The Extraction Method Of Crop Planting Structure In Irrigation District Based On Multi-Source Optical Remote Sensing Collaboration

The rational utilization of cultivated land resources is of great significance for ensuring food security and improving people’s living standard.Efficient and accurate monitoring of crop planting structure information is the prerequisite for ensuring rational utilization of cultivated land resources,realizing intelligent and efficient agricultural production management and accurate evaluation of grain output.Multi source remote sensing collaboration is an effective way to achieve efficient and accurate monitoring of regional crop planting structures.To clarify the"temporal-space-spectrum"requirements of multi-source remote sensing data for planting structure monitoring,how to realize multi-source remote sensing data collaborative monitoring and obtain suitable methods for large-scale planting structure extraction are urgent problems,which need to be solved by using multi-source remote sensing to achieve accurate monitoring of large-scale crop planting structure.This study takes Hetao Irrigation District of Inner Mongolia as the research area,and multi-temporal Sentinel-2 images in 2021 as the data source,it proposed a method to optimize the temporal and spectral features by constructing a global separable index(SI_global)combined with feature recursive elimination algorithm.The applicability of combining the preferred"temporal-spectrum"feature with multiple planting structure extraction algorithms（Random Forest（RF）,XGBoost,U-Net,Deep Lab V3+）was evaluated,so as to determine the critical period and key bands of multi-source collaborative monitoring;The UAV remote sensing image was used as the data source to obtain multi-scale spatial resolution images,extract crop texture features,construct field models with different shapes and extract models for different planting structures.Study the impact of spatial resolution on the internal texture and external shape of the plot and its adaptability of different algorithms to comprehensively analyze the appropriate spatial resolution for extracting planting structures,and thus determine the spatial resolution required for multi-source collaborative images;Using Sentinel-2 and GF-2 remote sensing images as data sources,a super resolution reconstruction model of Sentinel-2 visible light images with spatial resolution increased by 4 times is proposed through super resolution network.And the method of multi-source remote sensing data collaboration by fusion of the reconstructed visible light image with other characteristic bands of Sentinel-2 image;The planting structure extraction model was constructed using multi-source collaborative improved remote sensing images as the data source,and its applicability for planting structure extraction was evaluated.Aiming at the problems of the vulnerability of crop structure extraction in HDI to cloud influence and the low applicability of the model for multi-year planting structure extraction,a general planting structure extraction model was proposed by integrating multi-temporal remote sensing images and deep learning algorithms in key periods to increase applicability and avoid cloud influence.The main research conclusions are as follows:（1）Optimization of temporal-spectrum features for planting structure extraction based on Sentinel-2.The feature selection method proposed in this paper,which combines global separable index with feature recursive elimination algorithm,can accurately obtain planting structure and extract critical period in terms of temporal features,and effectively remove redundant and retained key features in terms of spectral features.Using Sentinel-2 images as the data source,it is more advantageous to extract the planting structure by combining the optimal features with U-Net algorithm.Compared with XGBoost and RF,it can effectively avoid the"salt and pepper"effect,and the edge information extracted from the results is more in line with the actual situation compared to Deep Lab V3+algorithm.The key periods of wheat planting structure extraction using feature optimization method are jointing stage and heading stage,and visible light（B2-B4 band of Sentinel-2 image,B2-B4）,red edge（B5）and short-wave infrared（B11-B12）are the corresponding characteristic bands;The key periods of maize planting structure extraction are milk maturity and maturity,and visible light（B2-B4）,red edge（B5）and short-wave infrared（B11-B12）are the corresponding characteristic bands.The key periods of sunflower structure extraction are flowering stage and maturity stage,and red edge（B5-B7）and near infrared（B8-B8A）were the corresponding characteristic bands.The key period for extraction of squash planting structure is the drying period,and the corresponding characteristic bands are visible light（B2-B4）,red edge（B5）and short-wave infrared（B11-B12）.（2）Research on optimal spatial resolution of planting structure extraction in irrigated district based on UAV remote sensing.This paper proposes to comprehensively analyze and study the appropriate spatial resolution of multi-source remote sensing collaboration,based on the perspective of the influence of spatial resolution on the internal texture and external shape of the block and the adaptability of different algorithms.In terms of texture,when the spatial resolution reached 2.5m,the texture information of the four crops had a strong auxiliary effect on the extraction of planting structure.When the spatial resolution is less than 2.5m,wheat texture information will be weakened.When the spatial resolution is less than 5m,the auxiliary effect of the texture information of the four crops on the extraction of planting structure will be weakened.In terms of plot shape,the deformation degree of plot in images with different resolutions is related to the shortest side of the plot.When the short side of the plot is 10m,25m and 50m respectively,the spatial resolution of the image should reach 2m,5m and above10m respectively to meet the monitoring requirements.In terms of extraction methods,the deep learning algorithm has more advantages in identifying crop plots with large internal variability.For wheat planting structure extraction,the resolution should be about 2m,and the extraction accuracy is higher when the resolution is above 5m for the extraction of maize,sunflower and squash planting structure.（3）Research on multi-source remote sensing collaborative methods based on super-resolution and"spatial spectral"fusion.In terms of super-resolution,the super-resolution model constructed by visible light images of GF-2 and Sentinel-2 can break through the limitations of spatial resolution and monitoring range of remote sensing images,and obtain high-resolution visible light images of irrigation areas during critical periods;The data set constructed by GF-2 and Sentinel-2 images combined with the super resolution model obtained by DRN network is applicable to different spatiotemporal images,and can restore the deformation and internal texture information of ground objects in the images,PSNR can be increased by about 5d B and the SSIM can be increased by about 0.2;In terms of"space-spectrum"fusion,the reconstructed visible light image can be fused with other feature bands of Sentinel-2 image,which can break through the"space-spectrum"limitation of remote sensing image,and obtain high-resolution images of key feature bands in key time periods of irrigation area;The fusion model constructed by Gram-Schmidt fusion method combined with red band can better recover the deformation and texture information of ground objects in the image,PSNR can be improved by about 3d B,and SSIM can be improved by about 0.1.（4）Research on application of planting structure extraction in irrigated area.The use of collaborative improved image construction model can improve the accuracy of planting structure extraction from the accurate identification of plot edge information and the accurate identification of broken plots,among which m Io U can increase by 0.14%-2.11%,F1-score can increase by 0.5%-1.7%,and OA can increase by 0.3%-1%.The method proposed in this paper can effectively avoid the influence of clouds,and the results of estimating the area of county cash crops,grain crops and cultivated land in 2017-2019 in HDI have a precision R² of 0.87and a RMSE of15622.05ha,which is higher than the existing research.In summary,this paper identified the“temporal-space-spectrum”requirements of multi-source remote sensing collaboration on images,proposed a multi-source remote sensing data collaboration method,and evaluated the applicability of collaborative improved images for extracting crop planting structure.In the aspect of multi-source remote sensing collaboration,the method presented in this paper can provide theoretical reference for the further study of multi-source remote sensing collaboration.In the application of regional crop planting structure extraction,this method can provide technical support for accurate and efficient extraction of regional planting structure.