Study Of Interpretable Universal Spectral Index Design Method

Spectral index is a widely used data processing method in optical remote sensing technology.Using spectral index can quickly extract information related to the study target from the complex spectral data.Compared with the original band,spectral index is a feature closer to the study target,and it is often used as an input feature instead of the original band in remote sensing applications such as feature classification,change monitoring,and biological variable inversion.However,it is a challenge to select the most suitable spectral indices for different remote sensing applications.Therefore,it is necessary to develop site-specific spectral index design methods to respond to the needs of remote sensing applications.The traditional method of designing spectral indices based on human experience is inefficient and human resource-intensive,which is not applicable when the data volume is too large or the data is too complex;the method of designing spectral indices based on machine learning end-to-end frees up human resources,but the spectral indices thus designed lack interpretability and are difficult to be used widely.In the face of the above problems,this paper proposes a general and interpretable spectral index design method,which can design reasonable spectral indices for spectral data of specific sensors,specific study areas and specific study targets,and the obtained indices have the features of highlighting the spectral characteristics of the features,being sensitive to the changes of target coverage,and attenuating or eliminating the non-target-related interference factors.Compared with the traditional method of designing spectral indices based on human experience,the method proposed in this paper is based on machine learning model without relying on human,which frees manpower and is more efficient.The research work in this paper is divided into three main parts as follows.(1)The structure and parameters of spectral data are investigated,and the common preprocessing steps of spectral data are analyzed to clarify the physical principles and mathematical models involved.The error terms generated during the conversion from raw sensor imaging data to surface reflectance data are sorted out,and the causes of the three main error sources are explained and improvement suggestions are made to provide ideas and targets for the subsequent design of the spectral index.(2)Based on the summarized objectives and ideas of spectral index design,an interpretable general spectral index design method is proposed.Firstly,the band selection algorithm of the spectral index is designed,an interpretable neural network is innovatively introduced,and a band selection algorithm based on weight analysis is developed as a result.Further,a twodimensional spectral index form satisfying the design requirements is derived based on the results of band selection,and then the sensitivity of the index is innovatively adjusted using a scaling factor.Finally,an auxiliary method is proposed for guiding and simplifying the index design.(3)The interpretability and generality of the proposed spectral index design method are verified.Vegetation indices were created based on actual data and study areas using the spectral index design method proposed in this paper,and the specific role and rationale behind the method was explained to verify the interpretability of the method.Based on multiple types of spectral data,the band selection algorithm proposed in this paper was used to select the characteristic bands for four common types of features: buildings,vegetation,water bodies,and soils,and the results all met the research objectives,verifying the generality of the method for different types of multispectral and hyperspectral data.