Delineation Of Farmland Boundaries And Estimation Of Aboveground Biomass In Rice Using High Resolution Satellite Imagery

With the development of social productivity,the smart farmland management system began to be established.The establishment of a smart farm management system provides new opportunities for the precision crops management at the field level.Farmland vector boundary is the basis and prerequisite for the establishment of the smart farmland management system.However,the processing of delineating small field boundaries is manual,which is time-consuming and laborious,and arbitrary.This ways cannot meet the demand of precision crop management.Therefore,setting up a new farmland boundary delineation method is vital to smart farmland management system.Otherwise,the non-destructive crops monitoring technology at the field scale supported by remote sensing technology is the core of the establishment of the smart farmland management system.How to use satellite imagery to obtain crop growth information more accurately and quickly at fieldscale has become the top priority in the establishment of smart farmland management system.Therefore,based on the actual demand of the smart farmland management system,this paper fouses on two important issues:1)the farmland boundaries automated delineation;and 2)rice aboveground biomass estimation at the field scale.The research results can be directly applied to the smart farmland management system.In addition,this research can provides strong theoretical and technical support for related research.In part of farmland boundaries automated delineation,we set up a new method for the delineation of farmland boundaries with spatial and phenological information derived from high resolution imagery.This method consists of the following four steps:1)Edge detection;2)Image segmentation;3)Identifing farmland objects based on phenology;4)Morphological post-processing.The results show that the spatial information based on WorldView-2 image and the phenological information of Planet image can be effectively complemented each other,which makes the overall identification accuracy of farmland objects increase by 6.24%～11.31%and Kappa coefficient by 0.13～0.22,respectively.Otherwise,the overall identification accuracy of farmland objects is between 94.98%and 98.84%,and the Kappa coefficient is between 0.90 and 0.98.This way is effective in term of objects indentifing.In addition,we notice that it is easy to indentify farmland objects automatically by using the spectral characteristics of the phenology phase of the "soil preperation" and the feature is very stable and has broad application prospects.By combining the spatial information derived from WorldView-2 imagery with the phenological information derived from Planet images,we can delineate small field boundaries efficiently,accurately and automatically;In particular,this method can work well only with two universal thresholds,and the requirements for high-resolution images are low.Additionally,the delineated farmland boundaries and the referenced boundaries are basically coincident,and most of them maintain a one-to-one relationship.The farmland boundary extraction accuracy is high.In terms of quantitative evaluation,more than 80%of the field under-segmentation rate(Sjunder)and over-segmentation rate(Sjover)are lower than 20%,and the extracted farmland boundaries and the referenced boundaries coincide with each other;more than 85%of the field.The matching degree(Sjlocal)is higher than 80%,and the farmland boundary is positioned accurately.In part of AGB estimation,we focus on three important issues in the AGB estimation at the field scale:1)the construction of a rice AGB estimation model involving multi-ecological zones;2)comparison of the performance of machine learning algorithms in rice AGB estimation;and 3)the application potential of object-oriented image analysis techniques in rice AGB estimation.And exploring the potential of WorldView-2 imagery in crop growth monitoring at the field scale.Results have demonstrated that there are obvious spectral differences in rice pixels extracted from satellite images at different ecological point.However,after histogram matching,the systematic spectral differences between different images disappear.The accuracy of the AGB estimated model based on the spectral index EVI and CIred-edge referring to different ecological point has been significantly improved,and the decision coefficients have been increased from the initial 0.06 and 0.53 to 0.52 and 0.68,respectively.In addition,we can find that the random forest algorithm(Random Forest,RF)has the highest estimation accuracy(R2=0.92,RMSE=0.52t/ha,RE=8%)among the six selected regression algorithms,and its computational efficiency is relatively high.Partial Least Squares Regression(PLSR)has obvious advantages in terms of computational efficiency,although the estimation accuracy of rice AGB is relatively low based on PLSR(R2=.71,RMSE=0.97t/ha,RE=16).It’s worth noting that the computational efficiency of PLSR algorithm is faster 18～1296 times than other machine learning algorithms.In conclusion,The RF and PLSR algorithms have important application value in rice AGB estimation.Additionally,compared with the traditional pixel-based image analysis method,object-oriented image analysis technology can further improve the estimation accuracy of rice AGB.Significantly,this improvement is particularly evident in the PLSR algorithm with R2 increasing from 0.72 to 0.80.RMSE and RE decreasing to 0.80t/ha and 13%,respectively.And the computational efficiency of obj ect-oriented parameter regression algorithm is more than 12 times that of pixel parameter regression algorithm.At the same time,the spatial distribution map of rice AGB based on obj ect-oriented image analysis technology is more in line with the actual needs of crop field management.These suggest that object-oriented image analysis technology has great potential in the field of crop precision management.