine-scale red tides.As the first satellite of Chinese high spatial resolution satellites series,GF-1 with four Wide Field of View(WFV)cameras on board,has the capability of large area monitoring,which meets the data requirements of red tide operational monitoring.However,the spectral resolution of GF-1 WFV is low(>50nm),including four broad spectral bands: blue,green,red and near-infrared.For the shortage of corresponding spectral bands,the red tide detection by GF-1 WFV images is of great challenge.First,there are lots of mixed pixels in red tide images.While,the existing image super-resolution models pay too much attention to low-frequency information and ignore the preservation of spectral information.Second,the response of red tide varies greatly among different dominant species and different growth states,resulting in the low detection accuracy.Third,the low biomass in the red tide distribution edge area,results in the weak response of satellite images,and the existing red tide detection algorithms with high threshold sensitivity has poor ability to extract the red tide edge.To address these issues,the deep learning methods,International Commission on Illumination(CIE)color space system,and side window convolution were introduced in GF-1 WFV images super-resolution and red tide detection.The details are as follows:1)The existing image super-resolution models pay too much attention to low-frequency information and ignore the preservation of spectral information.Therefore,a novel deep residual coordinate attention network(GFRCAN)was proposed to enhance the spatial resolution of GF-1 WFV images from 16 m to 8 m,to improve the detection in the red tide distribution edge areas.To form a very deep network,the residual-in-residual(RIR)structure consisting of several residual groups(RG)with long skip connections was used.Meanwhile,the residual coordinate attention block(RCOAB)and adaptive multi-scale spatial attention module(AMSA)were incorporated to focus on the high-frequency information and multi-scale features adaptive weighted fusion.Besides,the spectral and spatial details of SR images are improved by incorporating peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)into the loss function.Experiment results show that GFRCAN achieves good superresolution performance and has a good performance in spectral information preservation.GFRCAN also has good applicability and robustness,and it is proved to be suitable for the super resolution of other satellite images,such as GF-4.2)In view of the weak red tide characterization of the traditional CIE color system,a new red tide detection model based on pseudo hue angle named PHA-RI for GF-1 WFV was proposed in this paper.Different from the standard CIE color system,the false-color bands of near-infrared,red and green are involved in the calculation of the CIE tristimulus X,Y,Z,instead of the true-color bands of red,green and blue.Through this replacement,the distinction between red tide and non-red tide was doubled compared to that of true-color bands composite.For tristimulus Z is more sensitive to the suspended particulate matter,the tristimulus Z was selected as the indicator of turbid water to reduce the impact of sea water.Experiment results show that the PHA-RI model can effectively detect red tide.PHA-RI also has good applicability and robustness,and it is proved to be suitable for the detection of red tides with different dominant species including Noctiluca scintillans,Skeletonema costatum and Heterosigma akashiide,and different broad band high spatial resolution sensors,such as HY-1D CZI,Sentinel-2 MSI and Landsat 8 OLI.3)To address the problem of red tide detection in red tide distribution edge area and the strong threshold dependence of the traditional algorithms,a novel deep learning based red tide detection method was proposed incorporating multiple features and super resolution model.Based on the super resolution model GFRCAN,the side window convolution was also introduced to enhance the red tide edge and reduce the impaction of salt and pepper noise.Multiple features including GF1-RI and pseudo hue angle were also incorporated in the model to improve red tide detection accuracy.Experiments results show that the proposed method can effectively detect red tide,especially in red tide distribution edge area,without any thresholds.