CN-121999373-A - Multi-time sequence remote sensing data and machine learning fusion field crop yield estimation method

CN121999373ACN 121999373 ACN121999373 ACN 121999373ACN-121999373-A

Abstract

The application provides a field crop yield estimation method integrating multi-time-series remote sensing data and machine learning, which relates to the technical field of artificial intelligence and comprises the steps of obtaining a cross-year comprehensive data set, wherein the cross-year comprehensive data set comprises time-series remote sensing image data of a plurality of historical years, weather data of corresponding years and agricultural management records; the method comprises the steps of extracting robust features from a annual comprehensive data set to form a robust feature set, forming a training sample set based on the robust feature set and historical actual yield data, generating a resistance sample simulating climate disturbance and agricultural management change, training a crop yield estimation model by adopting the training sample set and the resistance sample to obtain a target crop yield estimation model, preprocessing time sequence remote sensing image data and air image data of a target year, and inputting the time sequence remote sensing image data and the air image data into the target crop yield estimation model to obtain a crop yield estimation value. The application improves the yield estimation precision by the field crop yield estimation method.

Inventors

ZHANG TINGTING
HONG YONG
WANG MI
GUO PEIPEI
CAI YULIN

Assignees

湖北珞珈实验室
武汉大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A field crop yield estimation method integrating multi-time-sequence remote sensing data and machine learning is characterized by comprising the following steps: Acquiring a cross-year comprehensive data set, wherein the cross-year comprehensive data set comprises time sequence remote sensing image data of a plurality of historical years, meteorological data of corresponding years and agricultural management records; Extracting robust features from the annual comprehensive data set to form a robust feature set, and forming a training sample set based on the robust feature set and historical actual yield data, wherein the historical actual yield data is used as a supervision tag; generating an antagonism sample for simulating climate disturbance and agricultural management change, and training a crop yield estimation model by adopting the training sample set and the antagonism sample to obtain a target crop yield estimation model; And preprocessing the time sequence remote sensing image data and the meteorological data of the target year, and inputting the preprocessed time sequence remote sensing image data and the meteorological data into the target crop yield estimation model to obtain a crop yield estimation value.
2. The method for estimating field crop yield based on multi-temporal remote sensing data and machine learning fusion of claim 1, further comprising: Acquiring temperature deviation, precipitation deviation and extreme event frequency according to the meteorological data of the target year and the historical average meteorological data; And carrying out weighted correction on the crop yield estimated value by adopting a dynamic correction function based on a meteorological deviation response mechanism based on the temperature deviation, the precipitation deviation and the extreme event frequency to obtain a crop yield correction value.
3. The method for estimating field crop yield by combining multi-temporal remote sensing data with machine learning according to claim 1, wherein the temporal remote sensing image data comprises a sequence of vegetation indexes derived from multi-spectral and hyperspectral remote sensing images, the meteorological data comprises daily average temperature, cumulative precipitation, sunshine hours and extreme weather event records, and the agricultural management records comprise irrigation time, fertilization amount and crop rotation information.
4. The method for estimating yield of field crops based on multi-temporal remote sensing data and machine learning fusion of claim 1, wherein said extracting robust features from said trans-annual synthetic dataset comprises: Extracting a plurality of candidate features from the annual comprehensive dataset, wherein the candidate features include remote sensing derived features and climate derived features; and calculating variation coefficients and stability scores of each candidate feature between years based on a preset evaluation algorithm, and selecting the candidate features with variation coefficients lower than a preset variation threshold and stability scores higher than a stability threshold as robust features.
5. The method for estimating yield of field crops by fusing multi-temporal remote sensing data and machine learning according to claim 4, wherein the calculating the variation coefficient and the stability score of each candidate feature between years based on a preset estimation algorithm comprises: Inputting standard deviation and mean values of the annual feature values in the candidate features into a preset variation calculation formula to obtain annual variation coefficients of the candidate features; And inputting the annual variation coefficient corresponding to each candidate feature, the standard deviation of the mean value difference of the features in the adjacent years and the mean value of the feature values in the years into an annual stability evaluation function to obtain the stability score of each candidate feature.
6. The method for estimating yield of field crops by fusion of multi-temporal remote sensing data and machine learning according to claim 1, wherein after preprocessing the temporal remote sensing image data and the meteorological data of a target year and inputting the processed temporal remote sensing image data and the meteorological data into the target crop yield estimation model, the method further comprises: determining an annual mean value and a standard deviation of similar features in a history stage according to input features input into the target crop yield estimation model; Determining the reliability level of the input features at each time step by adopting a preset data quality monitoring function based on the annual average value and standard deviation of the similar features at the history stage; and when the reliability level of the input features in the target time steps is smaller than a preset threshold value, dynamically interpolating and correcting the features of the target time steps in the input features.
7. The method for estimating field crop yield by fusing multi-temporal remote sensing data with machine learning of any one of claims 1 to 6, further comprising, prior to acquiring the trans-annual synthetic dataset: Acquiring multi-source time sequence remote sensing image data; Performing radiation calibration and space resampling on the multi-source time sequence remote sensing image data to unify the resolution and the band response range, and extracting a key vegetation index sequence from the multi-spectrum and hyperspectral images; And aiming at abnormal points and null value sections in the key vegetation index sequence, adopting a dynamic interpolation algorithm based on time sequence similarity, and realizing continuity recovery through weighted smoothing in a local time window.
8. A field crop yield estimation device integrating multi-time-series remote sensing data and machine learning, comprising: The acquisition module is used for acquiring a trans-annual comprehensive data set, wherein the trans-annual comprehensive data set comprises time sequence remote sensing image data of a plurality of historical years, meteorological data of corresponding years and agricultural management records; The extraction module is used for extracting robust features from the annual comprehensive data set to form a robust feature set, and forming a training sample set based on the robust feature set and historical actual yield data, wherein the historical actual yield data is used as a supervision tag; the training module is used for generating an antagonistic sample for simulating climate disturbance and agricultural management change, and training the crop yield estimation model by adopting the training sample set and the antagonistic sample to obtain a target crop yield estimation model; And the estimation module is used for preprocessing the time sequence remote sensing image data and the meteorological data of the target year and inputting the time sequence remote sensing image data and the meteorological data into the target crop yield estimation model to obtain a crop yield estimation value.
9. An electronic device comprising a processor and a memory, the memory having a stored computer program, wherein the computer program when executed by the processor implements a field crop yield estimation method of fusion of multi-temporal telemetry data and machine learning as claimed in any one of claims 1 to 7.
10. A non-transitory computer storage medium having a computer program stored thereon, wherein the computer program when executed by a processor implements a field crop yield estimation method of fusion of multi-temporal remote sensing data and machine learning as claimed in any one of claims 1 to 7.

Description

Multi-time sequence remote sensing data and machine learning fusion field crop yield estimation method Technical Field The application relates to the technical field of artificial intelligence, in particular to a field crop yield estimation method integrating multi-time-sequence remote sensing data and machine learning. Background In the technical fields of agricultural remote sensing and intelligent yield estimation, a crop yield estimation method is one of the core research directions of accurate agriculture and climate intelligent agriculture. The growth state and the yield level of crops can be quantitatively estimated through comprehensive analysis of multi-temporal remote sensing images, meteorological observation data and ground agricultural information, and data support is provided for farmland management, resource allocation and grain safety monitoring. In recent years, with the development of high space-time resolution improvement and machine learning algorithms of remote sensing satellites, crop yield estimation is gradually changed from an empirical model and a growth mechanism model to a data-driven intelligent fusion model, and especially, the combination of multi-time-sequence remote sensing data and deep learning enables dynamic tracking and multi-factor interactive modeling of the whole crop growth process to be possible. Currently, traditional crop yield estimation methods have limitations that are poorly adapted to climate and management changes. Because the single growth model is used as a core, parameter setting needs to rely on priori knowledge and specific area calibration, so that the estimation accuracy of the model is reduced under the condition of annual climate fluctuation. Disclosure of Invention In view of the above, the application provides a field crop yield estimation method integrating multi-time-sequence remote sensing data and machine learning. In a first aspect, the present application provides a method for estimating yield of field crops by fusing multi-time-series remote sensing data with machine learning, comprising: Acquiring a cross-year comprehensive data set, wherein the cross-year comprehensive data set comprises time sequence remote sensing image data of a plurality of historical years, meteorological data of corresponding years and agricultural management records; Extracting robust features from the annual comprehensive data set to form a robust feature set, and forming a training sample set based on the robust feature set and historical actual yield data, wherein the historical actual yield data is used as a supervision tag; generating an antagonism sample for simulating climate disturbance and agricultural management change, and training a crop yield estimation model by adopting the training sample set and the antagonism sample to obtain a target crop yield estimation model; And preprocessing the time sequence remote sensing image data and the meteorological data of the target year, and inputting the preprocessed time sequence remote sensing image data and the meteorological data into the target crop yield estimation model to obtain a crop yield estimation value. In an embodiment, the method further comprises: Acquiring temperature deviation, precipitation deviation and extreme event frequency according to the meteorological data of the target year and the historical average meteorological data; Acquiring temperature deviation, precipitation deviation and extreme event frequency according to the meteorological data of the target year and the historical average meteorological data; And carrying out weighted correction on the crop yield estimated value by adopting a dynamic correction function based on a meteorological deviation response mechanism based on the temperature deviation, the precipitation deviation and the extreme event frequency to obtain a crop yield correction value. In one embodiment, the time-series remote sensing image data comprises a sequence of vegetation indexes derived from multispectral and hyperspectral remote sensing images, the meteorological data comprises daily average temperature, accumulated precipitation, sunshine hours and extreme weather event records, and the agricultural management records comprise irrigation time, fertilization amount and crop rotation information. In an embodiment, the extracting robust features from the annual comprehensive dataset comprises: Extracting a plurality of candidate features from the annual comprehensive dataset, wherein the candidate features include remote sensing derived features and climate derived features; and calculating variation coefficients and stability scores of each candidate feature between years based on a preset evaluation algorithm, and selecting the candidate features with variation coefficients lower than a preset variation threshold and stability scores higher than a stability threshold as robust features. In an embodiment, the calculating the variation coefficient and the s