CN-121661513-B - Cultivation area extraction method based on time series filtering

CN121661513BCN 121661513 BCN121661513 BCN 121661513BCN-121661513-B

Abstract

The invention discloses a culture area extraction method based on time series filtering. The method comprises the steps of obtaining multi-temporal optical image data of a target area, preprocessing the multi-temporal optical image data to construct a time sequence image data set, constructing a sample data set for training an extraction model of a culture area, constructing an extraction model of the culture area with hierarchical progressive characteristics and space self-adaption capability, learning spatial structural features and time evolution features of the culture area by the training model to obtain an initial mask of the culture area, performing time sequence filtering processing, restraining transient noise by constructing a time sequence stability judging mechanism to obtain extraction results of the culture area with consistent time sequence, constructing a culture area change analyzing module to realize automatic comparison analysis of spatial distribution of the culture area in different time phases, and obtaining boundary extraction and quantitative characterization of space area of the culture area in each time phase. The invention can effectively relieve the problems of spectrum confusion and space-time variation interference and improve the stability and reliability of the extraction result of the culture area in the time dimension.

Inventors

HE SHUANGYAN
WANG WENSI
ZHAO YUCHEN
JIN WEIWEI
LU SHIMING
Gu Yanzhen

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (8)

1. The culture area extraction method based on time sequence filtering is characterized by comprising the following steps of: Step 1, acquiring multi-temporal optical remote sensing image data of a target area, and performing spatial registration and area clipping pretreatment on the multi-temporal optical remote sensing image data to construct a time sequence image dataset of the target area; Step2, labeling the culture area and the non-culture area based on the time sequence image dataset, and extracting space domain information from each labeling object according to a preset space scale to form a sample dataset containing space context relation; Step 3, constructing a culture area extraction model, and training the culture area extraction model by utilizing a sample data set to enable the model to learn the joint representation of the spatial structural characteristics and the time evolution characteristics of the culture area; Step 4, inputting the time sequence image data set into a training-completed culture area extraction model, and obtaining a culture area extraction result of each time phase; Step 5, carrying out joint analysis on the extraction results of the culture areas in the same spatial position in a plurality of time phases, and constructing a time sequence stability judging mechanism so as to inhibit transient noise and enhance the consistency of the culture areas in the time dimension, so as to obtain the extraction results of the culture areas with consistent time sequences; Step 6, comparing and analyzing spatial distribution of the culture areas in different time phases, extracting boundaries of the expansion area and the extinction area of the culture areas, calculating the spatial area and the variation of the culture areas in each time phase, and realizing quantitative characterization of the aerodynamic evolution of the culture areas; The culture area extraction model comprises the following steps: The tensor construction module is used for converting the multi-temporal optical image sample acquired by the research area into a multi-dimensional tensor form suitable for model input so as to represent the spatial structure characteristics and the time evolution characteristics of the sample; The feature extraction module is used for carrying out feature coding on the tensor form sample and extracting intermediate feature representation capable of representing spatial structural features and time evolution features of the culture area; The first central perception module is used for explicitly constructing a characteristic organization mode taking the central position as a reference in the space dimension, taking the core response of the central region as a structural anchor point, and carrying out directional reconstruction by combining the context information of the surrounding region, so as to form a characteristic expression mechanism taking the center as a leading part, so as to highlight the core structural characteristics of the target region; The second center perception module is used for repeatedly executing center guiding and context reorganizing processes on a plurality of feature levels to enable discrimination information of a center area to be propagated step by step and accumulated and expressed, and carrying out self-adaptive fusion by combining association relations between centers and space positions, so that center perception enhancement feature representation with level progressive characteristics and space self-adaptive capacity is formed; The spatial relation module is used for constructing a global spatial dependency relation based on the central perception enhanced feature representation and generating an enhanced global central vector containing spatial context association information; And the classification module is used for generating an initial classification result of the culture area and the non-culture area based on the global center vector.
2. The method for extracting a culture area based on time series filtering according to claim 1, wherein the first central sensing module and the second central sensing module each comprise a central sensing unit, the central sensing unit comprises a central path and a peripheral path, the central path is used for extracting central information features from features output by the third convolution module based on central projection, and the peripheral path is used for extracting peripheral information features from the features output by the third convolution module.
3. The method for extracting the culture area based on time series filtering according to claim 2, wherein the central information feature is extracted by the central path in the following manner: Extracting intermediate features And (3) the spatial position of the center, and then carrying out 1X 1 convolution and up-sampling operation on the extracted spatial position in sequence to obtain the center information characteristic.
4. A method of extracting a farming area based on time series filtering according to claim 2, wherein the circumferential path comprises m circumferential convolution kernels and the manner of extracting the circumferential information features is as follows: ; Wherein, the As a feature of the central information, Extracting a feature of the circumferential information for the circumferential path, Representing a splice calculation per channel dimension, The output of the kth cyclic convolution kernel is represented, k=1, 2,..m.
5. The method of claim 1, wherein the spatial relationship module comprises a fourth convolution module configured to project channels in the feature matrix after center enhancement to a relationship dimension of the center projection and the circumferential projection.
6. The method for extracting a culture area based on time series filtering according to claim 5, wherein the spatial relationship module further comprises a fifth convolution module, and the fifth convolution module is configured to generate a weight distribution of a spatial position according to an inner product of a central projection and a circumferential projection in a relationship dimension, specifically as follows: flattening loops Zhou Touying along a spatial dimension to obtain a flattened tensor ; Calculating center projections And ring Zhou Touying Similarity score matrix of (a) The method comprises the following steps: ; Wherein, the For real numbers, B represents a single training batch, N represents the total number of spatial pixels; computing weight distribution of spatial locations The method comprises the following steps: ; Wherein, the Representing an activation function.
7. The method of claim 1, wherein the classification module comprises a two-layer fully-connected network, the first layer fully-connected network is configured to project the enhanced global center vector to the hidden dimension, and the second layer fully-connected network is configured to map the output result of the first layer fully-connected network to probabilities of the classification of the farming area or the non-farming area to obtain a preliminary binary map of the preliminary farming area and the non-farming area.
8. The method for extracting a culture area based on time series filtering according to claim 1, wherein the method for generating the extraction result of the culture area in the step 5 is as follows: converting the extraction results of the multiple time-phase culture areas into a binary mask; Using a length of successive frame sequences Sampling a sliding time window of (a) to form a window tensor ; To window tensor Each pixel position of (2) is counted in an accumulation way to obtain an accumulation matrix For reflecting the stability of the local pixels; Accumulated count matrix based on minimum density threshold D Binarizing to generate local stability mask To eliminate sporadic noise; locally stabilizing the mask Building an accumulation mask on the basis , The representation is a union set of the numbers, The accumulated mask is the last moment; And (3) through the combination of the sliding time window and the local stability judgment, the suppression of instantaneous noise and local fluctuation in the remote sensing sequence is realized, and the culture area mask with consistent time sequence is obtained.

Description

Cultivation area extraction method based on time series filtering Technical Field The invention relates to the technical field of cultivation area extraction, in particular to a cultivation area extraction method based on time sequence filtering. Background The optical remote sensing technology can acquire the reflection or emission spectrum information of the same target in different wave bands at the same time in a photographing or scanning mode. Multispectral images typically comprise 4-15 discrete bands (e.g., visible light, near infrared, short wave infrared, etc.), which have a relatively wide band width (e.g., 50-200 nanometers), and a relatively low spectral resolution, and are suitable for large-scale surface monitoring tasks such as land utilization classification, vegetation coverage analysis, water surface target identification, water quality monitoring, etc. Mussel culture is one of important marine economic industries in coastal areas of China, and automatic monitoring of mussel culture is of great significance to marine pasture management and ecological environment protection. The traditional monitoring method mainly relies on manual field investigation, and has the advantages of high cost, low efficiency and poor timeliness. The existing remote sensing monitoring method is mostly based on threshold segmentation or traditional machine learning algorithm, but still has the following limitations: 1. spectrum confusion problem, namely, similar spectrum characteristics of the culture facilities and natural reefs, islands and the like, are easy to cause misclassification, and the automatic identification precision is further improved by combining space characteristic information; 2. space-time variation interference, namely, the spectral characteristics of different culture areas can dynamically change along with seasons and years, so that the traditional method is difficult to be applied, and an effective space-time analysis method is required to be introduced to reduce the influence of space-time interference on the identification result; 3. Most methods do not fully consider multi-phase data optimization results, and new models need to be developed to efficiently use limited timing information. Disclosure of Invention The invention aims to provide a culture area extraction method based on time series filtering, aiming at the defects in the prior art. In order to achieve the above object, the present invention provides a method for extracting a culture area based on time series filtering, comprising the steps of: Step 1, acquiring multi-temporal optical remote sensing image data of a target area, and performing spatial registration and area clipping pretreatment on the multi-temporal optical remote sensing image data to construct a time sequence image dataset of the target area; Step2, labeling the culture area and the non-culture area based on the time sequence image dataset, and extracting space domain information from each labeling object according to a preset space scale to form a sample dataset containing space context relation; Step 3, constructing a culture area extraction model, and training the culture area extraction model by utilizing a sample data set to enable the model to learn the joint representation of the spatial structural characteristics and the time evolution characteristics of the culture area; Step 4, inputting the time sequence image data set into a training-completed culture area extraction model, and obtaining a culture area extraction result of each time phase; Step 5, carrying out joint analysis on the extraction results of the culture areas in the same spatial position in a plurality of time phases, and constructing a time sequence stability judging mechanism so as to inhibit transient noise and enhance the consistency of the culture areas in the time dimension, so as to obtain the extraction results of the culture areas with consistent time sequences; And 6, carrying out contrast analysis on spatial distribution of the culture areas in different time phases, extracting boundaries of the expansion area and the extinction area of the culture areas, and calculating the spatial area and the variation quantity of the culture areas in each time phase to realize quantitative characterization of the aerodynamic evolution of the culture areas. Further, the culture area extraction model comprises: The tensor construction module is used for converting the multi-temporal optical image sample acquired by the research area into a multi-dimensional tensor form suitable for model input so as to represent the spatial structure characteristics and the time evolution characteristics of the sample; The feature extraction module is used for carrying out feature coding on the tensor form sample and extracting intermediate feature representation capable of representing spatial structural features and time evolution features of the culture area; The first central perception module is used