High-salt-tolerance rice planting management system based on image recognition

Abstract

The invention relates to the field of intelligent agriculture, in particular to a high-salt-tolerance rice planting management system based on image recognition, which comprises a data acquisition module, an image processing module, a growth condition detection module and an analysis management module, through the computer vision technology, close planting degree guidance and saline-alkaline degree control are intelligently provided for high-salt-tolerance rice planting, the management efficiency is improved, and the rice yield and quality are improved; according to the method, a Transformer coding method based on a window mechanism and global optimization is adopted, the recognition accuracy of the growth state of a rice plant is improved, the accuracy of analysis of the relationship between the growth condition of the rice and the salt and alkali content and the close planting degree of the rice is improved, and accurate guidance is provided for increasing the yield and quality of the rice; according to the method, a UNet network based on Mama block optimization is adopted, the influence of a field background on rice image feature extraction is removed, and the accuracy of rice growth condition-saline and alkaline content-rice close planting degree relation analysis is improved.

Description

Technical Field

[0001] The present invention belongs to the field of smart agriculture, and specifically relates to a high salt-tolerant rice planting management system based on image recognition. Background Art

[0002] The planting management of high salt-tolerant rice has important practical significance, which can help solve problems such as the utilization of saline-alkali land resources, food security, and environmental protection, and promote the development and sustainable development of modern agriculture. Through scientific planting management, the yield and quality of high salt-tolerant rice can be improved, and at the same time, contributions can be made to agricultural production and environmental protection. The existing planting management of high salt-tolerant rice is usually manually controlled, with low efficiency and it is difficult to effectively improve the yield and quality of high salt-tolerant rice. Summary of the Invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a high salt-tolerant rice planting management system based on image recognition. Aiming at the problems that the existing planting management method of high salt-tolerant rice is inefficient and difficult to effectively improve the yield and quality of high salt-tolerant rice, the present invention uses computer vision technology to reasonably analyze the relationship among rice growth status - salinity content - rice planting density, and intelligently provides planting density guidance and salinity control for high salt-tolerant rice planting, improving management efficiency and increasing rice yield and quality; the present invention creatively adopts a Transformer encoding method based on a window mechanism and global optimization to identify the growth status of rice plants, better captures the spatial characteristics of the growth status of rice plants through spatial-channel optimization, introduces a local window mechanism to selectively focus on local regions in the image, extracts key features of the growth status of rice plants, improves the recognition accuracy of the growth state of rice plants, and further increases the accuracy of the analysis of the relationship among rice growth status - salinity content - rice planting density, and further provides more accurate guidance for increasing rice yield and quality; the present invention creatively adopts a UNet network optimized based on Mamba blocks. On the basis of image feature extraction and classification, the spatial dimension is gradually reduced by the encoder of the cascade module, effectively capturing local features and global feature representations. At the same time, a convolutional Mamba hybrid block is introduced to mix convolutional operations with three-way Mamba to achieve hierarchical feature extraction with multiple receptive fields, and then the precise segmentation of rice images, removing the influence of the field background on the extraction of rice image features, and further increasing the accuracy of the analysis of the relationship among rice growth status - salinity content - rice planting density.

[0004] The high salt-tolerant rice planting management system based on image recognition provided by the present invention includes a data acquisition module, an image processing module, a growth status detection module, and an analysis and management module;

[0005] The data acquisition module uses a laser scanner to collect 3D images of the growth status of highly salt-tolerant rice in fields with different saline-alkali contents and collect the corresponding rice planting density in the fields;

[0006] The image processing module denoises the collected 3D images of the rice growth status and uses a UNet network optimized based on the three-way Mamba block to perform 3D image segmentation to remove the field background and obtain rice images;

[0007] The growth status detection module uses a Transformer encoding method based on a window mechanism and global optimization to classify and identify the growth status of rice images in fields with different saline-alkali contents and different planting densities, and obtains the relationship between rice growth status - saline-alkali content - rice planting density;

[0008] The analysis and management module collects the saline-alkali content of the field where highly salt-tolerant rice needs to be planted currently, and according to the relationship between rice growth status - saline-alkali content - rice planting density, changes the rice planting density or adjusts the saline-alkali content of the irrigation water to obtain the optimal rice growth status.

[0009] Furthermore, in the image processing module, a UNet network optimized based on the three-way Mamba block performs 3D image segmentation, which specifically includes the following steps:

[0010] Step Q1: Feature downsampling, perform D convolution operation on the denoised 3D image of the rice growth status to obtain downsampled features;

[0011] Step Q2: Convolution Mamba hybrid block optimization, optimize the downsampled features based on the convolution Mamba hybrid block to obtain optimized downsampled features, which specifically includes the following steps:

[0012] Step Q21: Spatial dimensionality reduction, use a convolution kernel to perform spatial dimensionality reduction and coarse-grained feature extraction on the downsampled features to obtain reduced-dimensional features;

[0013] Step Q22: Local context relationship optimization, further refine the reduced-dimensional features through convolution to capture local context relationships and obtain locally optimized features;

[0014] Step Q23: Three-way Mamba block processing, use three-way Mamba, that is, TOM Mamba, to perform three-directional feature dependence calculations on the locally optimized features to obtain three-way fusion features: ;

[0015] In the formula, , and respectively represent the forward, backward, and three-dimensional slice-to-slice feature modeling of the locally optimized features, Represents Mamba block processing, Represents local optimization features, Represents three-way fusion features;

[0016] Step Q24: Residual mechanism optimization, perform transposed convolution on the three-way fusion features and fuse them with the downsampled features based on residual connection, and perform three-way Mamba block processing on the fusion result again to obtain optimized downsampled features;

[0017] Step Q3: Multilayer downsampling, repeat steps Q1 - Q2 three times, perform downsampling on the optimized downsampled features repeatedly, and finally obtain four layers of optimized downsampled features;

[0018] Step Q4: Feature upsampling, perform multi-layer perceptron processing on the four layers of optimized downsampled features respectively, and perform upsampling of the features to obtain segmentation features;

[0019] Step Q5: Feature activation output, perform activation output on the segmentation features to obtain a segmentation mask, and obtain a segmentation map, that is, a rice image, according to the segmentation mask and the 3D image of the rice growth condition.

[0020] Furthermore, in the growth condition detection module, a Transformer encoding method based on a window mechanism and global optimization specifically includes the following steps:

[0021] Step S1: Three-view projection and color addition, perform three-view projection on the rice image, and color the three views according to the rice photos at the corresponding projection angles to obtain a colored rice image;

[0022] Step S2: Feature extraction, evenly divide each colored rice image into n*n sub-images, perform convolutional feature extraction on each sub-image to obtain a convolutional feature map;

[0023] Step S3: Global feature optimization, perform global feature optimization on the convolutional feature map to obtain global features, specifically including the following steps:

[0024] Step S31: Perform global pooling on the convolutional feature map to obtain a global feature vector in the channel dimension ;

[0025] Step S32: Multilayer perceptron processing, process and activate the global feature vector through two layers of multilayer perceptrons to obtain a feature vector with enhanced dependence : ;

[0026] Wherein, and are for multilayer perceptron processing, is the ReLU activation function, is the Sigmoid activation function;

[0027] Step S33: Reshape and fuse. Reshape the dimension of the feature vector to be the same as that of the convolutional feature map, and multiply the reshaped feature vector element-wise with the convolutional feature map to obtain the channel-enhanced feature map;

[0028] Step S34: Convolutional processing. Integrate the spatial information of the channel-enhanced feature map through two convolutional layers and activate it with an activation function to obtain the spatial attention vector;

[0029] Step S35: Global optimization. Multiply the spatial attention vector element-wise with the channel-enhanced feature map to obtain the global feature;

[0030] Step S4: Flatten the global features of all subgraphs to obtain a one-dimensional feature vector. Linearly map all the one-dimensional feature vectors of each rice color map through a linear layer to obtain n*n embedding blocks, and concatenate all the embedding blocks of each rice color map by position to obtain the embedding sequence of each rice color map;

[0031] Step S5: Position encoding embedding. Add position encoding to the embedding sequence according to the positions of all the embedding blocks in the rice color map to obtain the visual embedding sequence;

[0032] Step S6: Transformer encoding. Perform Transformer encoding on the visual embedding sequence based on the local window attention module to obtain the visual feature: ; ; ; ;

[0033] In the formula, represents the visual embedding sequence, represents the regularization process, represents the normalization process, , , are intermediate parameters, represents the visual feature, , , and represent multi-head self-attention processing, feed-forward neural network processing, multi-layer perceptron processing, and local window attention processing respectively;

[0034] Step S7: Classification and recognition of growth status. The visual features corresponding to each rice color image are weighted and fused to obtain the rice growth status features. The activation output and classification of the rice growth status features of all rice plants are performed to obtain the recognition results of the rice growth status and the corresponding relationship between rice growth status - salinity content - rice planting density.

[0035] The beneficial effects achieved by the present invention using the above solution are as follows:

[0036] (1) Through the technology of computer vision, the present invention reasonably analyzes the relationship between rice growth status - salinity content - rice planting density, and intelligently provides guidance on planting density and salinity control for high salt - tolerant rice planting, improving management efficiency, and increasing rice yield and quality.

[0037] (2) The present invention creatively uses a Transformer encoding method based on a window mechanism and global optimization to identify the growth status of rice plants. By spatial - channel optimization, it better captures the spatial characteristics of the growth status of rice plants, introduces a local window mechanism to selectively focus on local regions in the image, extracts key features of the growth status of rice plants, improves the recognition accuracy of the growth status of rice plants, further increases the accuracy of the analysis of the relationship between rice growth status - salinity content - rice planting density, and further provides more accurate guidance for increasing rice yield and quality.

[0038] (3) The present invention creatively uses a UNet network optimized based on Mamba blocks. On the basis of image feature extraction and classification, the encoder of the cascaded module gradually reduces the spatial dimension, effectively capturing local features and global feature representations. At the same time, a convolutional Mamba hybrid block is introduced to mix convolutional operations with three - way Mamba, realizing hierarchical feature extraction with multiple receptive fields, and then accurately segmenting rice images, removing the influence of the field background on rice image feature extraction, and further increasing the accuracy of the analysis of the relationship between rice growth status - salinity content - rice planting density. Description of the Drawings

[0039] Figure 1 It is a module diagram of the high salt - tolerant rice planting management system based on image recognition provided by the present invention;

[0040] Figure 2 It is a schematic flow diagram of a Transformer encoding method based on a window mechanism and global optimization;

[0041] Figure 3 It is a schematic processing flow diagram of a UNet network optimized based on Mamba blocks.

[0042] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. Detailed implementation manners

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0044] Embodiment 1, refer to Figure 1 , a high salt-tolerant rice planting management system based on image recognition, including a data acquisition module, an image processing module, a growth condition detection module, and an analysis and management module;

[0045] The data acquisition module uses a laser scanner to collect 3D images of the growth conditions of high salt-tolerant rice in fields with different saline-alkali contents and collect the corresponding rice density in the fields;

[0046] The image processing module denoises the collected 3D images of rice growth conditions and uses a UNet network optimized based on three-way Mamba blocks to perform 3D image segmentation to remove the field background and obtain rice images;

[0047] The growth condition detection module uses a Transformer encoding method based on a window mechanism and global optimization to classify and identify the growth conditions of rice images in fields with different saline-alkali contents and different rice densities, and obtain the relationship between rice growth conditions - saline-alkali content - rice density;

[0048] The analysis and management module collects the saline-alkali content of the field where high salt-tolerant rice needs to be planted currently, and according to the relationship between rice growth conditions - saline-alkali content - rice density, changes the rice density or adjusts the saline-alkali content of the irrigation water to obtain the optimal rice growth conditions.

[0049] By performing the above operations, reasonably analyze the relationship between rice growth conditions - saline-alkali content - rice density, and intelligently provide density guidance and salinity control for high salt-tolerant rice planting, improve management efficiency, and increase rice yield and quality.

[0050] Embodiment 2, refer to Figure 3 , based on the above embodiment, in the image processing module, a UNet network optimized based on three-way Mamba blocks performs 3D image segmentation, which specifically includes the following steps:

[0051] Step Q1: Feature downsampling. Perform 3D convolution operation with a convolution kernel of 7×7×7, a stride of 2×2×2, and a padding of 3×3×3 on the denoised 3D image of rice growth status to obtain downsampled features;

[0052] Step Q2: Convolution Mamba hybrid block optimization. Optimize the downsampled features based on the convolution Mamba hybrid block to obtain optimized downsampled features;

[0053] Step Q3: Multilayer downsampling. Repeat steps Q1 - Q2 three times, and repeat the downsampling on the optimized downsampled features to obtain, finally, a total of four layers of optimized downsampled features;

[0054] Step Q4: Feature upsampling. Perform multi-layer perceptron processing on the four layers of optimized downsampled features respectively, and perform feature upsampling to obtain segmentation features;

[0055] Step Q5: Feature activation output. Perform activation output on the segmentation features to obtain a segmentation mask, and obtain a segmentation map, that is, a rice image, according to the segmentation mask and the 3D image of rice growth status.

[0056] Example 3. This example is based on the above example, and step Q2 specifically includes the following steps:

[0057] Step Q21: Spatial dimensionality reduction. Use a convolution kernel of 5×5×5 to perform spatial dimensionality reduction and coarse-grained feature extraction on the downsampled features to obtain dimensionality-reduced features;

[0058] Step Q22: Local context relationship optimization. Further refine the dimensionality-reduced features through a convolution of 3×3×3 to capture local context relationships and obtain locally optimized features;

[0059] Step Q23: Three-way Mamba block processing. Use three-way Mamba, that is, TOM Mamba, to perform three-directional feature dependence calculations on the locally optimized features to obtain three-way fusion features: ;

[0060] In the formula, 、 and respectively represent the forward, backward, and feature modeling between three-dimensional slices of the locally optimized features, represents Mamba block processing, represents the locally optimized features, represents the three-way fusion features;

[0061] Step Q24: Residual mechanism optimization. Perform transposed convolutions of 3×3×3 and 5×5×5 on the three-way fusion features and fuse them with the downsampled features based on residual connections, and then perform three-way Mamba block processing on the fusion result again to obtain optimized downsampled features.

[0062] By performing the above operations, the spatial characteristics of the growth status of rice plants can be better captured through spatial-channel optimization, the local window mechanism is introduced to selectively focus on local regions in the image, key features of the growth status of rice plants are extracted, the recognition accuracy of the growth status of rice plants is improved, and further the accuracy of analyzing the relationship between rice growth status - salt content - rice planting density is increased, and further more accurate guidance is provided for increasing rice yield and quality.

[0063] Example 4, refer to Figure 2 , based on the above example, in the growth status detection module, a Transformer encoding method based on a window mechanism and global optimization specifically includes the following steps:

[0064] Step S1: Three-view projection and color addition. Perform three-view projection on the rice image and color the three views according to the rice photos at the corresponding projection angles to obtain a colored rice image;

[0065] Step S2: Feature extraction. Evenly divide each colored rice image into n*n sub-images, perform convolutional feature extraction on each sub-image to obtain a convolutional feature map;

[0066] Step S3: Global feature optimization. Perform global feature optimization on the convolutional feature map to obtain global features;

[0067] Step S4: Flatten the global features of all sub-images to obtain a one-dimensional feature vector, linearly map all the one-dimensional feature vectors of each colored rice image through a linear layer to obtain n*n embedding blocks, and splice all the embedding blocks of each colored rice image by position to obtain an embedding sequence of each colored rice image;

[0068] Step S5: Position encoding embedding. According to the positions of all embedding blocks in the colored rice image, add position encoding to the embedding sequence to obtain a visual embedding sequence;

[0069] Step S6: Transformer encoding. Perform Transformer encoding on the visual embedding sequence based on the local window attention module to obtain visual features: ; ; ; ;

[0070] In the formula, represents the visual embedding sequence, represents the regularization process, represents the normalization process, , , are intermediate parameters, represents the visual feature, , , and respectively represent multi-head self-attention processing, feed-forward neural network processing, multi-layer perceptron processing, and local window attention processing;

[0071] Step S7: Classification and recognition of growth status. The visual features corresponding to each rice color map are weighted and fused to obtain the rice growth status features. The rice growth status features of all rice are activated, output, and classified to obtain the recognition results of the rice growth status and the corresponding relationship between rice growth status - salinity content - rice planting density.

[0072] Example 5. Based on the above example, step S3 specifically includes the following steps:

[0073] Step S31: Perform global pooling on the convolutional feature map to obtain a global feature vector in the channel dimension ;

[0074] Step S32: Multi-layer perceptron processing. The global feature vector is processed and activated by two layers of multi-layer perceptrons to obtain a feature vector with enhanced dependence : ;

[0075] In the formula, and are multi-layer perceptron processing, is the ReLU activation function, is the Sigmoid activation function;

[0076] Step S33: Reshape and fuse. Reshape the dimension of the feature vector to be the same as that of the convolutional feature map, and multiply the reshaped feature vector element-wise with the convolutional feature map to obtain a channel-enhanced feature map;

[0077] Step S34: Convolution processing. Perform spatial information integration on the channel-enhanced feature map through two convolutional layers and activate with an activation function to obtain a spatial attention vector;

[0078] Step S35: Global optimization, perform element-wise multiplication on the spatial attention vector and the channel-enhanced feature map to obtain the global feature.

[0079] By performing the above operations, based on image feature extraction and classification, the spatial dimension is gradually reduced through the encoder of the cascading module, effectively capturing local and global feature representations. At the same time, a convolutional Mamba hybrid block is introduced to mix convolutional operations with three-way Mamba, realizing hierarchical feature extraction with multiple receptive fields, thereby accurately segmenting rice images, removing the influence of the field background on rice image feature extraction, and further increasing the accuracy of analyzing the relationship between rice growth conditions - salinity content - rice planting density.

[0080] Example Six, based on the above example, the implementation example of this solution is as follows:

[0081] (1) Data collection and preprocessing

[0082] Collection:

[0083] Drone: Collect high-resolution rice RGB images;

[0084] Laser scanner: 3D images of rice;

[0085] IoT device: Obtain soil salinity values in real time;

[0086] Image preprocessing:

[0087] Denoising: Use wavelet transform to remove noise;

[0088] Segmentation: Use a UNet network optimized based on the three-way Mamba block to extract the rice image area;

[0089] Feature enhancement: Improve image contrast through histogram equalization;

[0090] (2) Image recognition and analysis

[0091] Introduce a deep learning model based on a Transformer encoding method with a window mechanism and global optimization:

[0092] ResNet: Used for the classification of rice diseases and salt stress features;

[0093] YOLO: Used for quickly detecting pest and disease areas;

[0094] Model training:

[0095] Data source: Rice field images in a high-salt environment;

[0096] Data augmentation: Expand training data by means of random rotation, cropping, adding noise, etc.;

[0097] Key features:

[0098] Salt stress: changes in leaf color (yellowish RGB values) and decreased texture refinement, mainly reflected in the top view of the plant;

[0099] Disease: abnormal shape and texture of color patches (such as diamond-shaped lesions of rice blast);

[0100] Growth status: changes in plant height and density, mainly reflected in the front view and side view of the three-view drawings;

[0101] (3) Zoning management of saline-alkali land

[0102] Combination of GIS and images:

[0103] Generate a salt distribution map of saline-alkali land based on UAV images;

[0104] Use clustering algorithms (such as K-Means) to divide the fields into different salt regions;

[0105] Precision irrigation and reasonable close planting:

[0106] Combine the salt distribution map and the rice close planting requirements to generate a zoned irrigation plan and a close planting density management plan.

[0107] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0108] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made therein without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

[0109] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A high salt tolerance rice planting management system based on image recognition, characterized by: It includes data acquisition module, image processing module, growth status detection module and analysis management module; The data acquisition module uses a laser scanner to collect 3D images of the growth status of highly salt-tolerant rice in fields with different salinity and alkali content and collects the corresponding rice planting density in the fields; The image processing module performs denoising on the collected 3D image of rice growth status, and uses a UNet network based on three-way Mamba block optimization to perform 3D image segmentation to remove the field background, thereby obtaining a rice image; The growth status detection module uses a Transformer encoding method based on a window mechanism and global optimization to classify and identify the growth status of rice images in fields with different salinity and alkali content and different planting densities, and obtains the relationship between rice growth status, salinity and alkali content, and rice planting density. The analysis and management module collects the salinity content of the field where the highly salt-tolerant rice is currently planted, and manages the rice planting density to obtain the optimal rice growth condition based on the relationship between rice growth condition-salt-alkali content-rice planting density.

2. The high salt tolerance rice planting management system based on image recognition according to claim 1, characterized in that: In the image processing module, a UNet network based on three-way Mamba block optimization performs 3D image segmentation, which specifically includes the following steps: Step Q1: Perform 3D convolution operation on the denoised 3D image of rice growth status to obtain down-sampled features; Step Q2: Optimize the downsampled features based on the convolutional Mamba mixing block to obtain optimized downsampled features; Step Q3: Repeat step Q1-step Q2 three times, repeatedly downsample the optimized downsampled features, and finally obtain four layers of optimized downsampled features; Step Q4: The four layers of optimized downsampled features are processed by multi-layer perceptrons respectively, and the features are upsampled to obtain segmentation features; Step Q5: Activate and output the segmentation features to obtain a segmentation mask, and obtain a segmentation map, i.e., a rice image, based on the segmentation mask and the 3D image of the rice growth status.

3. The high salt tolerance rice planting management system based on image recognition according to claim 2, characterized in that: Step Q2 specifically includes the following steps: Step Q21: Use a 5×5×5 convolution kernel to perform spatial dimension reduction and coarse-grained feature extraction on the downsampled features to obtain reduced-dimensional features; Step Q22: further refine the reduced dimension features through 3×3×3 convolution to capture local contextual relationships and obtain local optimized features; Step Q23: Use three-way Mamba, i.e., TOM Mamba, to calculate the three-way features of the local optimization features to obtain the three-way fusion features: ； In the formula, , and The forward, reverse and three-dimensional slice feature modeling represent the local optimization features respectively. Represents Mamba block processing, represents the local optimization feature, It represents the three-way fusion feature; Step Q24: Perform transposed convolution on the three-way fusion feature and fuse it with the down-sampled feature based on residual connection, and perform three-way Mamba block processing on the fusion result again to obtain the optimized down-sampled feature.

4. The high salt tolerance rice planting management system based on image recognition according to claim 1, characterized in that: In the growth status detection module, a Transformer encoding method based on a window mechanism and global optimization specifically includes the following steps: Step S1: Project the rice image in three views, and color the three views according to the rice photos at corresponding projection angles to obtain a rice color map; Step S2: Each rice color image is evenly divided into n*n sub-images, and convolution feature extraction is performed on each sub-image to obtain a convolution feature map; Step S3: performing global feature optimization on the convolution feature map to obtain global features; Step S4: Flatten the global features of all sub-images to obtain a one-dimensional feature vector, linearly map all the one-dimensional feature vectors of each rice color image through a linear layer to obtain n*n embedding blocks, and splice all the embedding blocks of each rice color image by position to obtain an embedding sequence of each rice color image; Step S5: adding position encoding to the embedding sequence according to the positions of all embedded blocks in the rice color image to obtain a visual embedding sequence; Step S6: Perform Transformer encoding based on the local window attention module on the visual embedding sequence to obtain visual features; Step S7: weighted fusion is performed on the visual features corresponding to each rice color image to obtain the rice growth status features, and the rice growth status features of all rice are activated, output and classified to obtain the recognition result of the rice growth status and the corresponding rice growth status-salinity-rice planting density relationship.

Images