Rice and shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectrum unmixing

Abstract

The invention discloses a rice and shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectrum unmixing, and belongs to the technical field of rice and shrimp field extraction. The method comprises the following basic steps: firstly, embedding a channel-space collaborative attention mechanism related to the rice and shrimp field and spectral fidelity and morphological constraint terms into a super-resolution diffusion model, constructing a self-adaptive spectral index and spectral unmixing analysis for a super-divided image, and finally, extracting the rice and shrimp field in a multi-rule combined manner. The method solves the problems that the boundary of a medium-low resolution remote sensing image is fuzzy, the spectrum-space coupling modeling of the rice and shrimp field is insufficient in the existing algorithm, the conventional water body index and the vegetation index are misrecognized in the water area and the rice field staggered area in the rice and shrimp field, and the combined utilization of multiple features is lacked.

Description

Technical Field

[0001] The present invention relates to the technical field of rice-crayfish field extraction, and specifically discloses a remote sensing extraction method for rice-crayfish fields based on diffusion model super-resolution reconstruction and adaptive spectral unmixing. Background Art

[0002] Rice-crayfish fields are a typical agricultural ecosystem widely distributed in the middle and lower reaches of the Yangtze River in China. Their unique "circular ditch + internal paddy field" spatial structure provides a symbiotic environment for rice cultivation and crayfish farming. However, due to the complex spectral characteristics in the rice-crayfish field area, vegetation-water mixed pixels are widespread and spatially scattered. Traditional remote sensing image extraction methods face the following technical bottlenecks:

[0003] Currently widely used medium-resolution remote sensing images are difficult to capture the small paddy fields and circular ditch structures in rice-crayfish fields, resulting in blurred extraction results and lost details at the boundaries. Although super-resolution reconstruction technology can improve the spatial resolution of images, existing methods are mainly designed for general scenarios and not for super-resolution design for specific scenarios. Moreover, rice-crayfish fields are mainly composed of "circular ditches + internal paddy fields", which are related to water body index and vegetation index respectively. However, the NDWI / NDVI index is not optimized for the spectrum of rice-crayfish fields, leading to a high misclassification rate in the area where water and vegetation intersect. Traditional spectral unmixing ignores the spatial distribution law of rice-crayfish fields, mostly based on single spectral index threshold segmentation, lacking the joint utilization of multi-features such as spectrum, geometry, and space.

[0004] In summary, the existing technology is difficult to meet the requirements of high-precision extraction of rice-crayfish fields. There is an urgent need for an innovative method that combines super-resolution reconstruction, adaptive spectral index, and multi-rule joint extraction to solve problems such as blurred boundaries of low-resolution images, spectral confusion, and low decomposition accuracy of mixed pixels. The present invention is proposed precisely under this background. Summary of the Invention

[0005] The main purpose of the present invention is to embed a channel-space collaborative attention mechanism related to rice-crayfish fields, as well as spectral fidelity and morphological constraint terms, into a super-resolution diffusion model, construct an adaptive spectral index and spectral unmixing analysis for the super-resolved image, and finally jointly extract rice-crayfish fields by multiple rules to improve the recognition accuracy.

[0006] To achieve the above object, the present invention adopts the following technical means. A remote sensing extraction method for rice-crayfish fields based on diffusion model super-resolution reconstruction and adaptive spectral unmixing specifically includes the following technical innovations:

[0007] I. Super-resolution reconstruction of remote sensing images of rice-crayfish fields based on diffusion model

[0008] Step 1: Image acquisition and processing. The image data input into the model includes the original remote sensing images and high-resolution images of the paddy fields in the rice-crayfish field study area during the planting months. The original remote sensing image data uses the multi-spectral image data provided by the Sentinel-2 satellite. However, only select the key bands related to rice-crayfish field identification: blue, green, red, and near-infrared bands, with a spatial resolution of 10m. At the same time, obtain the high-resolution auxiliary remote sensing data of Jilin-1, with a spatial resolution of 0.8m. Based on the high-resolution remote sensing image dataset, perform blurring processing, generate low-resolution images that match the Sentinel-2 resolution through downsampling, and use a Gaussian kernel filter to blur the downsampled images to simulate the remote sensing image degradation process, and construct a high-low resolution image pair dataset.

[0009] Step 2: Diffusion model configuration and training. The diffusion model is a mathematical model, a generative model driven by non-equilibrium thermodynamics, which can be specifically divided into a forward process and a reverse process. The forward diffusion process realizes noise injection through parameterized Gaussian transfer, and its mathematical expression is:

[0010]

[0011] where x t is the image at time t, a t is the noise scheduling parameter, and ε is the standard normal distribution noise.

[0012] The reverse diffusion process uses a multi-channel deep denoising network p θ (x t-1 |x t ) to learn the latent distribution. The model weights are initialized based on transfer learning. A channel-spatial collaborative attention mechanism is designed in the residual module, and its calculation process is:

[0013] F ca = F ⊙ Softmax(MLP(GAP(F)))

[0014] where the channel weights are generated through global average pooling (GAP) and multi-layer perceptron (MLP), adaptively adjusting the importance of each band, so that the weights of the green and near-infrared bands are increased in the peripheral water area of the rice-crayfish field, and the weights of the red and near-infrared bands are increased in the central paddy field area, thereby enhancing the spectral feature expression ability of water bodies and vegetation.

[0015]

[0016] where M ditch is the pre-generated ditch mask, which combines the unique "ring ditch + internal paddy field" spatial structure characteristics of the rice-crayfish field to enhance the attention weight, enabling the model to pay more attention to the narrow strip structure of the ditch and the block distribution of the paddy field. The features after channel weighting (F ca) and the pre-generated trench mask (M ditch ) splicing, generate a spatial attention map through convolution, and finally combine the channel and spatial information.

[0017] Step 3: Physical constraint reconstruction and performance evaluation. Finally, in the reverse process, a multi-scale perceptual loss function is used to jointly optimize the pixel-level error, VGG feature similarity, and adversarial loss. Spectral fidelity constraints and morphological constraints are embedded, specifically:

[0018]

[0019] B3, B4, and B8 are green, red, and near-infrared bands, respectively. The sum of the L2 norm squares of these key bands is calculated to represent the difference between the high-resolution image HR and the super-resolution image SR, ensuring that the key bands of the reconstructed image still maintain the original spectral characteristics as much as possible at high resolution; TV is the abbreviation of Total Variation, where λ is a weight parameter used to control the importance of the total variation loss in the overall loss function. It imposes anisotropic TV constraints on the rice-shrimp field position (i, j) in the image in the horizontal and vertical directions, promoting the smoothness of the rice-shrimp field by minimizing the gradient while retaining the edge information.

[0020] A dynamic learning rate scheduler is used to control the training process. When the verification indicators converge, the optimization direction is switched. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators are used to quantify the reconstruction accuracy. The spatial-spectral consistency is evaluated by expert visual interpretation. Based on the evaluation results, data enhancement and model iterative optimization are performed on the high-frequency error areas.

[0021] 2. Remote sensing extraction of rice-shrimp fields based on adaptive spectral unmixing

[0022] Step 4: Construct adaptive spectral indexes NDWI_A and NDVI_A based on the characteristics of rice-shrimp fields. The green light and near-infrared bands related to the water body outside the rice-shrimp field are the same as the calculation bands of the traditional water body index NDWI, while the red light and near-infrared bands related to the central rice field are the same as the calculation bands of the traditional vegetation index NDVI. Therefore, the adaptive spectral indexes NDWI_A and NDVI_A are constructed. The specific formula is:

[0023]

[0024] The weight β is determined by fitting the reflectance curve of the rice-shrimp field samples (typical value β = 1.1), and the weight γ is determined by maximizing the separation between the rice-shrimp field and ordinary farmland (typical value γ = 1.3).

[0025] Step 5: Spectral unmixing analysis. Optimization of endmember selection: Pixels with NDVI_A > 0.7 and uniform texture are extracted as vegetation endmembers; pixels with NDWI_A > 0.6 and adjacent to vegetation endmembers (distance less than 50 m) are extracted as water body endmembers to avoid introducing non-rice-crayfish-pond water bodies. For the mixed pixels of water bodies and paddy fields in the rice-crayfish pond, the spectral characteristics of water bodies, paddy fields, and other ground objects are decomposed through spectral unmixing technology to obtain the abundance ratios of each endmember in the mixed pixels. Assume that the spectrum of a pixel is a linear combination of multiple endmembers, that is:

[0026]

[0027] where ρ P is the reflectance of pixel p, a k is the abundance of the k-th endmember, s k is the reflectance of the k-th endmember, and σ is the residual. At the same time, the mixed pixel decomposition model satisfies two constraint conditions: (1) Sum constraint: that is ensures that the total abundance of each pixel is 1. (2) Non-negativity constraint: that is a k ≥ 0. The water body abundance and vegetation abundance are calculated through the spectral unmixing model.

[0028] Step 6: Joint extraction of rice-crayfish ponds using multiple rules. Define the spectral-geometric-spatial three-element rules for rice-crayfish ponds: Based on the vegetation abundance and water body abundance obtained from spectral unmixing, combined with spectral, geometric, and spatial characteristics, accurately identify the rice-crayfish pond area. First, based on the spectral rule, pixels with vegetation abundance ≥ 0.6 and water body abundance ≥ 0.3 are selected as candidate areas; secondly, through geometric rule filtering, to exclude the interference of small water bodies, select an area ≥ 300 m 2 , match the shape of artificial fields with an aspect ratio ≤ 3, and ensure the regularity of the field shape with a circularity ≥ 0.5; finally, introduce the spatial rule, requiring the distance between the center of the field and the adjacent water body area to be ≤ 20 m to exclude isolated water bodies or non-shrimp-pond areas. The formula for calculating circularity is:

[0029]

[0030] The beneficial effects of the present invention are:

[0031] (1) A super-resolution reconstruction method based on a diffusion model is proposed. A channel-spatial collaborative attention module is embedded in the model for the spectral characteristic index of the rice-crayfish pond and the unique "circular ditch + internal paddy field" spatial structure characteristic, significantly improving the spatial resolution of the rice-crayfish pond area in the original Sentinel-2 remote sensing image.

[0032] (2) The adaptive spectral indices NDWI_A and NDVI_A are constructed, and by adjusting the band weights, the spectral separability between the vegetation and water bodies in the rice-crayfish fields is enhanced. Compared with the traditional NDWI and NDVI, the spectral confusion problem in the mixed pixel areas is effectively solved.

[0033] (3) By combining the multi-rule extraction strategies of spectrum, geometry, and space, through the threshold constraints of vegetation abundance and water body abundance, and combining geometric rules such as the area, length-width ratio, and circularity of the field plots, as well as the spatial distribution relationship between the field plots and the adjacent water bodies, the high-precision extraction of the rice-crayfish fields is achieved. Brief Description of the Drawings

[0034] Appendix Figure 1 : Flowchart of the remote sensing extraction method for rice-crayfish fields based on diffusion model super-resolution reconstruction and adaptive spectral unmixing Figure 2 : Super-resolution reconstruction algorithm based on the diffusion model Detailed Embodiments

[0035] For the convenience of those of ordinary skill in the art to understand and implement the present invention, and to make the objectives, contents, and advantages of the present invention clearer, the following further describes the method in detail with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the method, and are not used to limit the method.

[0036] The overall flowchart of a remote sensing extraction method for rice-crayfish fields based on diffusion model super-resolution reconstruction and adaptive spectral unmixing given by this method is as Figure 1 shown.

[0037] This method includes the following steps:

[0038] Step 1: Image acquisition and processing. The image data input into the model includes the original remote sensing image and the high-resolution image of the rice-crayfish field study area during the rice planting month. The original remote sensing image data is the multi-spectral image data provided by the Sentinel-2 satellite. The data includes four bands: red, green, blue, and near-infrared, with a spatial resolution of 10m. At the same time, the high-resolution auxiliary remote sensing data of Jilin-1 is obtained, with a spatial resolution of 0.8m, as the prior information source; the preprocessing includes radiometric correction, atmospheric correction, geometric correction, cropping, mosaicking, etc. Using the rice planting months from July to September, the rice-crayfish fields with obvious spatial structural features of "circular ditches + internal rice fields" are obtained. Based on the high-resolution remote sensing image dataset, blurring processing is performed. By downsampling, a low-resolution image matching the resolution of Sentinel-2 is generated, and a Gaussian kernel filter with a kernel size of 3x3 and a standard deviation of 1.0 is used to blur the downsampled image to simulate the degradation process of the remote sensing image and construct a high-low resolution image pair dataset.

[0039] Step 2: Diffusion Model Configuration and Training. The diffusion model is a mathematical model, a generative model driven by non-equilibrium thermodynamics, which can be specifically divided into a forward process and a reverse process. The forward diffusion process realizes noise injection through parameterized Gaussian transfer, and its mathematical representation is:

[0040]

[0041] where \(x\) t is the image at time \(t\), \(a\) t is the noise scheduling parameter, and \(\epsilon\) is the standard normal distribution noise.

[0042] The reverse diffusion process uses a multi-channel deep denoising network \(p\) θ (\(x\) t-1 | \(x\) t ) to learn the latent distribution. The model weights are initialized based on transfer learning, and a channel-spatial collaborative attention mechanism is designed in the residual module. Its calculation process is:

[0043] \(F\) ca = \(F\odot Softmax(MLP(GAP(F)))\)

[0044] where the channel weights are generated through global average pooling (GAP) and a multi-layer perceptron (MLP) to adaptively adjust the importance of each band, so that the weights of the green and near-infrared bands are increased in the peripheral water area of the rice-crayfish field, and the weights of the red and near-infrared bands are increased in the central paddy field area, thereby enhancing the spectral feature expression ability of water and vegetation.

[0045]

[0046] where \(M\) ditch is the pre-generated ditch mask. Combining the unique "circular ditch + internal paddy field" spatial structure characteristics of the rice-crayfish field enhances the attention weight, making the model pay more attention to the narrow strip structure of the ditch and the block distribution of the paddy field. The channel-weighted feature (\(F\) ca ) is concatenated with the pre-generated ditch mask (\(M\) ditch ), and a spatial attention map is generated through convolution. Finally, the information of channels and space is combined

[0047] Step 3: Physical Constraint Reconstruction and Performance Evaluation. Finally, in the reverse diffusion process, a multi-scale perceptual loss function is used to jointly optimize the pixel-level error, VGG feature similarity, and adversarial loss. The final loss combination:

[0048] \(L=\lambda_1L\) pixel +\(\lambda_2L\) VGG +\(\lambda_3L\) GAN

[0049] where the content loss \(L\) pixel constrains the pixel-level consistency, and the feature loss \(L\)VGG Extract high-level semantic features through pre-trained VGG network, and the adversarial loss L GAN Then enhance the visual authenticity through the discriminative network.

[0050] And embed the spectral fidelity constraint term and the morphological constraint term, specifically:

[0051]

[0052]

[0053] Among them, B3, B4, and B8 are the green light, red light, and near-infrared bands respectively. Calculate the sum of the squares of the L2 norms of these key bands, which represents the difference between the high-resolution image HR and the super-resolution image SR, and ensure that the reconstructed image still maintains the original spectral characteristics as much as possible in the key bands at high resolution.

[0054] TV is the abbreviation of Total Variation. Among them, λ is a weight parameter used to control the importance of the total variation loss in the overall loss function. Apply the anisotropic TV constraint to the position (i, j) of the rice-crayfish field in the horizontal and vertical directions of the image, and promote the smoothness of the rice-crayfish field by minimizing the gradient while retaining the edge information.

[0055] The training optimizer selects the Adam optimizer, sets the initial learning rate to 0.0001, and uses the cosine annealing strategy for learning rate scheduling to prevent the model from falling into local optima. The training batch size is 16, and the number of training epochs is 100. After completing the model training, input the image patches of each band into the model separately at a size of 256x256 pixels, generate a 0.5-meter resolution image through the inverse diffusion process, and then splice all the image patches into a complete high-resolution image. The peak signal-to-noise ratio PSNR of the reconstructed image reaches 33.5dB, and the structural similarity SSIM is 0.92. Combine expert visual interpretation to evaluate the spatial-spectral consistency, and perform data augmentation and model iterative optimization on the high-frequency error regions based on the evaluation results.

[0056] Step 4: Construct the adaptive spectral indices NDWI_A and NDVI_A based on the characteristics of the rice-crayfish field. The green light and near-infrared bands related to the water body outside the rice-crayfish field are the same as the calculation bands of the traditional water body index NDWI, while the red light and near-infrared bands related to the central paddy field are the same as the calculation bands of the traditional vegetation index NDVI. Therefore, construct the adaptive spectral indices NDWI_A and NDVI_A, and the specific formulas are:

[0057]

[0058] Among them, the weight β is determined by fitting the reflectance curve of the rice-crayfish field samples, and the weight γ is determined by maximizing the separation between the rice-crayfish field and ordinary farmland.

[0059] Select 100 sample points in the rice-crayfish field, and measure their vegetation coverage (70%-90%) and water body ratio (20%-40%). Conduct multiple regression on the reflectance of the sample points, and fit β = 1.1 and γ = 1.3.

[0060] Step 5: Spectral unmixing analysis. Optimization of endmember selection: Pixels with NDVI_A > 0.7 and uniform texture are extracted as paddy field endmembers; pixels with NDWI_A > 0.6 and adjacent to vegetation endmembers (distance less than 50m) are extracted as water body endmembers to avoid introducing non-rice-crayfish field water bodies. For the mixed pixels of water bodies and paddy fields in the rice-crayfish field, decompose the spectral characteristics of water bodies, paddy fields and other ground objects through spectral unmixing technology to obtain the abundance ratios of each endmember in the mixed pixels. Assume that the spectrum of a pixel is a linear combination of multiple endmembers, that is:

[0061]

[0062] where ρ P is the reflectance of pixel p, a k is the abundance of the k-th endmember, s k is the reflectance of the k-th endmember, and σ is the residual. At the same time, the mixed pixel decomposition model satisfies two constraint conditions: (1) Sum constraint: that is ensuring that the total abundance of each pixel is 1. (2) Non-negativity constraint: that is a k ≥ 0. Calculate the water body abundance and vegetation abundance through the spectral unmixing model.

[0063] Step 6: Extract rice-crayfish fields by combining multiple rules. Define the spectral-geometry-space three-element rules for rice-crayfish fields: Based on the vegetation abundance and water body abundance obtained from spectral unmixing, combined with spectral, geometric and spatial features, accurately identify the rice-crayfish field area. First, based on the spectral rule, select pixels with vegetation abundance ≥ 0.6 and water body abundance ≥ 0.3 as candidate areas; secondly, filter through geometric rules, select an area ≥ 300m 2 to exclude the interference of small water bodies, select an aspect ratio ≤ 3 by matching the shape of artificial fields, and ensure the regularity of the field shape by selecting a circularity ≥ 0.5; finally, introduce a spatial rule, requiring the distance between the center of the field and the adjacent water body area ≤ 20m to exclude isolated water bodies or non-shrimp field areas. The calculation formula for circularity is:

[0064]

Claims

1. A rice-shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectral unmixing, characterized in that: The following steps are involved:

1. Super-resolution reconstruction of rice-shrimp field remote sensing images based on diffusion model Step 1: The image data input to the model include the original remote sensing images and high-resolution images of the rice fields in the rice-shrimp field study area during the planting months. The original remote sensing image data uses the multispectral image data provided by the Sentinel-2 satellite with a spatial resolution of 10m. At the same time, the Jilin-1 high-resolution auxiliary remote sensing data with a spatial resolution of 0.8m is obtained. Fuzzification is performed based on the high-resolution remote sensing image dataset. A low-resolution image matching the Sentinel-2 resolution is generated by downsampling. The downsampled image is blurred using a Gaussian kernel filter to simulate the remote sensing image degradation process and construct a high-low resolution image pair dataset. Step 2: Diffusion model configuration and training. The diffusion model is a mathematical model, which is a generative model driven by non-equilibrium thermodynamics. It can be divided into a forward process and a reverse process. The forward diffusion process is based on parameterized Gaussian transfer to achieve noise injection, and the reverse diffusion process uses a multi-channel deep denoising network p θ (x t-1 |x t ) learns the potential distribution, and the model weights are initialized based on transfer learning; and in particular, in the residual module, a channel-space collaborative attention mechanism is constructed, which focuses more on the narrow strip structure of the ditch in the rice-shrimp field and the block distribution of the rice field to improve the resolution and optimize the image reconstruction quality; Step 3: Physical constraint reconstruction and performance evaluation. Finally, in the inverse diffusion process, a multi-scale perceptual loss function is used to jointly optimize the pixel-level error, VGG feature similarity, and adversarial loss. Spectral fidelity constraints and morphological constraints are embedded to ensure that the reconstructed image still maintains the original spectral characteristics as much as possible in the key bands at high resolution, and various anisotropic TV constraints are imposed along the horizontal and vertical directions of the ridge. A dynamic learning rate scheduler is used to control the training process. When the verification index converges, the optimization direction is switched. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators are used to quantify the reconstruction accuracy. Based on the evaluation results, data enhancement and model iterative optimization are performed on the high-frequency error area.

2. Remote sensing extraction of rice-shrimp fields based on adaptive spectral unmixing Step 4: Based on the characteristics of rice-shrimp fields, the adaptive spectral indexes NDWI_A and NDVI_A are constructed; the green light and near-infrared bands related to the water body outside the rice-shrimp fields are the same as the calculation bands of the traditional water body index NDWI, while the red light and near-infrared bands related to the central rice fields are the same as the calculation bands of the traditional vegetation index NDVI, so the adaptive spectral indexes NDWI_A and NDVI_A are constructed; Step 5: Spectral unmixing analysis and endmember selection optimization: NDVI_A>0.7 and uniform texture areas are extracted as vegetation endmembers; NDWI_A>0.6 and pixels adjacent to vegetation endmembers (less than 50m away) are extracted as water endmembers to avoid the introduction of non-rice-shrimp field water bodies; for the mixed pixels of water bodies and rice field vegetation in rice-shrimp fields, the spectral characteristics of water bodies, rice fields and other landforms are decomposed by spectral unmixing technology to obtain the abundance ratio of each endmember in the mixed pixels; Step 6: Combine multiple rules to extract rice-shrimp fields and define the spectrum-geometry-space three-element rules of rice-shrimp fields: Based on the vegetation abundance and water body abundance obtained by spectral unmixing, the spectral, geometric and spatial features are combined to accurately identify the rice-shrimp field area; first, based on the spectral rules, pixels with vegetation abundance ≥ 0.6 and water body abundance ≥ 0.3 are selected as candidate areas; secondly, the geometric rules are further filtered to select pixels with an area of ​​≥ 300m to exclude the interference of small water bodies. 2 , match the shape of artificial fields by selecting a length-to-width ratio ≤ 3, and ensure the regularity of the field shape by selecting a circularity ≥ 0.5; finally, introduce spatial rules, requiring that the distance between the center of the field and the adjacent water area be ≤ 20m to exclude isolated water bodies or non-shrimp field areas.

2. The rice-shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectral unmixing according to claim 1 is characterized in that: In step 1, only four key bands, namely blue light, green light, red light and near infrared, are selected for rice-shrimp field identification because they can not only accurately capture the spectral characteristics of the central rice field and the surrounding water body, but are also the core bands of the commonly used water body index NDWI and vegetation index NDVI.

3. The rice-shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectral unmixing according to claim 1 is characterized in that: In step 2, the channel-space collaborative attention mechanism includes: Fca=F⊙Softmax(MLP(GAP(F))) The global average pooling (GAP) and multi-layer perceptron (MLP) are used to generate channel weights and adaptively adjust the importance of each band, so that the weights of green light and near-infrared bands in the water area outside the rice-shrimp field are increased, and the weights of red light and near-infrared bands in the central rice field area are increased, thereby enhancing the spectral feature expression ability of water bodies and vegetation. Among them, M ditch For the pre-generated ditch mask, the attention weight is enhanced by combining the unique spatial structure features of "circular ditch + internal rice field" of rice-shrimp fields, so that the model pays more attention to the narrow strip structure of the ditch and the block distribution of the rice field, and the channel weighted features (F ca ) and the pre-generated trench mask (M ditch ) splicing, generate a spatial attention map through convolution, and finally combine the channel and spatial information.

4. The rice-shrimp field remote sensing extraction method based on diffusion model super-resolution reconstruction and adaptive spectral unmixing according to claim 1 is characterized in that: In step 3, the embedded spectral fidelity constraint term and morphological constraint term are specifically: B3, B4, and B8 are green, red, and near-infrared bands, respectively. The sum of the L2 norm squares of these key bands is calculated to represent the difference between the high-resolution image HR and the super-resolution image SR, ensuring that the key bands of the reconstructed image still maintain the original spectral characteristics as much as possible at high resolution; TV is the abbreviation of total variation, where λ is a weight parameter used to control the importance of total variation loss in the overall loss function. It imposes anisotropic TV constraints on the rice-shrimp field position (i, j) in the image in the horizontal and vertical directions, promoting the smoothness of the rice-shrimp field by minimizing the gradient while retaining the edge information.

Images