CN-121982549-A - Automatic construction method for crop remote sensing sample with time-space-spectrum combined characteristics

CN121982549ACN 121982549 ACN121982549 ACN 121982549ACN-121982549-A

Abstract

The invention discloses an automatic construction method of crop remote sensing samples with time-space-spectrum combined characteristics, which comprises the following steps of obtaining multi-period optical remote sensing images of target years in a research area, traversing unlabeled cultivated land pixels in the research area, calculating the similarity of time sequence characteristics, spectrum characteristics and space characteristics of each cultivated land pixel and a corresponding crop type characteristic signature library in time sequence dimensions, spectrum dimensions and space dimensions, judging the comprehensive similarity, performing space connectivity verification on candidate amplified sample points, and merging final amplified sample points with a core sample library to form a mass enhanced sample library. According to the invention, through a step-by-step filtering mechanism of time sequence consistency screening, spectrum consistency screening, space purity screening and neighborhood verification, multi-source characteristic information of time dimension, spectrum dimension and space dimension is fused, and a core sample library with high purity is constructed.

Inventors

LV ZHENG
WANG XIAOYAN
GONG YALI
CHEN WEIRONG
ZHENG JINJIN
ZHU XIAOBO

Assignees

中国资源卫星应用中心

Dates

Publication Date: 20260505
Application Date: 20260206

Claims (10)

1. The automatic construction method of the crop remote sensing sample with the time-space-spectrum combined characteristic is characterized by comprising the following steps of: acquiring multiple optical remote sensing images of a target year in a research area, and reconstructing a vegetation index time sequence generated by the optical remote sensing images to obtain a reconstructed time sequence vegetation index curve; Extracting time sequence characteristics from each sample point in the initial sample pool through the time sequence vegetation index curve, extracting spectral characteristics from the optical remote sensing image, extracting spatial characteristics from a neighborhood analysis window, and sequentially performing time sequence consistency screening, spectral consistency screening, spatial purity screening and neighborhood verification to form a core sample library; Constructing a characteristic signature library of each crop type according to the space-time spectrum combined characteristics of each crop type sample point in the core sample library; Traversing unmarked cultivated land pixels in the research area, calculating the similarity of time sequence features, spectrum features and space features of each cultivated land pixel and a corresponding crop type feature signature library in time sequence dimension, spectrum dimension and space dimension, and carrying out fusion calculation on the similarity of the three dimensions to obtain the comprehensive similarity of each cultivated land pixel for each crop type; And judging the comprehensive similarity to obtain candidate amplified sample points, performing space connectivity verification on the candidate amplified sample points, taking the verified candidate amplified sample points as final amplified sample points, and combining the final amplified sample points with the core sample library to form a mass enhanced sample library.
2. The method for automatically constructing a remote sensing sample of a crop with combined time-space-spectrum features according to claim 1, wherein the step of reconstructing a time sequence of vegetation indexes generated by the optical remote sensing image comprises: Calculating normalized vegetation indexes, enhanced vegetation indexes, normalized red edge indexes and normalized difference moisture indexes from the optical remote sensing image in a scene-by-scene manner, and generating a time sequence of each vegetation index; And processing the time sequence of each vegetation index, filling up missing data points in the time sequence, and smoothing existing noise data points to obtain a complete continuous time sequence vegetation index curve.
3. The method for automatically constructing a remote sensing sample of a crop with combined time-space-spectrum features according to claim 2, wherein the step of timing consistency screening comprises: establishing a standard time sequence template of the target crop type; Calculating the curve similarity distance between the time sequence characteristics of each sample point and the standard time sequence template for each sample point in the initial sample pool; And eliminating sample points with curve similarity distance larger than a first threshold value, and reserving sample points with curve similarity distance smaller than or equal to the first threshold value to obtain a time sequence consistency sample set.
4. The automatic construction method of crop remote sensing samples with time-space-spectrum combined characteristics according to claim 3, wherein a time warping algorithm is adopted when calculating the curve similarity distance; the time normalization algorithm searches a matching path with the minimum accumulated cost in a corresponding relation matrix by establishing the corresponding relation matrix between a sample point time sequence curve and a standard time sequence template; the cumulative cost value of the matching path is used as the curve similarity distance.
5. The method for automatically constructing a remote sensing sample of a crop characterized by a time-space-spectrum combination as claimed in claim 4, wherein the step of screening for spectral consistency comprises: extracting multispectral wave band reflectivity values of each sample point screened by time sequence consistency in a critical growth period of crops, and taking the multispectral wave band reflectivity values as spectral feature vectors; clustering operation is carried out on the spectral feature vectors of all the sample points, the sample points with similar spectral features are divided into the same cluster, and the sample points with deviated spectral features are marked as outlier sample points; and removing the outlier sample points, and reserving sample points belonging to the cluster to obtain a spectrum consistency sample set.
6. The method for automatically constructing a remote sensing sample of a crop with combined time-space-spectrum features according to claim 5, wherein the step of screening spatial purity comprises: for each sample point screened by spectrum consistency, establishing a neighborhood analysis window by taking the sample point as a center, and calculating a gray level co-occurrence matrix in the neighborhood analysis window; extracting texture feature parameters through the gray level co-occurrence matrix; And judging the homogeneity degree of the area around the sample point according to the texture characteristic parameters, and removing the sample point with the homogeneity degree smaller than a second threshold value, or with the contrast degree larger than a third threshold value, or with the variance larger than a fourth threshold value, so as to obtain a spatially pure sample set.
7. The method for automatically constructing a remote sensing sample of a crop with combined time-space-spectrum features of claim 6, wherein the step of neighborhood verification comprises: Extracting 8 adjacent neighborhood pixels around each sample point which is screened by the spatial purity; respectively calculating curve similarity distances between the time sequence features of the 8 neighborhood pixels and a crop standard time sequence template, and judging whether the curve similarity distances are smaller than a fifth threshold value; Respectively calculating the spatial homogeneity indexes of the 8 neighborhood pixels, and judging whether the spatial homogeneity indexes are larger than a sixth threshold; counting the number of neighborhood pixels in the 8 neighborhood pixels, wherein the number of neighborhood pixels simultaneously satisfies that the curve similarity distance is smaller than the fifth threshold and the spatial homogeneity index is larger than the sixth threshold; and when the number of the neighborhood pixels meeting the condition reaches 6 or more, reserving the sample point to form the core sample library.
8. The method for automatically constructing a remote sensing sample of a crop with combined time-space-spectrum features of claim 7, wherein the step of constructing a signature library of features for each crop type comprises: Respectively counting time sequence characteristics, spectrum characteristics and space characteristics of all sample points of each crop type aiming at each crop type in the core sample library; Carrying out statistical analysis on the time sequence characteristics, and calculating an average curve of time sequence vegetation index curves of all sample points of the crop type as a standard time sequence template of the crop type; Carrying out statistical analysis on the spectral characteristics, and calculating the mean value and covariance of the spectral characteristic vectors of all sample points of the crop type as the spectral characteristic statistical parameters of the crop type; and carrying out statistical analysis on the spatial characteristics, and calculating the numerical range of texture characteristic parameters of all sample points of the crop type as the spatial characteristic reference range of the crop type.
9. The method for automatically constructing a remote sensing sample of a crop characterized by a time-space-spectrum combination as claimed in claim 8, wherein the step of obtaining the integrated similarity of each cultivated land pixel for each crop type comprises: For each cultivated land pixel, extracting a time sequence vegetation index curve as a time sequence feature, calculating a curve similarity distance between the time sequence feature and a standard time sequence template in a feature signature library, and converting the curve similarity distance into a time sequence dimension similarity; Extracting a multispectral wave band reflectivity value of each cultivated land pixel in a critical growth period of crops as a spectral feature vector, calculating a feature space distance between the spectral feature vector and a spectral feature statistical parameter in a feature signature library, and converting the feature space distance into spectral dimension similarity; Extracting texture feature parameters in each cultivated land pixel neighborhood analysis window as space features, calculating the deviation degree between the texture feature parameters and a space feature reference range in a feature signature library, and converting the deviation degree into space dimension similarity; And respectively normalizing the time sequence dimension similarity, the spectrum dimension similarity and the space dimension similarity to a numerical value interval of 0 to 1, and carrying out weighted summation on the three normalized dimension similarities to obtain the comprehensive similarity.
10. The method for automatically constructing a remote sensing sample of a crop for a time-space-spectrum joint feature of claim 9, wherein the step of determining the integrated similarity and performing a spatial connectivity verification comprises: Setting a comprehensive similarity threshold, screening cultivated land pixels with comprehensive similarity larger than the comprehensive similarity threshold, and marking the screened cultivated land pixels as candidate amplified sample points; Carrying out spatial position analysis on the candidate amplified sample points, and judging whether each candidate amplified sample point is adjacent to a sample point in the core sample library in spatial position; Judging whether each candidate amplified sample point is adjacent to other candidate amplified sample points in space position; reserving candidate amplified sample points adjacent to sample points in the core sample library or adjacent to other candidate amplified sample points as the final amplified sample points; candidate amplified sample points that exist in isolation at spatial locations and are not adjacent to any sample point are removed.

Description

Automatic construction method for crop remote sensing sample with time-space-spectrum combined characteristics Technical Field The invention relates to the technical field of remote sensing data processing, in particular to an automatic construction method of a crop remote sensing sample with time-space-spectrum combined characteristics. Background The remote sensing monitoring of the crop planting area is an important technical means for agricultural resource investigation and grain safety evaluation. The quality and quantity of the remote sensing samples are used as training data for crop classification and identification, and the classification accuracy and the monitoring reliability are directly affected by the remote sensing samples. Traditional remote sensing sample acquisition mainly relies on manual field investigation and visual interpretation, which is time-consuming, labor-consuming and costly, and makes it difficult to acquire a sufficient number of high quality samples over a wide area. In the existing remote sensing sample construction method, the sample screening process only considers single dimension characteristics, such as sample selection only according to spectrum characteristics or time sequence characteristics, and the lack of comprehensive utilization of time, space and spectrum information leads to insufficient purity of samples and more confusing samples, and the sample amplification process lacks strict space connectivity constraint and multi-dimensional similarity verification, so that isolated noise pixels or pixels with inconsistent characteristics are easily incorporated into a sample library, and the reliability and representativeness of amplified samples are reduced. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a crop remote sensing sample automatic construction method with time-space-spectrum combined characteristics, which solves the problems that the existing sample construction method only considers single dimension characteristics, the purity of the sample is insufficient, the sample amplification process lacks strict space connectivity constraint and multi-dimension similarity verification, and the reliability of the amplified sample is low. In order to solve the technical problems, the invention provides the following technical scheme: the invention provides an automatic construction method of crop remote sensing samples with time-space-spectrum joint characteristics, which comprises the following steps of obtaining multi-period optical remote sensing images of target years in a research area, carrying out reconstruction processing on vegetation index time sequences generated by the optical remote sensing images to obtain reconstructed time sequence vegetation index curves, extracting time sequence characteristics from each sample point in an initial sample pool through the time sequence vegetation index curves, extracting spectral characteristics from the optical remote sensing images, extracting spatial characteristics from a neighborhood analysis window, sequentially carrying out time sequence consistency screening, spectral consistency screening, spatial purity screening and neighborhood verification to form a core sample library, constructing a characteristic signature library of each crop type according to the time-space spectrum joint characteristics of each crop type sample point in the core sample library, traversing cultivated land pixels which are not marked in the research area, calculating the similarity of the time sequence characteristics, the spectral characteristics and the spatial characteristics of each cultivated land pixel and the corresponding crop type characteristic signature library in the time sequence dimension, the spectral dimension and the spatial dimension, carrying out fusion calculation on the similarity of three dimensions to obtain the cultivated land pixels, carrying out comprehensive amplification of each crop type sample point by aiming at each crop type comprehensive characteristics through a neighborhood analysis window, carrying out amplification point amplification candidate amplification, and carrying out final amplification on the amplification point amplification candidate sample amplification by carrying out the comprehensive amplification on the amplification point by the candidate sample amplification. The method for automatically constructing the crop remote sensing sample with the time-space-spectrum combined characteristic comprises the steps of calculating normalized vegetation indexes, enhanced vegetation indexes, normalized red edge indexes and normalized difference moisture indexes from scene to scene for the optical remote sensing image to generate time sequences of all the vegetation indexes, processing the time sequences of all the vegetation indexes, filling missing data points in the time seque