CN-121982552-A - Soybean high-temperature-resistant grading method based on vegetation index priori and self-supervision learning

CN121982552ACN 121982552 ACN121982552 ACN 121982552ACN-121982552-A

Abstract

The invention provides a soybean high temperature resistant grading method based on vegetation index priori and self-supervision learning, which comprises the steps of utilizing SVM to fit mapping relation between vegetation indexes and high temperature resistant grades, introducing SHAP interpretability analysis to calculate contribution degree weight of each vegetation index, generating vegetation index priori significance map reflecting physiological importance of blades, mapping the vegetation priori significance map to an RGB image coordinate system to construct a non-uniform mask strategy, guiding a mask self-encoder to reconstruct a high physiological significance region preferentially, extracting high temperature resistant texture features of robustness, extracting features according to RGB images and multispectral images, constructing feature vectors, fusing the feature vectors by utilizing a cross-modal cross-attention module, calculating errors by utilizing a weighted gating loss module based on contrast loss and cross-entropy loss, and updating front-end network parameters by adopting back propagation. The invention effectively reduces the manual marking cost and realizes the identification of the high-temperature resistant phenotype of the large-scale soybean germplasm.

Inventors

JIN XIU
YANG WEIZHI
WANG XIAOBO
LI JIAJIA
MIAO LONG
DENG YOUHUI
YAO JIAHUI

Assignees

安徽农业大学

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The soybean high temperature resistant grading method based on vegetation index priori and self-supervision learning is characterized by comprising the following steps: obtaining a visible light RGB image and a multispectral image of a soybean canopy; Calculating a vegetation index and generating a soybean high-temperature-resistant grade label by combining with an agronomic character investigation result in a maturity stage; using SVM to fit the mapping relation between the vegetation indexes and the high temperature resistant level, introducing SHAP interpretability analysis to calculate the contribution degree weight of each vegetation index, and generating a vegetation index priori significance map reflecting the physiological significance of the leaves; Mapping the vegetation index priori significance map to an RGB image coordinate system to construct a non-uniform mask strategy, guiding a mask to reconstruct a high physiological significance region preferentially from an encoder, and extracting robust high temperature resistant texture features; And extracting features according to the RGB images and the multispectral images, constructing feature vectors, fusing the feature vectors by using a cross-modal cross-attention module, calculating errors by using a weighted gating loss module based on contrast loss and cross-entropy loss, updating front-end network parameters by using back propagation, and outputting the soybean high-temperature resistant grade by using an optimization model.
2. The soybean high temperature resistant grading method based on vegetation index priori and self-supervised learning according to claim 1, wherein the wave bands of the multispectral image at least comprise a green light wave band, a red edge wave band and a near infrared wave band.
3. The soybean high temperature resistant grading method based on vegetation index priors and self-supervised learning according to claim 2, wherein in the step of calculating the vegetation index, spectral features are extracted based on the preprocessed images, and the calculated vegetation index at least comprises a normalized vegetation index, a red-edge normalized vegetation index, an optimized soil adjustment vegetation index and a green normalized vegetation index.
4. The method of claim 3, wherein the step of generating a soybean high temperature resistant grade label comprises: calculating a high temperature resistance coefficient which is the ratio of the property value of the experimental group to the property value of the control group; Determining weights using principal component analysis; Calculating a high-temperature-resistant comprehensive evaluation value through membership function standardized analysis, and determining a high-temperature grade label according to the evaluation value, wherein the evaluation value H is expressed as: ; Wherein, the The membership function value of the j-th index is represented, The weight coefficient indicating the j-th index, and n indicating the total number of indexes.
5. The method of claim 4, wherein the step of generating a vegetation index prior significance map reflecting physiological significance of the leaves comprises: constructing a Support Vector Machine (SVM) classification model, and fitting a mapping relation between a vegetation index and a high-temperature-resistant level; Calculating contribution values of each vegetation index to different high temperature resistant categories by using SHAP algorithm, and obtaining global weight by weighted summation Expressed as: ; In the formula, C represents a class of high temperature resistant grades; A weighting coefficient representing category c; The calculation formula of SHAP, which represents the absolute contribution value of the ith vegetation index to class c, is: ; In the formula, Representing the model prediction result of the feature i under the input sample x F represents a set of all features, S represents any subset of features i not included in F, Representing output values when the model uses only features in subset S to participate in the prediction; representing the output values when the model uses only feature subset S plus feature i to participate in the prediction; Using the calculated global SHAP weight to obtain F vegetation index feature patterns Performing weighted linear combination at pixel level for each pixel position in the image Multiplying the numerical values of the vegetation index feature graphs with corresponding weights and summing to obtain original response values, wherein the original response values are expressed as follows: ; In the formula, Representing global weights; representing a response value; Representing the numerical value of the ith vegetation index feature map at the pixel position (x, y), N representing the total number of feature maps; Nonlinear transformation is carried out by adopting Sigmoid activation function to generate physical priori significance map 。
6. The method for classifying soybean high temperature resistance based on vegetation index priors and self-supervised learning according to claim 5, wherein the step of mapping vegetation index priors significance maps to an RGB image coordinate system to construct a non-uniform masking strategy comprises: dividing RGB image into non-overlapping fixed-size image blocks, calculating average saliency value of saliency map in each image block area, and constructing mask probability matrix ; According to the mask probability matrix Performing Bernoulli sampling on all image blocks to generate a mask view; Wherein: Mask probability matrix In (i) th image block retention probability Calculated by linear mapping, expressed as: ; In the formula, An average saliency value within an i-th image block region; reserving a probability for the set basis for maintaining the minimum information amount of the background area; Is a scaling factor; Maximum retention probability upper limit; determining a high-saliency region and a low-saliency region according to the average saliency value, and respectively endowing the high-saliency region and the low-saliency region with corresponding retention probabilities; The MAE decoder receives the potential features of the encoder and the mask flags, reconstructs the original RGB image, and reconstructs the loss function Mean square error of reconstructed image and original image on visible image block by minimizing While reconstructing the critical leaf structure, deducing the overall growth state of the soybean according to the context information.
7. The soybean high temperature resistant grading method based on vegetation index priors and self-supervised learning according to claim 6, the method is characterized in that the step of extracting the characteristics and constructing the characteristic vector comprises the following steps: Loading a pre-trained encoder ViT by using an RGB coding path as a backbone network, taking an RGB image block as input, and extracting a structural feature vector containing fine textures and morphological structures; The multispectral coding path adopts a lightweight convolutional neural network ResNet-18 as backbone, takes a multispectral vegetation index graph as input, and extracts physiological feature vectors containing biochemical information such as chlorophyll, moisture and the like; And (3) constructing a contrast loss function introducing vegetation index priori to align structural features and physiological features, and mapping structural feature vectors and physiological feature vectors with different dimensions to a public potential space with the same dimension.
8. The method for classifying soybean high temperature resistance based on vegetation index priors and self-supervised learning as recited in claim 7, wherein in introducing a contrast loss function of vegetation index priors, for the first in a lot Samples, structural features And multispectral physiological features Regarding the characteristics of other samples in the batch as negative samples as positive sample pairs, calculating contrast loss Expressed as: ; In the formula, Representing the cosine similarity calculation, The temperature coefficient is given, and N is the batch size; And (3) with Is the structural characteristics and physiological characteristics of the same soybean plant; is the physiological characteristic of other soybean samples in the batch.
9. The soybean high temperature resistant classification method based on vegetation index priors and self-supervised learning according to claim 8, wherein in the step of fusing feature vectors by using a cross-modal cross-attention module, information complementation is performed by using a cross-attention mechanism, RGB structural features are used as queries, multispectral physiological features are used as keys and values, and fusion features are calculated Expressed as: ; In which RGB structural features are used As a query, with multispectral physiological features As keys and values, where T represents the transpose of the matrix; The Softmax is a nonlinear activation function and is used for normalizing the scaled characteristic dot product result into attention weight probability distribution between 0 and 1; finally, the original structure information is superimposed through residual connection And outputting the fusion characteristic.
10. The method for classifying soybean high temperature resistance based on vegetation index priors and self-supervised learning according to claim 9, the step of calculating an error by a weighted gating loss module based on the contrast loss and the cross entropy loss and updating the front end network parameters by using back propagation comprises the following steps: Constructing a gating network; fusing multiple views features As input of a pre-constructed gating network, the gating network dynamically calculates self-adaptive weight coefficients of a feature alignment task and a classification task through multi-layer perceptron MLP and Softmax functions And , wherein, ; Based on the self-adaptive weight coefficient, the characteristic alignment contrast loss Cross entropy loss with classification tasks Weighted summation is carried out to obtain the total gating loss Expressed as: ; performing a back propagation mechanism using the calculated total gating weight loss Simultaneously updating the extracted encoder parameters and ResNet-18 parameters, and performing end-to-end model joint optimization.

Description

Soybean high-temperature-resistant grading method based on vegetation index priori and self-supervision learning Technical Field The invention belongs to the technical field of artificial intelligence, and particularly relates to a soybean high-temperature-resistant classification method based on vegetation index priori and self-supervision learning. Background The development of high-tolerance Wen Biaoxing identification in large-scale soybean germplasm resource groups, and the screening of excellent-resistance germplasm are key links of genetic improvement and new variety breeding. The traditional field heat resistance identification mainly depends on a manual visual inspection method, is time-consuming and labor-consuming, has low efficiency, is greatly influenced by subjective experience of observers, and is difficult to meet the requirements of modern breeding on high-throughput and objective phenotype analysis. In recent years, unmanned Aerial Vehicle (UAV) low-altitude remote sensing technology gradually replaces traditional manual investigation by virtue of high space-time resolution, non-destructive and flexible maneuvering, and becomes a mainstream means for high-throughput monitoring of soybean field phenotypes. In the existing crop resistance evaluation research based on unmanned aerial vehicle images, the inversion method based on a single physical vegetation index is most widely applied. The method generally utilizes multispectral or thermal infrared sensor data to calculate single indexes such as normalized vegetation index (NDVI) or Crop Water Stress Index (CWSI) and the like to quantitatively represent the growth state of crops. Although these physical indices have clear agronomic interpretations, there are significant limitations in early diagnosis of soybean high temperature stress. Soybeans often exhibit complex morphological-physiological coupling reactions such as leaf flipping (LEAF FLIPPING) to reduce the light receiving area when subjected to high temperature stresses, while causing a rise in canopy temperature with closed pores. Under a complex field environment, a single spectrum or temperature index is extremely easy to be interfered by soil background, illumination shadow and canopy structure heterogeneity, the micro phenotype change is difficult to be captured sharply, the robustness of resistance identification is insufficient, and misjudgment is extremely easy to generate. In order to overcome the limitation of physical indexes, a data driving method based on deep learning is gradually becoming a research hotspot. With the development of Convolutional Neural Networks (CNN) and Vision Transformer (ViT), significant progress has been made in phenotyping using RGB image training end-to-end models. However, the supervised learning paradigm relies heavily on massive amounts of high quality annotation data, whereas expert resistance annotation in the agricultural field is extremely costly and difficult to obtain on a large scale, resulting in models facing serious overfitting risks. In order to solve the difficult problem of data annotation, self-supervision learning techniques such as a mask self-encoder (MAE) and the like are introduced into an agricultural scene, and aim to pretrain a model by utilizing massive unlabeled data. Unfortunately, existing MAE methods generally employ a random masking strategy, i.e., randomly occluding image regions to allow the model to reconstruct. In a soybean field remote sensing image, crop foreground including core high temperature resistant texture features is intermixed with soil background including a large amount of invalid information. The random masking strategy can not distinguish the two, so that a large amount of calculation force is wasted on the model to reconstruct useless background textures, even wrong background noise characteristics are learned, and the extraction and characterization capability of the model on the key high-temperature resistant phenotype is seriously weakened. In addition, in the aspect of multi-mode data utilization, most of the existing researches adopt a simple channel splicing or pixel-level weighting mode to fuse the visible light RGB image and the multi-spectrum image. Because the RGB image has high spatial resolution but lacks spectral information, the multispectral image has lower spatial resolution but is rich in physiological and biochemical information, a gap exists between the two, and deep feature semantic alignment cannot be realized by simple splicing, so that high-frequency texture structure information and low-frequency physiological decay information are difficult to effectively complement, and the fine analysis and comprehensive evaluation of the antagonistic phenotype are limited. Disclosure of Invention The embodiment of the invention aims to provide a soybean high-temperature resistant grading method based on vegetation index priori and self-supervision learning, aiming at solving the techn