Search

CN-122024101-A - Soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning

CN122024101ACN 122024101 ACN122024101 ACN 122024101ACN-122024101-A

Abstract

The invention discloses a soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning, which comprises the steps of obtaining unmanned aerial vehicle multi-spectrum data to extract local reflection features, feature pyramid fusion multi-scale information, densely connecting a network to aggregate multi-source remote sensing features, processing dynamic changes by a time sequence convolution network to obtain a space-time evolution sequence, geometrically correcting a space transformation module, obtaining a pixel level estimation result by an end-to-end prediction model, iteratively adjusting to obtain optimized positioning distribution when in dynamic abnormality, determining final component estimation by matching the multi-source features, and extracting residual salinity spectrum fingerprint features by the hyperspectral data to realize soil salinity identification. The invention realizes high-precision estimation and type identification of soil salinity components.

Inventors

  • HOU FUJIANG
  • Sang Yazhuan

Assignees

  • 兰州大学

Dates

Publication Date
20260512
Application Date
20260203

Claims (10)

  1. 1. A soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning is characterized by comprising the following steps: The method comprises the steps of acquiring multispectral data acquired by an unmanned aerial vehicle, and extracting local reflection characteristics through a convolution layer to obtain a preliminary soil salinity reflection mode; According to the preliminary soil salinity reflection mode, adopting a characteristic pyramid structure to fuse multi-scale distribution information, and determining a multi-scale salinity distribution trend; aggregating multisource remote sensing features and the multiscale salinity distribution trend through a dense connection network to obtain a comprehensive salinity feature set; if the comprehensive salinity characteristic set has time sequence correlation, processing dynamic change by adopting a time sequence convolution network to obtain a salinity space-time evolution sequence; extracting a spatial distribution mode from the salt space-time evolution sequence, compensating displacement invariance through a spatial transformation module, and determining a geometrically corrected salt spectrum mode; According to the geometrically corrected salt spectrum mode, an end-to-end prediction model is constructed to integrate a global reflection mode, and a pixel-level salt content estimation result is obtained; If the pixel-level salt content estimation result shows that the dynamic change is abnormal, iteratively cycling to the time sequence convolution network to adjust an evolution sequence, and obtaining optimized accurate positioning distribution; Judging the accuracy of salt type identification through matching the optimized accurate positioning distribution with the multi-source remote sensing characteristics, and determining final soil salt component estimation; acquiring hyperspectral imaging data from the final soil salinity component estimation, and adopting a three-dimensional convolutional neural network to process the joint information to acquire spectral fingerprint characteristics of the residual salinity type; And classifying according to the spectral fingerprint characteristics of the residual salinity types to obtain a soil salinity identification result.
  2. 2. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for obtaining the preliminary soil salinity reflective mode comprises the following steps: acquiring multispectral data of the unmanned aerial vehicle as an input item, calculating a reflection value in a local area through a convolution layer, and generating a feature map containing local reflection features; overlapping the feature images by adopting a preset convolution kernel size, and enhancing the space association information in the feature set; if the spatial correlation information exceeds a preset threshold, judging that a significant soil salinity reflection mode exists in the area; extracting a corresponding feature set from the feature map according to the salient pattern area, and generating a pattern map describing soil salt graduation; and carrying out channel dimension combination on the pattern diagram and the spectrum data to form a final data stream containing multi-dimension information.
  3. 3. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for determining the multi-scale salt distribution trend comprises the following steps: Acquiring preliminary data through the preliminary soil salinity reflection mode, and performing layering arrangement on the preliminary data to obtain an initial mapping result of salinity distribution; Constructing a feature pyramid structure according to the initial mapping result, decomposing the multi-scale distribution layer by layer, and determining the distribution information of each level; Carrying out data recombination by combining the decomposed distribution information with the scale change to obtain a multi-scale salt distribution characteristic set; performing deep analysis on the distribution information by using a convolutional neural network through the multi-scale salinity distribution feature set, and judging the change rule of salinity trend under different scales; extracting key distribution information according to the change rule, and performing data screening to obtain core influence factors of salt distribution; Based on the core influence factors, fusing a multi-scale analysis result, and determining an overall distribution model of the salinity trend; and if the overall distribution model has deviation from the primary analysis result, carrying out secondary correction on the data to obtain a final salt distribution trend result.
  4. 4. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for obtaining the comprehensive salinity characteristic set comprises the following steps: acquiring multi-source remote sensing image data, and extracting spectral features and texture features by adopting a preset wave band combination rule; calculating local area statistics according to a multi-scale sliding window strategy to obtain a multi-scale distribution trend; and inputting the spectral features, the texture features and the multi-scale trends into a preset dense connection network, performing feature aggregation and nonlinear transformation based on the dense connection network, and outputting a comprehensive salinity feature set.
  5. 5. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for obtaining the salt space-time evolution sequence comprises the following steps: extracting time dimension information from the comprehensive salinity characteristic set, judging whether time sequence correlation characteristics exist or not, and obtaining a time sequence correlation evaluation result; If the time sequence correlation evaluation result shows that the correlation exists, modeling processing is carried out on the dynamic change by adopting a time sequence convolution network, and the characteristic representation of the dynamic change is obtained; carrying out serialization processing on the time-space evolution according to the characteristic representation of the dynamic change to generate preliminary evolution sequence data; the preliminary evolution sequence data is subjected to smoothing processing, noise interference is eliminated, and optimized evolution sequence content is determined; extracting time sequence characteristic distribution of the salinity data by adopting the optimized evolution sequence content to obtain key nodes of time-space evolution; and generating a final salt space-time evolution sequence according to the key nodes of the space-time evolution.
  6. 6. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of determining geometrically corrected salt spectrum patterns includes: Acquiring time sequence information from the salt space-time evolution sequence, and constructing an initial spatial distribution data set; Extracting a distribution rule according to the initial spatial distribution data set to form a preliminary distribution rule record; According to the preliminary distribution rule record, the displacement influence is processed by using a space transformation technology, and transformed space distribution data are generated; If the transformed spatial distribution data have deviation, the spatial distribution information after correction is obtained by comparing with a preset threshold value to adjust; extracting spectral distribution characteristics from the corrected spatial distribution information, and determining a corresponding spectral distribution mode; After the spectrum distribution mode is obtained, carrying out final processing on the data by combining a geometric correction method, and judging a correction result meeting the requirements; and integrating the correction results, classifying the distribution modes of the salt data by adopting a support vector machine algorithm, and determining the final distribution mode category.
  7. 7. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for obtaining the pixel-level salt content estimation result comprises the following steps: Removing noise interference by adopting a preprocessing method according to the geometrically corrected salt spectrum data, and smoothing a spectrum curve by a filtering technology to obtain a processed spectrum data set; Acquiring global reflection characteristic data according to the processed spectrum data set, extracting key wave band information in the reflection characteristic data by using a statistical analysis method, and determining a spectrum response interval related to the salt content; Constructing an end-to-end prediction model according to the spectrum response interval, and training the salinity spectrum data by adopting a support vector regression method to obtain a preliminary salinity content prediction value; According to the preliminary salt content predicted value, fusing spatial distribution information of pixel levels, if the predicted value is within a preset threshold range, reserving predicted data of pixels, and if the predicted value is beyond the threshold range, correcting by a neighborhood interpolation method to obtain an adjusted predicted data set; according to the adjusted prediction data set, carrying out secondary calibration on the salinity content of each pixel by combining with the global reflection characteristic, and if the deviation between the calibrated value and the initial prediction value exceeds a preset threshold value, carrying out smoothing treatment by adopting a weighted average method to determine final pixel level salinity content data; and generating a grid result of salt distribution according to the final pixel level salt content data, and obtaining a spatially continuous estimation result.
  8. 8. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of obtaining the optimized precise positioning distribution comprises the following steps: Acquiring salt content estimation data of pixel levels, primarily monitoring the dynamic changes, and judging whether abnormal performance exists or not to obtain abnormal marking data; if abnormal marking data are detected, analyzing the evolution sequence of dynamic change through a time sequence convolution network, and determining specific time period distribution of abnormal expression; According to the abnormal time interval distribution, locally correcting the evolution sequence by adopting a convolution adjustment method to obtain adjusted sequence data; the distribution model with accurate positioning is constructed by optimizing the adjusted sequence data, so that the optimized distribution characteristics are obtained; according to the optimized distribution characteristics, recalculating an estimation result of the salt content, judging whether the salt content accords with a preset threshold range, and determining a final distribution result; And generating a corresponding salt content space map according to the final distribution result, and obtaining comprehensive and accurate positioning information.
  9. 9. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process of determining the final soil salinity estimation comprises: acquiring remote sensing image data and soil sample point space information of a target area, optimizing positioning distribution according to the soil sample point space information, and generating a sample point space weight matrix; Spectral features of positions corresponding to sample points are extracted from the remote sensing image data, and a multi-source feature set is constructed; matching the multi-source feature set with a spectrum feature library with known salinity types by adopting a preset matching rule; If the matching degree is higher than a preset threshold, judging that the salt type identification is accurate, and obtaining a salt type label; And determining the estimated value of the soil salinity component by adopting a regression model according to the salinity type label and the spectral characteristics.
  10. 10. The method of claim 1, wherein the step of determining the position of the substrate comprises, The process for obtaining the spectral fingerprint characteristics of the residual salt type comprises the following steps: acquiring hyperspectral imaging data from a soil salinity component estimation result, and performing depth feature extraction on hyperspectral images by adopting a three-dimensional convolutional neural network to acquire preliminary spectral features of residual salinity in soil; According to the preliminary spectral characteristics, combining soil salinity distribution data in the combined information, and preliminarily distinguishing the types of residual salinity by a characteristic matching method to determine a type classification result; if the confidence coefficient of the type classification result is lower than a preset threshold value, carrying out secondary analysis on the primary spectrum characteristics to obtain fine spectrum fingerprint information; Judging the specific attribution of the residual salt type by comparing the fine spectrum fingerprint information, and obtaining a final type judging result; and generating a spectral fingerprint characteristic data set of residual salt according to the final type judgment result, and storing the spectral fingerprint characteristic data set into a pre-established database.

Description

Soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning Technical Field The invention belongs to the field of soil salinity identification, and particularly relates to a soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning. Background Soil salinity is an important challenge in agricultural production and ecological environment protection, and directly affects sustainable use of land and grain safety. Salinization not only can lead to the soil fertility to be reduced, but also can threaten vegetation growth, thereby aggravating land degradation. In the global scope, especially in arid and semiarid regions, soil salinity monitoring and management have become key problems to be solved urgently, and research of the soil salinity monitoring and management has non-negligible value for improving agricultural production efficiency and maintaining ecological balance. However, the current methods for soil salinity identification have significant limitations in practical applications. Many conventional techniques often rely on a single observation means or a fixed analysis framework, and are difficult to adapt to complex and diverse natural environments, especially when faced with different terrains, climatic conditions and soil types, recognition accuracy and adaptability are often limited. More importantly, when the method processes multi-source information, the lack of mining capability of deep association between different data features leads to insufficient understanding of salt distribution rules, and the requirements of accurate agriculture and dynamic monitoring are difficult to meet. In this field, core technical difficulties have focused on how to efficiently integrate data from a variety of sensing devices and extract representative features therefrom. In particular, the information acquired by different devices has great difference in spatial scale and expression form, for example, high-resolution images can capture fine textures of soil surfaces, and medium-low resolution data are more suitable for reflecting a wide-range distribution trend. The difference makes fusion information become extremely complex under a unified frame, and further influences the comprehensive characterization of salt distribution from microscopic to macroscopic. In addition, the complexity of information fusion brings about another problem that the expression form of salt characteristics can be changed under different environmental conditions, for example, the soil with the same salt level can present distinct appearance characteristics under different illumination or shooting angles, and the difficulty of identification is increased. Therefore, how to construct an identification system capable of processing microcosmic details and macroscopic patterns simultaneously on the basis of multi-source information and ensuring stable identification of salinity characteristics under different environmental conditions becomes a key problem of accurate monitoring and dynamic evaluation of soil salinity. The solution of this problem will directly affect the timely intervention and scientific decision-making of salinized lands in agricultural management. Disclosure of Invention In order to solve the technical problems, the invention provides a soil salinity identification method based on unmanned aerial vehicle multi-source remote sensing feature fusion and deep learning, which comprises the following steps: The method comprises the steps of acquiring multispectral data acquired by an unmanned aerial vehicle, and extracting local reflection characteristics through a convolution layer to obtain a preliminary soil salinity reflection mode; According to the preliminary soil salinity reflection mode, adopting a characteristic pyramid structure to fuse multi-scale distribution information, and determining a multi-scale salinity distribution trend; aggregating multisource remote sensing features and the multiscale salinity distribution trend through a dense connection network to obtain a comprehensive salinity feature set; if the comprehensive salinity characteristic set has time sequence correlation, processing dynamic change by adopting a time sequence convolution network to obtain a salinity space-time evolution sequence; extracting a spatial distribution mode from the salt space-time evolution sequence, compensating displacement invariance through a spatial transformation module, and determining a geometrically corrected salt spectrum mode; According to the geometrically corrected salt spectrum mode, an end-to-end prediction model is constructed to integrate a global reflection mode, and a pixel-level salt content estimation result is obtained; If the pixel-level salt content estimation result shows that the dynamic change is abnormal, iteratively cycling to the time sequence convolution network