Hamming melon growth state evaluation method and system based on artificial intelligence
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-12
Smart Images
Figure CN122200638A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image enhancement technology, specifically referring to an artificial intelligence-based method and system for assessing the growth status of cantaloupes. Background Technology
[0002] The Hami melon growth status assessment method is a technical solution that uses image enhancement and deep learning technology to analyze multispectral image data of Hami melons and assess their current growth status in real time. It provides scientific basis for growers, agricultural enterprises and agricultural technology extension departments, helps to formulate precise planting and management measures such as irrigation, fertilization and pest and disease control, thereby improving the yield and quality of Hami melons and promoting the development of smart agriculture.
[0003] However, existing methods for assessing the growth status of Hami melons suffer from several problems. Firstly, hyperspectral images of Hami melons lack sufficient spatial-spectral correlation representation. Secondly, subtle key regional features are easily overwhelmed by redundant information. Thirdly, image enhancement struggles to balance resolution and detail fidelity, resulting in a lack of high-quality input data and insufficient accuracy in assessing the growth status of Hami melons. Furthermore, existing methods for assessing the growth status of Hami melons also struggle to simultaneously and efficiently represent spatial morphology and spectral information. Fourthly, global spatial-spectral long-range dependencies are difficult to capture, and subtle features are easily lost, making it difficult to accurately distinguish between various growth states of Hami melons and leading to unstable assessment results. Summary of the Invention
[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based method and system for assessing the growth status of cantaloupes. Addressing the problems in existing cantaloupe growth status assessment methods, such as insufficient spatial-spectral correlation representation of hyperspectral images, the easy submersion of small key region features by redundant information, and the difficulty in balancing resolution and detail fidelity in image enhancement, leading to a lack of high-quality input data and insufficient accuracy in cantaloupe growth status assessment, this solution obtains corrected multi-band images through a Gaussian function and a positive value suppression strategy. It then obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences, and through adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of growth stages is introduced to generate global enhancement features. Finally, through spatial local window segmentation and spectral band grouping, refined representation features are generated, resulting in a hyperspectral enhanced image of the cantaloupe. High-resolution images ensure clear edges and spectral continuity, providing high-quality input for accurate assessment of growth status and effectively improving the accuracy of cantaloupe growth status assessment. Addressing the problems in existing cantaloupe growth status assessment methods, such as difficulty in simultaneously and efficiently representing spatial morphology and spectral information, difficulty in capturing long-range global spatial-spectral dependencies, and easy loss of subtle features, leading to inaccurate differentiation of various growth states and unstable assessment results, this scheme extracts spatial and spectral feature maps, generates a weighted spectral feature map, and obtains a comprehensive feature map through stitching. It introduces a cantaloupe growth topology soft mask, injects dual-domain positional encoding, adds a growth status classification token, guides the global sequence of growth stages, and obtains a dual-domain perceptual global aggregated feature sequence through global feature modeling. It then extracts fine-grained spatial features and spectral consistency features respectively, generates a refined high-order feature map, and outputs a growth status prediction label, significantly improving the accuracy and stability of cantaloupe growth status assessment.
[0005] The technical solution adopted by this invention is as follows: The artificial intelligence-based method for evaluating the growth status of cantaloupe provided by this invention includes the following steps:
[0006] Step S1: Acquiring images of cantaloupes;
[0007] Step S2: Hami melon image enhancement processing;
[0008] Step S3: Construction of a cantaloupe growth status assessment model;
[0009] Step S4: Assessment of Hami melon growth status.
[0010] Further, in step S1, the cantaloupe image acquisition involves acquiring historical cantaloupe hyperspectral images, removing the first and last noisy bands to obtain the effective spectral bands, and then performing preprocessing and image annotation; the preprocessing includes size unification, band-by-band grayscale image extraction, illumination equalization, and grayscale normalization; the image annotation includes labeling the growth stage and growth status.
[0011] Further, in step S2, the cantaloupe image enhancement processing specifically includes the following steps:
[0012] Step S21: Band-by-band illumination correction; For the preprocessed cantaloupe hyperspectral image, calculate the gray mean and standard deviation for each band, adaptively compensate for shadows in dark areas using a Gaussian function, and compress overexposed bright areas using a positive value suppression strategy. Finally, normalize the correction results and stitch them together along the band dimension to obtain the corrected multi-band image.
[0013] Step S22: Dual-branch feature extraction; Three-dimensional convolution is used to downsample the corrected multi-band image. The downsampled features are extracted using two parallel branches with the same structure but independent parameters. Both branches use plant-aware decomposition convolution to obtain gated modulation features and global scanning features.
[0014] Step S23: Topological rearrangement selective scanning; Perform topological rearrangement selective scanning on the global scanning features to reconstruct the global scanning features into four global sequences: spatial forward, spatial backward, spectral forward, and spectral backward. Then, independently calculate the adaptive scale parameters for each global sequence and perform state space transformation to achieve global unified feature mapping and obtain four-way global modeling features.
[0015] Step S24: Stage Adaptive Modulation Enhancement; Global average pooling is performed on the global scanning features, and they are concatenated with the one-hot encoding of the growth stage and then input into a multilayer perceptron to generate growth stage adaptive weights. The gated modulation features are aligned to the spatial size and number of channels of the global modeling features through convolution and bilinear interpolation. Then, the Hadamard product is used to perform weighted modulation of the aligned gated modulation features, global modeling features and growth stage adaptive weights in sequence. The four weighted modulation features are summed and then input into the plant perception decomposition convolution to generate global enhancement features.
[0016] Step S25: Local Dual Attention Refinement; The global enhancement features are spatially segmented into local windows and grouped into spectral bands. Spatial local features and band group features are extracted separately. Spatial local attention weights and spectral channel attention weights are learned for each local window and each band group through independent small convolutional network branches. The learned weights are then expanded to the original feature map size through a broadcasting mechanism. The two attention weights are then fused with the global enhancement features element-wise and normalized and smoothed. Finally, the features are refined through plant perception decomposition convolution to generate fine representation features.
[0017] Step S26: Enhanced image generation; channel integration of the sum of global enhancement features and fine representation features is performed by convolution, and then bilinear interpolation upsampling is performed to generate a hyperspectral enhanced image of cantaloupe.
[0018] Furthermore, in step S3, the construction of the cantaloupe growth status assessment model specifically includes the following steps:
[0019] Step S31: Spatial-Spectral Dual Branch; The hyperspectral enhanced image of the cantaloupe is fed into a dual-branch CNN backbone network, including the EfficientNet-B0 branch and the ResNet-50 branch, to obtain spatial feature maps and spectral feature maps.
[0020] Step S32: Channel attention splicing and fusion; The spatial dimension of the spectral feature map is compressed by global average pooling to generate a global description vector of the spectral channel. Then, the importance weight of each band is learned through two fully connected layers. The importance weight is weighted element-wise with the original spectral feature map in the channel dimension to obtain a weighted spectral feature map. The spatial feature map and the weighted spectral feature map are spliced in the channel dimension to generate a comprehensive feature map.
[0021] Step S33: Topology-guided sequence encoding; Generate a cantaloupe growth topology soft mask based on spatial gradient and average gray level, input the comprehensive feature map into the growth topology-aware convolutional layer, and obtain the growth topology-aware sequence after block segmentation and projection processing. Then, inject dual-domain position encoding, which simultaneously encodes spatial position information and spectral band order information to obtain a dual-domain position-aware sequence. Add a growth state classification token to the beginning of the sequence to obtain the growth stage guidance global sequence.
[0022] Step S34: Global feature modeling; The global sequence guided by the growth stage is input into a 4-layer stacked Transformer encoder for global feature modeling to obtain a dual-domain perceptual global aggregated feature sequence;
[0023] Step S35: Gated Feature Refinement; The dual-domain perception global aggregated feature sequence is reshaped into a two-dimensional feature map. A structure with parallel spatial detail convolution branch and spectral smoothing convolution branch is adopted to extract spatial fine-grained features and spectral consistency features respectively. Then, a cantaloupe growth topological soft mask is introduced to mask the features. At the same time, the one-hot encoding of the cantaloupe growth stage is dynamically modulated by global pooling and adaptive gating weights in the activation generation stage to generate a refined high-order feature map.
[0024] Step S36: Growth state classification; global average pooling is performed on the refined high-order feature map to generate a global feature vector. The global feature vector is then subjected to linear mapping and Dropout regularization to suppress overfitting. Finally, the probability distribution of the four growth states of cantaloupe is obtained through the Softmax activation function, and the predicted label of the growth state is output.
[0025] Further, in step S4, the assessment of the growth status of the cantaloupe involves acquiring a real-time hyperspectral image of the cantaloupe, removing the first and last noise bands to obtain the effective spectral bands, preprocessing and labeling the growth stages, performing image enhancement processing based on step S2 to generate a real-time hyperspectral enhanced image of the cantaloupe, and inputting it into the cantaloupe growth status assessment model constructed in step S3 for classification to obtain the growth status corresponding to the real-time hyperspectral image of the cantaloupe, thus completing the assessment of the growth status of the cantaloupe.
[0026] The artificial intelligence-based cantaloupe growth status assessment system provided by the present invention includes a cantaloupe image acquisition module, a cantaloupe image enhancement and processing module, a cantaloupe growth status assessment model construction module, and a cantaloupe growth status assessment module.
[0027] The cantaloupe image acquisition module acquires historical hyperspectral images of cantaloupes, performs preprocessing and image annotation, and sends the data to the cantaloupe image enhancement processing module;
[0028] The cantaloupe image enhancement processing module obtains the corrected multi-band image through Gaussian function and positive value suppression strategy. It obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences. After adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of the growth stage is introduced to generate global enhancement features. Through spatial local window segmentation and spectral band grouping, fine characterization features are generated to obtain the cantaloupe hyperspectral enhanced image. The data is then sent to the cantaloupe growth status assessment model construction module.
[0029] The Hami melon growth status assessment model construction module extracts spatial feature maps and spectral feature maps, generates a weighted spectral feature map, and obtains a comprehensive feature map by splicing them together. It introduces a Hami melon growth topology soft mask, injects dual-domain position encoding and adds a growth status classification token, guides the global sequence of growth stages, obtains a dual-domain perception global aggregation feature sequence through global feature modeling, extracts spatial fine-grained features and spectral consistency features respectively, generates a refined high-order feature map, outputs growth status prediction labels, and sends the data to the Hami melon growth status assessment module.
[0030] The cantaloupe growth status assessment module enhances and classifies real-time hyperspectral images of cantaloupes to obtain the growth status corresponding to the real-time hyperspectral images of cantaloupes.
[0031] The beneficial effects achieved by the present invention using the above solution are as follows:
[0032] (1) To address the problems in existing methods for assessing the growth status of Hami melons, such as insufficient spatial-spectral correlation representation of hyperspectral images, easy submersion of small key regional features by redundant information, and difficulty in balancing resolution and detail fidelity in image enhancement, resulting in a lack of high-quality input data and insufficient accuracy in assessing the growth status of Hami melons, this scheme obtains corrected multi-band images through Gaussian functions and positive value suppression strategies to eliminate field light noise; it obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches, respectively, preserving spectral dimensional correlation, while extracting spatial texture and band correlation; and it further refines the global scanning features. The descriptive features are reconstructed into four global sequences. Through adaptive scaling and state space transformation, four global modeling features are obtained, establishing a stable correlation between the entire spatial domain and the entire spectral band, and strengthening the features of key regions. One-hot encoding of the growth stage is introduced to generate global enhanced features and enhance spatial-spectral correlation. Through spatial local window segmentation and spectral band grouping, fine characterization features are generated to accurately focus on subtle discrimination regions and suppress non-target noise. The resulting hyperspectral enhanced image of Hami melon restores the high-resolution image, ensuring clear edges and spectral continuity, providing high-quality input for accurate assessment of growth status, and effectively improving the accuracy of Hami melon growth status assessment.
[0033] (2) To address the problems in existing methods for assessing the growth status of Hami melons, such as difficulty in simultaneously and efficiently representing spatial morphology and spectral information, difficulty in capturing long-range global spatial-spectral dependence, and easy loss of subtle features, which lead to difficulty in accurately distinguishing various growth statuses of Hami melons and unstable assessment results, this scheme extracts spatial feature maps and spectral feature maps to achieve independent and efficient representation of the spatial and spectral domains; generates weighted spectral feature maps, and obtains comprehensive feature maps by splicing them together, integrating spatial morphology and spectral physicochemical information to improve the comprehensive discrimination ability of features; introduces Hami melon growth topology soft mask, injects dual-domain position encoding and adds growth status classification tokens, guides the global sequence of growth stages, synchronously perceives spatial layout and band order, incorporates prior knowledge of growth period, and enhances the structural rationality of the sequence; obtains dual-domain perception global aggregated feature sequence through global feature modeling, and realizes global joint modeling of the spatial morphology, spectral response and growth stage prior of the whole Hami melon; extracts spatial fine-grained features and spectral consistency features respectively, generates refined high-order feature maps, and repairs detailed information; outputs growth status prediction labels, which significantly improves the accuracy and stability of assessing different growth statuses of Hami melons. Attached Figure Description
[0034] Figure 1 A flowchart illustrating the artificial intelligence-based method for assessing the growth status of cantaloupe provided by this invention;
[0035] Figure 2 A schematic diagram of the artificial intelligence-based cantaloupe growth status assessment system provided by the present invention;
[0036] Figure 3 This is a flowchart illustrating step S2;
[0037] Figure 4 This is a flowchart illustrating step S3.
[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0040] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0041] Example 1, see Figure 1 The present invention provides an artificial intelligence-based method for assessing the growth status of cantaloupe, which includes the following steps:
[0042] Step S1: Hami melon image acquisition; acquire historical hyperspectral images of Hami melons, and perform preprocessing and image annotation;
[0043] Step S2: Hami melon image enhancement processing; a corrected multi-band image is obtained through Gaussian function and positive value suppression strategy. Gated modulation features and global scanning features are obtained through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences. After adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of the growth stage is introduced to generate global enhancement features. Fine characterization features are generated through spatial local window segmentation and spectral band grouping to obtain a hyperspectral enhanced image of Hami melon.
[0044] Step S3: Construction of Hami melon growth status assessment model; extract spatial feature map and spectral feature map, generate weighted spectral feature map, obtain comprehensive feature map by splicing, introduce Hami melon growth topology soft mask, inject dual-domain position encoding and add growth status classification token, guide global sequence of growth stage, obtain dual-domain perception global aggregated feature sequence through global feature modeling, extract spatial fine-grained features and spectral consistency features respectively, generate refined high-order feature map, and output growth status prediction label;
[0045] Step S4: Assessment of the growth status of cantaloupe; enhance and classify the real-time hyperspectral image of cantaloupe to obtain the growth status corresponding to the real-time hyperspectral image of cantaloupe.
[0046] Example 2, see Figure 1This embodiment is based on the above embodiment. In step S1, the Hami melon image acquisition involves acquiring historical hyperspectral images of Hami melons, removing the first and last noise bands to obtain the effective spectral bands, and then performing preprocessing and image annotation. The preprocessing includes size unification, band-by-band grayscale image extraction, illumination equalization, and grayscale normalization. Size unification involves adjusting the image spatial size to 512×512 pixels using bilinear interpolation. Band-by-band grayscale image extraction involves directly extracting the single-channel grayscale image corresponding to each effective spectral band. Illumination equalization involves using adaptive histogram equalization to eliminate brightness differences caused by uneven field illumination, leaf shadows, and overexposure of strong light. Grayscale normalization involves linearly scaling the grayscale values of the grayscale image to the [0, 1] range. Image annotation includes labeling growth stages and growth states. The growth stages include seedling stage, vine extension stage, flowering and fruit setting stage, fruit enlargement stage, and maturity stage. The growth states include healthy and vigorous state, mild stress state, moderate stress state, and unhealthy and diseased state.
[0047] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the cantaloupe image enhancement processing specifically includes the following steps:
[0048] Step S21: Band-by-band illumination correction; The preprocessed hyperspectral image of the cantaloupe still suffers from issues such as overlapping shadows from canopy leaves, localized overexposure, and non-uniform illumination noise. For the preprocessed hyperspectral image, the mean and standard deviation of grayscale are calculated band by band. Adaptive compensation for shadows in dark areas is performed using a Gaussian function, and a positive value suppression strategy is employed to compress overexposed bright areas. Finally, the correction results are renormalized, and the corrected multi-band image is stitched together along the band dimension. This further eliminates field illumination interference, restores the true reflectance characteristics of cantaloupe leaves and fruits, and provides image data with stronger illumination consistency and higher spectral fidelity for subsequent feature extraction. The formulas used are as follows:
[0049] ;
[0050] ;
[0051] ;
[0052] ;
[0053] In the formula, This is the normalized grayscale image of the d-th spectral band, where d is the band index, and D... total It is the total number of bands, μ d and σ d These are the global grayscale mean and grayscale standard deviation of the d-th band image, respectively, σ globalIt is all bands σ d The mean, k s It is the adaptive compensation coefficient for shadows, which controls the enhancement level of shadow areas. ; It is the Gaussian compensation range coefficient, which controls the width of the effective range of shadow compensation. ;k c It is the overexposure suppression coefficient, which controls the compression strength of overexposed areas. ; It is a positive operation. It is a smoothing term. ; and These are the intermediate images after adaptive shadow compensation and overexposure suppression processing for the d-th band, respectively. and They are The minimum and maximum grayscale values of all pixels in the array. , and They are the 1st, the dth, and the Dth, respectively. total Standard images after illumination correction and renormalization for each band, I corr This is the image after multi-band correction. It's a splicing operation;
[0054] Step S22: Dual-branch feature extraction; Hyperspectral data has complex spatial-spectral correlations, and conventional convolution easily severs band correlations, failing to simultaneously extract effective plant features and suppress background noise. Three-dimensional convolution is used to downsample the corrected multi-band image, compressing spatial resolution and increasing feature channel dimensions while maintaining spectral dimensional correlations. To simultaneously extract spectral correlations and spatial texture features from the cantaloupe hyperspectral image, enhance the effective plant region, and suppress background noise, two parallel branches with identical structures but independent parameters are used for feature extraction of the downsampled features. Both branches use plant-aware decomposition convolution. The plant-aware decomposition convolution first performs a one-dimensional convolution along the spectral dimension to extract inter-band correlations, then performs a two-dimensional convolution along the spatial dimensions (height and width) to extract spatial texture information. Finally, Sigmoid gating is used to enhance feature expression, resulting in gated modulation features and global scanning features. The formulas used are as follows:
[0055] ;
[0056] ;
[0057] ;
[0058] ;
[0059] In the formula, Fdown It is a downsampling feature. This is a 3D convolution operation. `kernel` is the kernel size, and `stride` is the convolution stride. It is plant perception decomposition convolution, F in These are input features. and These are two-dimensional convolution along the spatial dimension and one-dimensional convolution along the spectral dimension, respectively. It is the Hadamard product operation. It is the Sigmoid activation function. It is a 1×1 convolution operation. and These are the plant perception decomposition convolutions of the gated branch and the feature branch, respectively, where Z and X are the gated modulation feature and the global scanning feature, respectively.
[0060] Step S23: Topological Rearrangement Selective Scanning; To address the problem that traditional single-direction scanning is insufficient to fully characterize the spatial structure and spectral correlation of cantaloupe and easily generates information bias, and to fully model the global spatial-spectral dependence of hyperspectral images while maintaining low computational overhead, topological rearrangement selective scanning is performed on the global scanning features. Based on the cantaloupe growth topology (root-stem-leaf-fruit), the global scanning features are reconstructed into four global sequences: spatially forward, spatially backward, spectrally forward, and spectrally backward. Adaptive scaling parameters are then independently calculated for each global sequence, and state-space transformation is performed to achieve a globally unified feature mapping. This establishes a stable correlation between the entire spatial domain and the entire spectral band, effectively eliminating the information bias caused by unidirectional scanning, strengthening the feature expression of key regions, and improving the accuracy and robustness of image enhancement and growth status assessment, resulting in four-way global modeling features. The formulas used are as follows:
[0061] ;
[0062] ;
[0063] ;
[0064] ;
[0065] ;
[0066] In the formula, X1 is the spatially positive global sequence feature. X1 is a topological dimension rearrangement operation for cantaloupe growth, unfolding features according to the root-stem-leaf-fruit spatial structure; X2 is the spatial reverse global sequence feature. This is a sequence reversal operation; X3 is the spectral forward global sequence feature. This is a spectral dimension rearrangement operation; X4 is the inverse global sequence feature of the spectrum. It is the Softplus activation function. It is the adaptive scaling parameter corresponding to the i-th global sequence, where i is the sequence index. It is a linear mapping layer specifically for learning scale parameters. It is a scale information embedding vector. and These are the discretized state transformation matrix and discretized input mapping matrix corresponding to the i-th global sequence, respectively. A and B are the original state-space system matrix and the original state-space input matrix, respectively. H i It is the global hidden state feature corresponding to the i-th global sequence. Y is the global feature vector obtained by performing global average pooling on X in the spatial and spectral dimensions. C and D are the state-space output mapping matrix and the state-space pass-through mapping matrix, respectively. i It is the global modeling feature of the i-th global sequence after state-space transformation;
[0067] Step S24: Stage-Adaptive Modulation Enhancement; To enable the global modeling features to adaptively match the evaluation needs of different growth stages of cantaloupe, and to enhance the feature expression of key areas such as leaves, fruits, and lesions within the corresponding growth period, while suppressing the interference of background and invalid information, global average pooling is performed on the global scanning features. This is then concatenated with the one-hot encoding of the growth stage and input into a multilayer perceptron to generate growth stage adaptive weights. The gated modulation features are aligned to the spatial size and number of channels of the global modeling features through convolution and bilinear interpolation. Then, the Hadamard product is used to sequentially perform weighted modulation of the aligned gated modulation features, global modeling features, and growth stage adaptive weights. The sum of the four weighted modulation features is then input into the plant perception decomposition convolution to generate a global enhanced feature with spatial-spectral global correlation, growth stage adaptive enhancement, and strong response of the target region. The feature expression is highly matched with the growth period, suppressing invalid background interference and significantly improving the distinguishability of different growth stages. The formula used is as follows:
[0068] ;
[0069] ;
[0070] ;
[0071] In the formula, W cycle It is an adaptive weight during the growth stage. It is a multilayer perceptron. It is global average pooling. It is a feature concatenation operation, G stage It is the one-heat encoding vector of the cantaloupe growth stage. It is the weighted modulation feature of the i-th path. , It is the aligned gated modulation feature. This is a size and channel alignment operation, F sum It is the result of summing the four-way weighted modulation characteristics, F global It is a global enhancement feature;
[0072] Step S25: Local Dual Attention Refinement; To address the issues of insufficient response of global enhancement features to small, critical areas such as cantaloupe lesions and fruits, and their susceptibility to being overwhelmed by redundant information from leaves and background, while avoiding the drawbacks of high computational cost and low focusing accuracy of traditional global attention features, spatial local window segmentation and spectral band grouping are performed on the global enhancement features. Spatial local features and band group features are extracted separately. Independent small convolutional network branches learn spatial local attention weights and spectral channel attention weights for each local window and each band group. The learned weights are then broadcast to the original feature map size, adaptively enhancing the feature response of sensitive areas such as lesions and fruits, and suppressing noise in non-target areas. The two attention weights are then element-wise weighted and fused with the global enhancement features, followed by normalization and smoothing. Finally, feature refinement is completed through plant perception decomposition convolution, generating refined representation features that accurately focus on small discrimination areas and improve the feature representation ability of subtle growth states such as stress and disease. The formulas used are as follows:
[0073] ;
[0074] ;
[0075] ;
[0076] ;
[0077] ;
[0078] ;
[0079] ;
[0080] In the formula, R spa and R spe These are spatial local features and band grouping features. This is a spatial local window segmentation operation; `size` is the size parameter for the spatial local window segmentation. This is a spectral band grouping operation, where P is the number of spectral band groups, and W... spa and W spe These are spatial local attention weights and spectral channel attention weights, respectively. and These are spatial attention convolutional layers and spectral attention convolutional layers, F attn It is a dual attention-weighted fusion feature. It's a broadcast operation, F norm It is a normalized smoothing feature. and They are F attn The minimum and maximum values of all pixel feature values, F fine It is a refined representation feature;
[0081] Step S26: Image Enhancement Generation; To restore spatial resolution while maintaining global structural consistency and local detail fidelity, a 3×3 convolution is used to integrate the sum of global enhancement features and fine representation features, followed by bilinear interpolation upsampling (upsampling factor of 2) to restore the feature map spatial size from 256×256 to 512×512, while keeping the number of bands unchanged. This generates a hyperspectral enhanced image of the cantaloupe, effectively avoiding information bias caused by single feature upsampling. The resulting enhanced image has clear edges, continuous spectrum, and consistent global response, providing more complete input information for assessing the growth status of cantaloupes. The formula used is as follows:
[0082] ;
[0083] ;
[0084] In the formula, F merge It is a feature of fusion. It is a 3×3 convolution. It is bilinear interpolation, I enh This is a hyperspectral enhanced image of a cantaloupe.
[0085] By performing the above operations, this method addresses the problems in existing methods for assessing the growth status of Hami melons. These problems include insufficient spatial-spectral correlation representation of hyperspectral images, the easy submersion of small key region features by redundant information, and the difficulty in balancing resolution and detail fidelity in image enhancement, leading to a lack of high-quality input data and insufficient accuracy in assessing Hami melon growth status. This solution obtains corrected multi-band images using a Gaussian function and a positive value suppression strategy to eliminate field light noise. It also obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches, preserving spectral dimensional correlations while extracting spatial texture and band correlations. The global scanning features are reconstructed into four global sequences. Through adaptive scaling and state space transformation, four global modeling features are obtained, establishing a stable correlation between the entire spatial domain and the entire spectral band, and strengthening the features of key regions. One-hot encoding of the growth stage is introduced to generate global enhanced features and enhance the spatial-spectral correlation. Through spatial local window segmentation and spectral band grouping, fine characterization features are generated to accurately focus on subtle discrimination regions and suppress non-target noise. The resulting hyperspectral enhanced image of Hami melon restores the high-resolution image, ensuring clear edges and spectral continuity, providing high-quality input for accurate assessment of growth status, and effectively improving the accuracy of Hami melon growth status assessment.
[0086] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the construction of the cantaloupe growth status assessment model specifically includes the following steps:
[0087] Step S31: Spatial-Spectral Dual Branch; A single network cannot simultaneously and efficiently represent the spatial morphological texture and multi-band spectral reflectance patterns of Hami melon, resulting in simplistic feature representation. Therefore, the hyperspectral enhanced image of Hami melon is fed into a dual-branch CNN backbone network, including the EfficientNet-B0 branch and the ResNet-50 branch, to obtain spatial and spectral feature maps. The EfficientNet-B0 branch is used to extract spatial domain features such as spatial texture, morphology, and edges of Hami melon leaves, fruits, and lesion areas. The ResNet-50 branch is used to extract spectral domain features such as spectral reflectance patterns, band correlations, and spectral response differences between multiple bands. Both branches remove the original classification heads, outputting only deep feature maps to achieve independent and efficient representation of spectral and spatial information, providing a foundation for subsequent fusion modeling. The formulas used are as follows:
[0088] ;
[0089] ;
[0090] In the formula, O spa and O spe These are spatial feature maps and spectral feature maps, It is the EfficientNet-B0 feature extraction function. It is the ResNet-50 feature extraction function;
[0091] Step S32: Channel Attention Concatenation and Fusion; To highlight the effective spectral bands highly correlated with the growth state of cantaloupe and suppress noise band interference, the spatial dimension of the spectral feature map is compressed by global average pooling to generate a global description vector for the spectral channels. Then, the importance weights of each band are learned through two fully connected layers. The importance weights are then element-wise weighted with the original spectral feature map along the channel dimension to obtain a weighted spectral feature map that enhances the effective spectral information. This makes the model more focused on spectral features valuable for assessing the growth state. The spatial feature map and the weighted spectral feature map are concatenated along the channel dimension to fuse spatial morphological information and spectral physicochemical information, generating a comprehensive feature map that simultaneously contains local spatial details, global spatial structure, and inter-band correlation characteristics. The formula used is as follows:
[0092] ;
[0093] ;
[0094] ;
[0095] ;
[0096] In the formula, s spe This is the global description vector for the spectral channels, where M and N are the height and width of the spectral feature map, respectively, and m and n are the pixel indices in the height and width directions of the feature map, respectively. It is the spatial location of the spectral feature map The eigenvectors of all spectral channels at point ω spe W1 is the spectral channel attention importance weight vector, and W2 and W1 are the weight matrices of the second and first fully connected layers, respectively. It is a modified linear unit activation function. It is a weighted spectral feature map, O fusion It is a comprehensive feature map;
[0097] Step S33: Topology-guided sequence encoding; Conventional block encoding does not conform to the physiological structure of cantaloupe, lacks spatial and spectral position information, and lacks prior guidance on growth stages for classification. A soft mask for cantaloupe growth topology is generated based on spatial gradient and average grayscale. The comprehensive feature map is input into a growth topology-aware convolutional layer. After block segmentation and projection processing, a growth topology-aware sequence is obtained. This mask is then used to guide feature segmentation, ensuring the model conforms to the plant's physiological structure from the initial stage. Subsequently, dual-domain position encoding is injected, simultaneously encoding spatial position information and spectral band order information, resulting in a dual-domain position-aware sequence. This allows the model to simultaneously perceive spatial layout and band correlation. A growth state classification token is added to the beginning of the sequence, incorporating the one-hot encoding of the cantaloupe growth stage into the token initialization. This ensures that the classification features possess prior knowledge of the growth period from the global aggregation stage, resulting in a global sequence guided by the growth stage. The formula used is as follows:
[0098] ;
[0099] ;
[0100] ;
[0101] ;
[0102] ;
[0103] In the formula, L emb It is a growing topological sensing sequence. It is a topology-aware convolutional function, where Q is the hidden layer feature dimension (Q=128), U is the block size (U=16), and M... topo It is a soft mask for the growth topology of Hami melons. This is a standard 2D convolution operation, where K is the number of output channels, S is the stride, and J is the padding size. It is a Gaussian blur function. It is an indicator function. It is the spatial gradient of the hyperspectral enhanced image of Hami melon in the d-th band. This is the average grayscale image of a cantaloupe in all bands using hyperspectral enhancement. part This is a weighted map of physiological regions in Hami melons. V emb and V pos These are spectral band position coding and spatial position coding, respectively. V0 is a dual-domain position-aware sequence, and x... cls It is a growth state classification token, x base It is a basic learnable categorized token. It is the growth phase that guides the global sequence;
[0104] Step S34: Global Feature Modeling; CNNs only possess local receptive fields and cannot capture long-range spatial dependencies across regions and global spectral correlations across bands, resulting in insufficient high-order discriminative information. The growth stage-guided global sequence is input into a 4-layer stacked Transformer encoder for global feature modeling, yielding a dual-domain perceptual global aggregated feature sequence. The encoder, through a multi-head self-attention mechanism, fully captures the long-range spatial dependencies across regions and the global spectral correlations across bands in the hyperspectral image of the cantaloupe, compensating for the limitation of CNNs in extracting only local receptive field features. Simultaneously, by combining feedforward networks, residual connections, and layer normalization, stable training is achieved. This process enables the model to complete a global joint modeling of the spatial morphology, spectral response, and growth stage priors of the entire cantaloupe in a unified feature space, providing high-order features with global discriminative power for subsequent accurate assessment of growth status. The formulas used are as follows:
[0105] ;
[0106] ;
[0107] In the formula, and These are the dual-domain perceptual global aggregated feature sequences output by the (l-1)th and lth layer Transformer encoders, respectively. It is layer normalization. It is the self-attention of the bulls. It is the intermediate global feature of the l-th layer after self-attention, residual connection and layer normalization;
[0108] Step S35: Gated Feature Refinement; The dual-domain perception global aggregated feature sequence is reshaped into a two-dimensional feature map to restore spatial layout relationships and accurately repair leaf edges, fruit textures, subtle lesion areas, and local spectral abrupt changes lost in global feature modeling. Simultaneously, hyperspectral distortion is avoided. A parallel structure of spatial detail convolution and spectral smoothing convolution is adopted to extract spatial fine-grained features and spectral consistency features respectively. Then, a soft mask of cantaloupe growth topology is introduced to mask the features, forcing the model to focus only on effective physiological regions such as roots, stems, leaves, and fruits while suppressing background noise. Simultaneously, the one-hot encoding of the cantaloupe growth stage is dynamically modulated using global pooling and adaptive gating weights in the activation generation stage to generate a refined high-order feature map, providing more discriminative and stable features for subsequent growth status assessment. The formulas used are as follows:
[0109] ;
[0110] ;
[0111] ;
[0112] ;
[0113] ;
[0114] ;
[0115] In the formula, TD re It is a two-dimensional feature map. It is a dimensional reshaping operation, TD spa and TD spe These are spatial fine-grained features and spectral consistency features, respectively. Padding is the fill length, and TD... mask It is a mask feature map. It is global average pooling. It is a linear fully connected layer, g stage It is an adaptive gating weight during the growth phase, ADV refine It is a refined high-order feature map;
[0116] Step S36: Growth State Classification; Global average pooling is performed on the refined high-order feature map. Global mean aggregation is then applied to features at all locations in the spatial dimension to eliminate feature shifts caused by spatial location differences. Simultaneously, the global high-order semantic information of the entire cantaloupe plant in both the spectral and spatial domains is preserved, generating a global feature vector. This global feature vector is then sequentially processed through linear mapping and Dropout regularization to suppress overfitting. Finally, the Softmax activation function is used to obtain the probability distribution of the four growth states of the cantaloupe, outputting the predicted growth state label. During model training, the cross-entropy loss function is used to construct the optimization objective. By reducing the difference between the predicted probability distribution and the true labeled label, iterative updates of the network parameters are achieved. The formulas used are as follows:
[0117] ;
[0118] ;
[0119] ;
[0120] In the formula, EF global It is a global feature vector. It is the probability vector for predicting the growth status of cantaloupes. It is a random deactivation regularization operation. It is a normalized exponential activation function. It is the cross-entropy loss value, N b Here, z is the training batch size, z is the sample index within the batch, and c is the growth state category index. and These are the true label and predicted label of the z-th sample in class c, respectively.
[0121] By performing the above operations, this solution addresses the problems in existing methods for assessing the growth status of Hami melons, such as difficulty in simultaneously and efficiently representing spatial morphology and spectral information, difficulty in capturing long-range global spatial-spectral dependencies, and easy loss of subtle features, leading to inaccurate differentiation of various growth states and unstable assessment results. This solution extracts spatial and spectral feature maps to achieve independent and efficient representation of the spatial and spectral domains; generates weighted spectral feature maps, and obtains a comprehensive feature map by splicing them together, integrating spatial morphology and spectral physicochemical information to improve the comprehensive discriminative ability of features; introduces a Hami melon growth topology soft mask, injects dual-domain positional encoding and adds a growth status classification token, guides the global sequence of growth stages, synchronously perceives spatial layout and band order, incorporates prior knowledge of the growth period, and enhances the structural rationality of the sequence; obtains a dual-domain perceived global aggregated feature sequence through global feature modeling, realizing global joint modeling of the spatial morphology, spectral response, and prior knowledge of the growth stages of the entire Hami melon; extracts fine-grained spatial features and spectral consistency features respectively, generates refined high-order feature maps, and repairs detailed information; outputs growth status prediction labels, significantly improving the accuracy and stability of assessing different growth states of Hami melons.
[0122] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the assessment of the growth status of cantaloupe involves acquiring real-time hyperspectral images of cantaloupe, removing the first and last noise bands to obtain the effective spectral bands, and performing preprocessing and labeling of the growth stage. Based on step S2, cantaloupe image enhancement processing is performed to generate real-time hyperspectral enhanced images of cantaloupe, which are then input into the cantaloupe growth status assessment model constructed in step S3 for classification to obtain the growth status corresponding to the real-time hyperspectral image of cantaloupe, thus completing the assessment of the growth status of cantaloupe.
[0123] Example 6, see Figure 2 Based on the above embodiments, the artificial intelligence-based cantaloupe growth status assessment system provided by the present invention includes a cantaloupe image acquisition module, a cantaloupe image enhancement processing module, a cantaloupe growth status assessment model construction module, and a cantaloupe growth status assessment module.
[0124] The cantaloupe image acquisition module acquires historical hyperspectral images of cantaloupes, performs preprocessing and image annotation, and sends the data to the cantaloupe image enhancement processing module;
[0125] The cantaloupe image enhancement processing module obtains the corrected multi-band image through Gaussian function and positive value suppression strategy. It obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences. After adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of the growth stage is introduced to generate global enhancement features. Through spatial local window segmentation and spectral band grouping, fine characterization features are generated to obtain the cantaloupe hyperspectral enhanced image. The data is then sent to the cantaloupe growth status assessment model construction module.
[0126] The Hami melon growth status assessment model construction module extracts spatial feature maps and spectral feature maps, generates a weighted spectral feature map, and obtains a comprehensive feature map by splicing them together. It introduces a Hami melon growth topology soft mask, injects dual-domain position encoding and adds a growth status classification token, guides the global sequence of growth stages, obtains a dual-domain perception global aggregation feature sequence through global feature modeling, extracts spatial fine-grained features and spectral consistency features respectively, generates a refined high-order feature map, outputs growth status prediction labels, and sends the data to the Hami melon growth status assessment module.
[0127] The cantaloupe growth status assessment module enhances and classifies real-time hyperspectral images of cantaloupes to obtain the growth status corresponding to the real-time hyperspectral images of cantaloupes.
[0128] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0130] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. An artificial intelligence-based method for assessing the growth status of Hami melons, characterized by: The method includes the following steps: Step S1: Hami melon image acquisition; acquire historical hyperspectral images of Hami melons, and perform preprocessing and image annotation; Step S2: Hami melon image enhancement processing; a corrected multi-band image is obtained through Gaussian function and positive value suppression strategy. Gated modulation features and global scanning features are obtained through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences. After adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of the growth stage is introduced to generate global enhancement features. Fine characterization features are generated through spatial local window segmentation and spectral band grouping to obtain a hyperspectral enhanced image of Hami melon. Step S3: Construction of Hami melon growth status assessment model; extract spatial feature map and spectral feature map, generate weighted spectral feature map, obtain comprehensive feature map by splicing, introduce Hami melon growth topology soft mask, inject dual-domain position encoding and add growth status classification token, guide global sequence of growth stage, obtain dual-domain perception global aggregated feature sequence through global feature modeling, extract spatial fine-grained features and spectral consistency features respectively, generate refined high-order feature map, and output growth status prediction label; Step S4: Assessment of the growth status of cantaloupe; enhance and classify the real-time hyperspectral image of cantaloupe to obtain the growth status corresponding to the real-time hyperspectral image of cantaloupe.
2. The method for evaluating the growth status of Hami melons based on artificial intelligence according to claim 1, characterized in that: In step S2, the cantaloupe image enhancement process specifically includes the following steps: Step S21: Band-by-band illumination correction; For the preprocessed cantaloupe hyperspectral image, calculate the gray mean and standard deviation for each band, adaptively compensate for shadows in dark areas using a Gaussian function, and compress overexposed bright areas using a positive value suppression strategy. Finally, normalize the correction results and stitch them together along the band dimension to obtain the corrected multi-band image. Step S22: Dual-branch feature extraction; Three-dimensional convolution is used to downsample the corrected multi-band image. The downsampled features are extracted using two parallel branches with the same structure but independent parameters. Both branches use plant-aware decomposition convolution to obtain gated modulation features and global scanning features. Step S23: Topological rearrangement selective scanning; Perform topological rearrangement selective scanning on the global scanning features to reconstruct the global scanning features into four global sequences: spatial forward, spatial backward, spectral forward, and spectral backward. Then, independently calculate the adaptive scale parameters for each global sequence and perform state space transformation to achieve global unified feature mapping and obtain four-way global modeling features. Step S24: Stage-adaptive modulation enhancement; Step S25: Local dual-attention refinement; Step S26: Enhanced image generation; channel integration of the sum of global enhancement features and fine representation features is performed by convolution, and then bilinear interpolation upsampling is performed to generate a hyperspectral enhanced image of cantaloupe.
3. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 2, characterized in that: In step S24, the stage adaptive modulation enhancement involves performing global average pooling on the global scanning features, concatenating them with the one-hot encoding of the growth stage, and inputting them into a multilayer perceptron to generate growth stage adaptive weights. The gated modulation features are then aligned to the spatial size and number of channels of the global modeling features through convolution and bilinear interpolation. The aligned gated modulation features, global modeling features, and growth stage adaptive weights are then weighted and modulated sequentially using Hadamard product. The four weighted modulation features are summed and input into the plant perception decomposition convolution to generate global enhancement features.
4. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 2, characterized in that: In step S25, the local dual attention refinement involves spatially segmenting the global enhancement features into local windows and grouping them into spectral bands. Spatial local features and band group features are extracted separately. Spatial local attention weights and spectral channel attention weights are learned for each local window and each band group through independent small convolutional network branches. The learned weights are then expanded to the original feature map size through a broadcast mechanism. The two attention weights are then fused with the global enhancement features element-wise and normalized and smoothed. Finally, the feature refinement is completed through plant perception decomposition convolution to generate refined representation features.
5. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 1, characterized in that: In step S3, the construction of the cantaloupe growth status assessment model specifically includes the following steps: Step S31: Spatial-Spectral Dual Branch; The hyperspectral enhanced image of the cantaloupe is fed into a dual-branch CNN backbone network, including the EfficientNet-B0 branch and the ResNet-50 branch, to obtain spatial feature maps and spectral feature maps. Step S32: Channel attention splicing and fusion; The spatial dimension of the spectral feature map is compressed by global average pooling to generate a global description vector of the spectral channel. Then, the importance weight of each band is learned through two fully connected layers. The importance weight is weighted element-wise with the original spectral feature map in the channel dimension to obtain a weighted spectral feature map. The spatial feature map and the weighted spectral feature map are spliced in the channel dimension to generate a comprehensive feature map. Step S33: Topology-guided sequence encoding; Step S34: Global feature modeling; The global sequence guided by the growth stage is input into a 4-layer stacked Transformer encoder for global feature modeling to obtain a dual-domain perceptual global aggregated feature sequence; Step S35: Gating feature refinement; Step S36: Growth state classification; global average pooling is performed on the refined high-order feature map to generate a global feature vector. The global feature vector is then subjected to linear mapping and Dropout regularization to suppress overfitting. Finally, the probability distribution of the four growth states of cantaloupe is obtained through the Softmax activation function, and the predicted label of the growth state is output.
6. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 5, characterized in that: In step S33, the topology-guided sequence encoding is based on the spatial gradient and average gray level to generate a soft mask for cantaloupe growth topology. The comprehensive feature map is input into the growth topology-aware convolutional layer, and after block segmentation and projection processing, a growth topology-aware sequence is obtained. Then, dual-domain position encoding is injected, which simultaneously encodes spatial position information and spectral band order information to obtain a dual-domain position-aware sequence. A growth state classification token is added to the beginning of the sequence to obtain a global sequence for guiding the growth stage.
7. The method for evaluating the growth status of Hami melons based on artificial intelligence according to claim 5, characterized in that: In step S35, the gated feature refinement involves reshaping the dual-domain perceptual global aggregated feature sequence into a two-dimensional feature map. A structure with parallel spatial detail convolutional branches and spectral smoothing convolutional branches is adopted to extract spatial fine-grained features and spectral consistency features respectively. Then, a cantaloupe growth topological soft mask is introduced to mask the features. At the same time, the one-hot encoding of the cantaloupe growth stage is dynamically modulated by global pooling and adaptive gating weights in the activation generation stage to generate a refined high-order feature map.
8. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 1, characterized in that: In step S1, the cantaloupe image acquisition involves acquiring historical hyperspectral images of cantaloupe and performing preprocessing and image annotation. The preprocessing includes size unification, band-by-band grayscale image extraction, illumination equalization, and grayscale normalization. The image annotation includes labeling the growth stage and growth status.
9. The method for evaluating the growth status of cantaloupe based on artificial intelligence according to claim 1, characterized in that: In step S4, the assessment of the growth status of cantaloupe involves acquiring real-time hyperspectral images of cantaloupe, preprocessing and labeling the growth stages, performing image enhancement processing based on step S2 to generate real-time hyperspectral enhanced images of cantaloupe, and inputting them into the cantaloupe growth status assessment model constructed in step S3 for classification to obtain the growth status corresponding to the real-time hyperspectral images of cantaloupe, thus completing the assessment of the growth status of cantaloupe.
10. An artificial intelligence-based cantaloupe growth status assessment system, used to implement the artificial intelligence-based cantaloupe growth status assessment method as described in any one of claims 1-9, characterized in that: It includes a cantaloupe image acquisition module, a cantaloupe image enhancement and processing module, a cantaloupe growth status assessment model construction module, and a cantaloupe growth status assessment module; The cantaloupe image acquisition module acquires historical hyperspectral images of cantaloupes, performs preprocessing and image annotation, and sends the data to the cantaloupe image enhancement processing module; The cantaloupe image enhancement processing module obtains the corrected multi-band image through Gaussian function and positive value suppression strategy. It obtains gated modulation features and global scanning features through two parallel plant perception decomposition convolution branches. The global scanning features are reconstructed into four global sequences. After adaptive scaling and state space transformation, four-way global modeling features are obtained. One-hot encoding of the growth stage is introduced to generate global enhancement features. Through spatial local window segmentation and spectral band grouping, fine characterization features are generated to obtain the cantaloupe hyperspectral enhanced image. The data is then sent to the cantaloupe growth status assessment model construction module. The Hami melon growth status assessment model construction module extracts spatial feature maps and spectral feature maps, generates a weighted spectral feature map, and obtains a comprehensive feature map by splicing them together. It introduces a Hami melon growth topology soft mask, injects dual-domain position encoding and adds a growth status classification token, guides the global sequence of growth stages, obtains a dual-domain perception global aggregation feature sequence through global feature modeling, extracts spatial fine-grained features and spectral consistency features respectively, generates a refined high-order feature map, outputs growth status prediction labels, and sends the data to the Hami melon growth status assessment module. The cantaloupe growth status assessment module enhances and classifies real-time hyperspectral images of cantaloupes to obtain the growth status corresponding to the real-time hyperspectral images of cantaloupes.