Remote sensing precise monitoring method of rice bacterial blight based on deep learning spatial spectrum super-resolution

Abstract

The present invention relates to a remote sensing fine monitoring method for rice bacterial blight based on deep learning spatial-spectral super-resolution, comprising: acquiring image data and preprocessing the image data; enhancing the spatial resolution and spectral resolution of the image data by combining a spatial-spectral resolution enhancement model, and acquiring a spatial-spectral resolution enhancement map; analyzing the spatial-spectral characteristics of the image to achieve rice disease classification. The beneficial effects of the present invention are: the present invention significantly improves the accuracy of disease detection by effectively fusing spectral and texture information, and can accurately identify subtle disease features; and the present invention utilizes a learnable texture extraction structure to enable the model to adapt to different data features, thereby enhancing the generalization ability of the model on diversified data; the present invention also reduces unnecessary calculations through an optimized network structure and attention mechanism, thereby improving the efficiency of disease detection.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing monitoring of rice diseases, and more specifically, to a remote sensing fine monitoring method of rice bacterial blight based on deep learning spatial spectrum super-resolution. Background Art

[0002] Rice is an important food crop in my country, and stable rice production is crucial to ensuring national food security. However, worldwide, rice production reductions due to diseases are as high as 30% of total production. Bacterial leaf blight is one of the most serious bacterial diseases in rice production in my country, and is known as the "three major diseases" of rice together with rice blast and sheath blight. The "Forecast of Trends in Crop Diseases and Pests" released by the National Agricultural Technology Center shows that rice diseases will be more serious in 2023, with an area of ​​​​diseases reaching 360 million mu. Among them, the "three major diseases" of the key rice prevention and control targets in Zhejiang Province are all listed. Rice is the main food crop in Ningbo. The Ningbo Statistical Yearbook in 2022 shows that rice production accounts for 77.2% of the total grain output. However, the high temperature and high humidity climate conditions have led to frequent occurrences of rice bacterial leaf blight, which has become a prominent bottleneck restricting the development of agriculture in our city. The 2022 Ningbo Grain Production Situation Special Report pointed out that in recent years, the area of ​​rice bacterial blight in our city has gradually increased, especially during the heading period and after heading, which seriously affects the high and stable yield of rice and the quality of rice, endangering the food security of the city and even the province. Therefore, carrying out large-scale early and precise monitoring of rice bacterial blight is of great significance for the precise prevention and control of rice bacterial blight in Ningbo and ensuring food security.

[0003] Traditional early and accurate disease monitoring relies on field surveys, but it is difficult to achieve accurate disease monitoring in large areas. The rapid development of remote sensing technology has made large-area disease monitoring possible. Existing remote sensing disease monitoring can be divided into three scales: ground, aerial, and satellite. Most ground disease monitoring research focuses on leaf-scale disease identification and classification, and the technology is relatively mature. Specifically, a spectrometer is first used to obtain hyperspectral data of diseased leaves, and then spectral index, random forest, convolutional neural network, wavelet transform and other methods are used to extract disease features. Machine learning methods such as VGGnet, long short-term memory recursive neural network (LSTM), artificial neural network, and support vector machine

[18] are widely used in disease identification and classification. The accuracy of ground disease monitoring results is generally high, but it can only achieve disease monitoring at a point scale, and cannot achieve accurate disease monitoring at a regional scale. UAVs can obtain high-resolution remote sensing data in a small area, so they are gradually applied to disease monitoring and have achieved relatively satisfactory monitoring accuracy. However, large-scale disease monitoring relies on satellite remote sensing, and existing satellite disease monitoring focuses on the construction of disease indices [8], or disease identification based on commonly used spectral indices. Early and accurate disease monitoring based on satellite data is difficult to identify due to its limited spatial / spectral resolution. Therefore, how to use deep learning algorithms to achieve spatial-spectral super-resolution of remote sensing data, improve the spatial resolution and spectral resolution of remote sensing data, obtain remote sensing data with high-altitude-spectral resolution, and then realize precise remote sensing monitoring of rice bacterial blight is the main goal of this study. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies in the prior art and propose a remote sensing fine monitoring method for rice bacterial blight based on deep learning spatial spectrum super-resolution.

[0005] First, a remote sensing fine monitoring method for rice bacterial blight based on deep learning spatial spectrum super-resolution is provided, including:

[0006] S1, acquiring image data, and preprocessing the image data;

[0007] S2, enhancing the spatial resolution and spectral resolution of the image data by combining the spatial-spectral resolution enhancement model to obtain a spatial-spectral resolution enhancement map;

[0008] S3. Analyze the spatial-spectral characteristics of the image to achieve rice disease classification.

[0009] Preferably, in S2, the spatial resolution of the enhanced image data is expressed as:

[0010] SR1=Conv+ReLU(up(RDB(Conv+ReLU(LR1))))

[0011] LR2′=Conv+ReLU(LR2)

[0012] Q=VGG(up(LR2′))

[0013] K=V=VGG(AdaIN(up x2 (LR2)+noise1,w1))

[0014] SR2=SoftAttention(Q,Hard Attention(K,V))

[0015] Among them, LR1 is low-resolution image 1, that is, the bands of Sentinel 2 with a resolution of 10m (bands B2, B3, B4, B8). LR2 is low-resolution image 2, that is, the Sentinel bands with a resolution of 20m (B5, B6, B7, B8a, B11, B12). Q is the result after VGG feature extraction, K and V are the results after style transfer and VGG feature extraction, RDB stands for dense residual block, Conv+ReLU is convolution and ReLU operation, SR1 is the upscaled feature obtained by LR1, SR2 is the upscaled feature obtained by LR2; VGG is the VGG feature extraction structure, AdaIN is the AdaIN algorithm, Soft Attention is the soft attention structure, and HardAttention is the hard attention structure.

[0016] Preferably, in S2, the spectral resolution of the image data is enhanced, which is expressed as:

[0017] SSR=DFF(SR1,SR2)

[0018] SR=STANet(LR1,SSR)

[0019] Among them, DFF is a feature extractor based on dense residual structure, SSR represents the fusion result of two upscaled features, STANet is a spectral super-resolution network; SR is the final spatial-spectral resolution enhancement map.

[0020] Preferably, S3 includes:

[0021] S301, performing preliminary feature extraction on the input spatial-spectral resolution enhancement map and the low-resolution image through 3D convolution, group convolution and point convolution;

[0022] S302, using the attention mechanism to capture the subtle relationship between different positions in the sequence and introduce interactive information;

[0023] S303, mapping the extracted features to the scores of the output categories through a linear layer to achieve accurate classification and extraction of diseased rice and non-disease-infected rice.

[0024] As a preference, it also includes:

[0025] S4. Test on public datasets and data collected in the field.

[0026] Preferably, in S1, the image data includes image data acquired by Sentinel-2 and VENμS satellites at the same time.

[0027] In a second aspect, a remote sensing fine monitoring system for rice bacterial blight based on deep learning spatial spectrum super-resolution is provided, which is used to perform any remote sensing fine monitoring method for rice bacterial blight described in the first aspect, including:

[0028] An acquisition module, used for acquiring image data and preprocessing the image data;

[0029] An enhancement module is used to enhance the spatial resolution and spectral resolution of image data through a joint spatial-spectral resolution enhancement model to obtain a spatial-spectral resolution enhancement map;

[0030] The classification module is used to analyze the spatial-spectral characteristics of images and classify rice diseases.

[0031] In a third aspect, a computer storage medium is provided, wherein a computer program is stored in the computer storage medium; when the computer program is executed on a computer, the computer executes the method for remote sensing fine monitoring of rice bacterial blight as described in any one of the first aspects.

[0032] The beneficial effects of the present invention are:

[0033] 1. Improve the accuracy of disease detection: The present invention significantly improves the accuracy of disease detection by effectively fusing spectral and texture information, and can accurately identify subtle disease features.

[0034] 2. Enhance model adaptability: The present invention utilizes a learnable texture extraction structure to enable the model to adapt to different data features, thereby enhancing the generalization ability of the model on diverse data.

[0035] 3. Improve detection efficiency: The present invention reduces unnecessary calculations and improves the efficiency of disease detection through optimized network structure and attention mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of a spatial-spectral resolution enhancement algorithm;

[0037] Figure 2 It is a flow chart of a classification algorithm based on super-resolution;

[0038] Figure 3A schematic diagram showing the comparison of super-resolution results in the dataset;

[0039] Figure 4 This is a schematic diagram of the super-resolution results of real Sentinel-2 images;

[0040] Figure 5 This is a schematic diagram of disease identification results. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for ordinary persons in the art, without departing from the principle of the present invention, the present invention can also be modified in some ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0042] Embodiment 1:

[0043] The emergence of super-resolution rice disease recognition technology provides a new solution to the limitations of traditional methods. By improving the spatial spectral resolution of images, this technology can effectively enhance image details, making tiny disease features more visible, thereby improving the accuracy and reliability of disease recognition. This not only helps farmers to detect and control diseases in a timely manner, protect crop growth, and improve agricultural production efficiency, but also provides rich data resources for related scientific research and promotes the development of agricultural science and technology. Therefore, super-resolution rice disease recognition technology has important practical value and scientific significance.

[0044] In this regard, Example 1 of the present application provides a remote sensing fine monitoring method for rice bacterial blight based on deep learning spatial spectrum super-resolution, through a series of carefully designed steps, from data processing to feature enhancement, to classification and recognition, and finally system deployment and verification, each step is aimed at improving the performance and application scope of the model. Through this structured processing flow, not only the accuracy and efficiency of the model are improved, but also the stability and generalization ability of the model in practical applications are ensured.

[0045] The method provided in Example 1 of the present application comprises:

[0046] S1. Data collection and processing of relevant data sets. Through meticulous data cleaning and standardization, the consistency and availability of data are ensured, laying a solid foundation for model training.

[0047] This step improves the quality of the data and provides high-quality input for model training, thereby improving the learning efficiency and final performance of the model.

[0048] S2. A deep network is used to enhance features, especially the details of land cover types. Through adversarial training, the model can generate more detailed and realistic land cover images.

[0049] By enhancing the details of the image, the model's recognition accuracy of surface cover types is improved, especially in areas with complex terrain and diverse cover types.

[0050] S3 uses convolutional neural networks and attention networks in deep learning for feature extraction, and combines the advantages of super-resolution and multi-feature fusion to achieve accurate classification of surface cover types.

[0051] In S3, the embodiment of the present application adopts a method that combines super-resolution and multi-feature classification, which not only improves the model's ability to capture spatial features, but also can effectively process remote sensing data, thereby improving the model's classification performance in a dynamic environment.

[0052] S4. We conducted field tests on a large scale to verify the generalization ability and practical application effect of the model. We fine-tuned the model based on the test results to ensure that it can achieve optimal performance in different environments.

[0053] In S4, field testing and fine-tuning were performed to ensure the stability and adaptability of the model, enabling it to maintain efficient and accurate monitoring capabilities under different geographical and environmental conditions.

[0054] Embodiment 2:

[0055] Based on Example 1, Example 2 of the present application provides a more specific method for remote sensing fine monitoring of rice bacterial blight based on deep learning spatial spectrum super-resolution, including:

[0056] S1. Acquire image data and pre-process the image data.

[0057] In S1, the image data includes image data acquired by Sentine1-2 and VENμS satellites at the same time. Through appropriate preprocessing, the data quality is improved to prepare for subsequent processing.

[0058] S2. By combining a spatial-spectral resolution enhancement model, the spatial resolution and spectral resolution of the image data are enhanced to obtain a spatial-spectral resolution enhancement map.

[0059] In S2, for the improvement of spatial resolution, this application mainly focuses on how to process low spatial resolution (LR) images into high spatial resolution (SR) images. The core technology of spatial resolution enhancement is implemented through residual dense network (RDN). RDN restores high-frequency detail information by learning the nonlinear mapping relationship between LR images and their corresponding HR images. Figure 1 As shown, the formula is expressed as:

[0060] SR1=Conv+ReLU(up(RDB(Conv+ReLU(LR1))))

[0061] LR2′=Conv+ReLU(LR2)

[0062] Q=VGG(up(LR2′))

[0063] K=V=VGG(AdaIN(up x2 (LR2)+noise1,w1))

[0064] SR2=SoftAttention(Q,Hard Attention(K,V))

[0065] Among them, LR1 is low-resolution image 1, that is, the resolution is 10m sentinel band. LR2 is low-resolution image 2, that is, the resolution is 20m sentinel band. Q is the result after VGG feature extraction, K and V are the results after style transfer and VGG feature extraction, RDB stands for dense residual block, Conv+ReLU is convolution and ReLU operation, SR1 is the upscaled feature obtained by LR1, SR2 is the upscaled feature obtained by LR2; VGG is the VGG feature extraction structure, AdaIN is the AdaIN algorithm, S0ft Attention is the soft attention structure, and Hard Attention is the hard attention structure.

[0066] Preferably, in S2, the spectral resolution of the image data is enhanced, which is expressed as:

[0067] And enhance the spectral resolution of the data. To improve the spectral resolution, this application uses the VGG and Attention models in style transfer. First, the VGG model is used to extract image features, and then the attention mechanism is used to enhance the model's ability to capture spectral details. The formula is expressed as:

[0068] SSR=DFF(SR1,SR2)

[0069] SR=STANet(LR1,SSR)

[0070] Among them, DFF is a feature extractor based on dense residual structure, SSR represents the fusion result of two upscaled features, STANet is a spectral super-resolution network; SR is the final spatial-spectral resolution enhancement map.

[0071] S3. Analyze the spatial-spectral characteristics of the image to achieve rice disease classification.

[0072] S3 includes:

[0073] S301, performing preliminary feature extraction on the input spatial-spectral resolution enhancement map and the low-resolution image through 3D convolution, group convolution and point convolution;

[0074] S302, using the attention mechanism to capture the subtle relationship between different positions in the sequence, and introducing interactive information to further improve the performance of the model;

[0075] S303, mapping the extracted features to the scores of the output categories through a linear layer to achieve accurate classification.

[0076] It can be seen that the entire process of S3 is divided into three stages, each of which processes the data to varying degrees, and finally outputs the final classification decision by summarizing the classification results of each stage. This structured processing flow not only enhances the model's ability to capture features, but also ensures the model's efficiency and accuracy in processing complex data through detailed stage division.

[0077] S4. Test on public datasets and data collected in the field.

[0078] Specifically, the super-resolution disease monitoring system was applied in the actual scene of Ningbo through field verification and parameter adjustment, the performance of the system was verified in the field, and the model parameters were adjusted according to the requirements of the actual application to ensure the best monitoring effect. Finally, the integration of software and hardware was achieved through system integration and deployment, and the algorithm model was deployed to the end-user equipment, including satellite data receiving equipment and computer analysis system, to ensure the stable operation of the system and the user-friendly operation interface.

[0079] This application has conducted in-depth verification on the non-training areas of the SEN2VENμS dataset and found that the model of this application has achieved resolution improvement in all 8 bands, especially in the b5, b6, b7 and b8A bands, where the enhancement effect is more prominent. By showing the reconstruction result graphs, quantitative indicators and RGB images of each band, this application can clearly observe the significant enhancement effect of the data, as shown in the attached figure. Figure 3 These verification results not only prove the effectiveness of the model in this application, but also provide strong support for subsequent research.

[0080] In addition, this application was tested in a real environment, and the Sentinel-2 images of Ningbo were selected as experimental data, which were downloaded and processed through Google Earth Engine. The experimental results show that the model of this application can effectively improve the resolution in various scenarios, such as Figure 4 This experimental result further confirms the efficiency and feasibility of the proposed model in processing actual data, laying a solid foundation for subsequent classification and recognition tasks.

[0081] By comparing the performance of different models, as shown in Table 1, the SR+CTransformer model proposed in this application has improved the accuracy of disease recognition and non-disease recognition, which is 1.64% and 0.63% higher than the CTransformer model respectively. Overall, the comprehensive classification performance of SR+CTransformer has also been enhanced, and the overall accuracy (OA) has increased by 0.48% to 95.7%. The average classification accuracy (ACA) has increased by 0.85% to 93.49%, and the Kappa coefficient has also increased by 1.24% to 88.41%. These data show that the SR+CTransformer model has superior performance when dealing with disease recognition tasks compared with other models and the original CTransformer model.

[0082] Table 1 Comparison of disease recognition model performance

[0083]

[0084] In order to ensure the generalization ability of the model and avoid overfitting, this application only used 5% of the training sample data to train the model and performed disease identification on a large area of ​​scenes. According to the map of the study area and the urban profile of Ningbo City, this application learned that the southeast of the study area is a mountainous area, while the northeast is an urban area. The date of the observed satellite image is October 11, 2022, which is the late growth stage of rice, and diseases occur more frequently. Figure 5 The disease identification map displayed clearly reveals that the area is widely affected by the disease, especially in the lower right part of the area, where the disease damage is particularly serious.

[0085] It should be noted that the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be described in detail in this application.

[0086] Embodiment 3:

[0087] On the basis of Examples 1 and 2, Example 3 of the present application provides a remote sensing fine monitoring system for rice bacterial blight based on deep learning spatial spectrum super-resolution, including:

[0088] An acquisition module, used for acquiring image data and preprocessing the image data;

[0089] An enhancement module is used to enhance the spatial resolution and spectral resolution of image data through a joint spatial-spectral resolution enhancement model to obtain a spatial-spectral resolution enhancement map;

[0090] The classification module is used to analyze the spatial-spectral characteristics of images and classify rice diseases.

[0091] Specifically, the system provided in this embodiment is a system corresponding to the method provided in Embodiments 1 and 2. Therefore, the parts in this embodiment that are the same or similar to Embodiments 1 and 2 can be referenced to each other and will not be repeated in this application.

[0092] Embodiment 4:

[0093] On the basis of Examples 1 and 2, Example 4 of the present application further provides a disease detection system based on spatial spectrum resolution enhancement, including:

[0094] Improved deep learning network structure: The improved convolutional network is used through the downscaling module to extract deep features of different scales; the spectral aggregation module and the spectral attention block are used to spectrally aggregate the deep features; the texture and spectral extraction module is used to perform texture extraction and spectral cross-attention operations on the aggregated features to extract texture detail information and spectral information; finally, the feature fusion of texture and spectral information is realized through the fusion module.

[0095] Learnable convolutional Gabor filter and gradient Sobel filter structure: This paper innovatively designs a texture extraction structure that combines convolutional Gabor filter and gradient Sobel filter, so that the model can efficiently extract texture features of low-resolution images and effectively integrate them with spectral features.

[0096] Spectral and texture attention fusion mechanism: This disease detection system reconstructs feature information more accurately through the spectral and texture attention fusion mechanism, thereby achieving high-precision disease detection. This mechanism combines the cross-correlation of spectral dimensions with the texture attention structure to simultaneously enhance the spatial and spectral details of the reconstructed features.

Claims

1. A remote sensing fine monitoring method for rice bacterial blight based on deep learning spatial spectrum super-resolution, characterized in that: include: S1, acquiring image data, and preprocessing the image data; S2, enhancing the spatial resolution and spectral resolution of the image data by combining the spatial-spectral resolution enhancement model to obtain a spatial-spectral resolution enhancement map; In S2, the spatial resolution of the enhanced image data is expressed as: SR1=Conv+ReLU(up(RDB(Conv+ReLU(LR1)))) LR2′=Conv+ReLU(LR2) Q=VGG(up(LR2′)) K=V=VGG(AdaIN(up x2 (LR2)+noise1,w1)) SR2=SoftAttention(Q,Hard Attention(K,V)) Among them, LR1 is a low-resolution image with a resolution of 10m Sentinel band; LR2 is a low-resolution image with a resolution of 20m Sentinel band; Q is the result after VGG feature extraction, K and V are the results after style transfer and VGG feature extraction, RDB stands for dense residual block, Conv+ReLU is convolution and ReLU operation, SR1 is the upscaled feature obtained by LR1, and SR2 is the upscaled feature obtained by LR2; VGG is the VGG feature extraction structure, AdaIN is the AdaIN algorithm, Soft Attention is the soft attention structure, and Hard Attention is the hard attention structure; S3: Analyze the spatial-spectral characteristics of images to classify rice diseases; S3 includes: S301, performing preliminary feature extraction on the input spatial-spectral resolution enhancement map and the low-resolution image through 3D convolution, group convolution and point convolution; S302, using the attention mechanism to capture the subtle relationship between different positions in the sequence and introduce interactive information; S303, mapping the extracted features to the scores of the output categories through a linear layer to achieve accurate classification.

2. The method for remote sensing fine monitoring of rice bacterial blight based on deep learning spatial spectrum super-resolution according to claim 1, characterized in that: In S2, the spectral resolution of the image data is enhanced, expressed as: SSR=DFF(SR1,SR2) SR=STANet(LR1,SSR) Among them, DFF is a feature extractor based on dense residual structure, SSR represents the fusion result of two upscaled features, STANet is a spectral super-resolution network; SR is the final spatial-spectral resolution enhancement map.

3. The method for remote sensing fine monitoring of rice bacterial blight based on deep learning spatial spectrum super-resolution according to claim 2, characterized in that: Also includes: S4. Test on public datasets and data collected in the field.

4. A remote sensing fine monitoring system for rice bacterial blight based on deep learning spatial spectrum super-resolution, characterized by: The method for remote sensing fine monitoring of rice bacterial blight according to any one of claims 1 to 3 comprises: An acquisition module, used for acquiring image data and preprocessing the image data; An enhancement module is used to enhance the spatial resolution and spectral resolution of image data through a joint spatial-spectral resolution enhancement model to obtain a spatial-spectral resolution enhancement map; The classification module is used to analyze the spatial-spectral characteristics of images and classify rice diseases.

5. A computer storage medium, characterized in that: The computer storage medium stores a computer program; when the computer program is executed on a computer, the computer executes the method for remote sensing fine monitoring of rice bacterial blight according to any one of claims 1 to 3.

Images