Melon leaf powdery mildew phenotype classification method

Abstract

The application discloses a melon leaf powdery mildew phenotype classification method, and relates to the field of disease monitoring, comprising the following steps: collecting standardized plant samples, pathogenic materials, calibrated collection equipment and baseline environmental data sets, obtaining multi-time point and multi-modal original image data sets and synchronous environmental records; performing cleaning and preliminary feature extraction on the multi-time point and multi-modal original image data sets and the synchronous environmental records, obtaining a standardized feature data set, constructing and quantifying a composite topology network of leaf wax layer and villus layer, and obtaining an enhanced fingerprint data set; performing cross-modal generative causal alignment processing on the enhanced fingerprint data set and early physiological fingerprints, obtaining mechanism-level phenotype feature vectors after fusing topology barrier features and cross-modal alignment; and inputting the mechanism-level phenotype feature vectors into a classification model. The application realizes early, accurate and mechanism-explaining phenotype classification of melon powdery mildew.

Description

Technical Field

[0001] This invention relates to the field of disease monitoring, and in particular to a method for classifying the phenotypic characteristics of powdery mildew on cucurbit leaves. Background Technology

[0002] In the field of disease monitoring, automatic leaf disease identification technology based on multispectral, chlorophyll fluorescence, and RGB imaging has been widely applied. Existing typical schemes usually rely on multi-temporal image acquisition to extract features such as color, texture, and spectral indices, and combine this with machine learning models to classify leaf diseases such as powdery mildew. While this method can effectively identify diseases in the symptomatic stage, it essentially passively quantifies already formed lesions, making it difficult to achieve high-sensitivity detection during the latent or early infection stages of diseases, and it cannot reveal phenotypic differences caused by different resistance mechanisms.

[0003] The limitations of existing technologies mainly lie in their insufficient ability to elucidate the mechanisms of disease occurrence. The occurrence of powdery mildew in cucurbits involves spore attachment, germination, and early spread, which are closely related to micro-topological features such as leaf wax layer structure and pubescence distribution, as well as early physiological responses such as photosynthetic efficiency. Traditional methods lack quantitative characterization of the three-dimensional microstructure of the leaf surface and have failed to effectively establish a causal relationship between early physiological responses and later visible symptoms. This results in classification results that only address severity levels, failing to distinguish between structurally tolerant and physiologically resistant subtypes, thus limiting their application value in early screening and precise control of disease-resistant breeding. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a phenotypic classification method for powdery mildew on cucurbit leaves to solve the technical problems of insufficient analysis of the early mechanism of the disease and inability to distinguish resistance subtypes in the existing technology.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for classifying the phenotypic characteristics of powdery mildew on cucurbit leaves, which includes collecting standardized plant samples, pathogenic raw materials, calibrated collection equipment and baseline environmental datasets, and acquiring multi-time-point, multi-modal raw image datasets and synchronous environmental records. Cleaning and preliminary feature extraction are performed on the original image datasets and synchronous environmental records from multiple time points and multiple modalities to obtain a standardized feature dataset. A composite topological network of leaf wax layer and villi layer is constructed and quantized to obtain an enhanced fingerprint dataset. The enhanced fingerprint dataset and early physiological fingerprints were subjected to cross-modal generative causal alignment to obtain the mechanism-level phenotypic feature vector after fusing topological barrier features and cross-modal alignment. Input the mechanism-level phenotypic feature vector into the classification model to obtain the phenotypic classification results and mechanism label report of the basic disease level, refined subtype and mechanism contribution label; The phenotypic classification results and mechanism label reports were validated to generate a phenotypic classification report and decision-making recommendations for powdery mildew on cucurbit leaves.

[0007] As a preferred embodiment of the powdery mildew phenotypic classification method for cucurbit leaves described in this invention, the data acquisition device and baseline environmental dataset include: Representative cucurbit varieties were selected, including known disease-resistant and disease-susceptible control varieties, and cultured until the leaves were fully expanded to obtain standardized plant samples. The pathogen of powdery mildew in cucurbits was collected using the standardized plant samples, and standard inoculation materials were prepared or the pathogen raw materials were obtained using natural disease conditions. Portable multispectral imaging equipment, chlorophyll fluorescence imager, RGB camera and environmental sensors were prepared, and whiteboard reflectance calibration and baseline testing were performed to obtain standardized plant samples, diseased raw materials, calibrated collection equipment and baseline environmental datasets.

[0008] As a preferred embodiment of the powdery mildew phenotypic classification method for cucurbit leaves described in this invention, the original image dataset and synchronization environment record include: Based on the standardized plant sample, target leaves are selected from the upper functional leaves of the plant, and fixed measurement positions are marked on the target leaves. Based on standardized plant samples with fixed measurement locations, the diseased raw materials, the calibrated acquisition equipment, and baseline environmental datasets; Before and after inoculation, the calibrated acquisition device was used to simultaneously acquire visible light RGB images, multispectral images, chlorophyll fluorescence parameters, and corresponding environmental data of the target leaves marked with fixed measurement positions at multiple time points, resulting in a multi-time point, multi-modal raw image dataset and synchronous environmental records.

[0009] As a preferred embodiment of the phenotypic classification method for powdery mildew on cucurbit leaves described in this invention, the standardized feature dataset includes: Background removal, leaf region segmentation, and multimodal image registration are performed on visible light RGB images, multispectral images, and chlorophyll fluorescence parameter images from multi-time point and multimodal raw image datasets and synchronous environmental records. Based on the registered preprocessed image dataset, a preliminary feature dataset of conventional spatial texture features and basic spectral indices is constructed and obtained by extracting spatial statistical features describing the texture patterns of leaf surfaces and spectral combination indices reflecting the physiological state of leaves. Noise and outlier removal are performed on the preliminary feature dataset of conventional spatial texture features and basic spectral indices to obtain a standardized feature dataset.

[0010] As a preferred embodiment of the powdery mildew phenotypic classification method for cucurbit leaves described in this invention, the enhanced fingerprint dataset includes: Near-infrared sensitive band enhancement processing was performed on the multispectral images in the standardized feature dataset to separate the three-dimensional microstructure signal characterizing the wax crystals and villi distribution on the leaf surface, thus obtaining the microstructure signal of the leaf surface. Connectivity analysis and network modeling were performed on the microstructure signals of the blade surface to construct a wax-villi composite topology network that reflects the integrity of the wax layer coverage and the connection relationship of villi nodes, thus obtaining the wax-villi composite topology network structure. By analyzing the integrity of barrier paths and the uniformity of pore distribution in the waxy-velvety composite topological network structure, topologically robust and topologically fragile phenotypic subtypes are identified, and an enhanced fingerprint dataset of topological network barrier features is obtained.

[0011] As a preferred embodiment of the phenotypic classification method for powdery mildew on cucurbit leaves described in this invention, the mechanism-level phenotypic feature vector includes, Physiological response fingerprints were extracted from early chlorophyll fluorescence parameters associated with enhanced fingerprint datasets, and early physiological abnormal signals, including dynamic change patterns of maximum photochemical efficiency, were identified to obtain early physiological fingerprints. Construct a causal transfer network based on interpretable generative adversarial mechanisms; The generator encodes early physiological fingerprints as potential causal variables and generates structure-aware simulated visible phenotypic sequences based on the leaf surface structure constraints revealed by the topological network barrier features in the enhanced fingerprint dataset. The discriminator distinguishes between generated sequences and real phenotypic sequences based on adversarial training, obtains the causal rationality score of the generated sequence, and obtains the trained generative causal transfer model. The generative causal transfer model trained by early physiological fingerprint input is used to obtain a simulated visible phenotypic sequence that is compatible with the topological barrier features of the leaf surface. Combined with the causal rationality score given by the discriminator, the causal consistency between physiological response and visible phenotypic under specific structural constraints is evaluated to obtain the structure-aware causal alignment features. The structure-aware causal alignment features are fused with the topological network barrier features in the enhanced fingerprint dataset, and the physiological-structural coupling relationship represented by the latent causal variables of the generative causal transfer model is introduced to construct the mechanism-level phenotypic feature vector explained by the physiological-structural coupling mechanism. The mechanism-level phenotypic feature vector after fusing the topological barrier features and cross-modal alignment is obtained.

[0012] As a preferred embodiment of the phenotypic classification method for powdery mildew on cucurbit leaves described in this invention, the phenotypic classification results and mechanism label reports include, Based on mechanism-level phenotypic feature vectors, a lightweight ensemble classification model for basic disease level classification and refined subtype classification is trained. The mechanism-level phenotypic feature vectors are then input into the trained ensemble classification model to obtain the basic disease level and refined subtype classification results. Based on the contribution distribution of physiological-structural coupling mechanisms in the mechanism-level phenotypic feature vectors, the influence weights of different mechanisms on the primary classification results of basic disease severity and refined subtypes are calculated, and mechanism contribution labels are generated. The primary classification results of basic disease severity and refined subtypes are integrated with mechanism contribution labels to form a phenotypic classification result and mechanism label report.

[0013] As a preferred embodiment of the powdery mildew phenotypic classification method for cucurbit leaves described in this invention, the powdery mildew phenotypic classification report and decision recommendations for cucurbit leaves include: The basic disease level and refined subtype classification results in the phenotypic classification results and mechanism label reports were compared and verified with the manual visual grading results based on traditional RGB images during the same period to obtain the verification comparison results. Based on the verification and comparison results, a visual and data-driven phenotypic classification report of powdery mildew on cucurbit leaves is generated, which includes image annotations, phenotypic subtypes, and mechanism contribution labels. Then, combined with the physiological barrier mechanism and structural tolerance mechanism characteristics corresponding to different subtypes, targeted field control or breeding screening decision-making suggestions are formulated.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the phenotypic classification method for powdery mildew on cucurbit leaves as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the phenotypic classification method for powdery mildew on cucurbit leaves as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By collecting standardized plant samples, pathogenic raw materials, calibrated collection equipment, and baseline environmental datasets, multi-time-point multimodal images and environmental data are simultaneously collected, and standardized feature datasets are obtained through cleaning and preliminary feature extraction; thereby constructing and quantifying leaf wax content. A villous composite topological network is used to obtain an enhanced fingerprint dataset. Then, early physiological fingerprints and topological barrier features are fused through cross-modal generative causal alignment to construct a mechanism-level phenotypic feature vector. This feature vector is input into a classification model, which outputs classification results and reports containing basic disease severity, refined subtypes, and mechanism contribution labels. Finally, based on the validation results, a visual classification report and targeted decision-making suggestions are generated to achieve early, accurate, and mechanism-explained phenotypic classification of powdery mildew in cucurbits. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the phenotypic classification method for powdery mildew on cucurbit leaves. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0022] Reference Figure 1 This is one embodiment of the present invention, which provides a method for classifying the phenotypic characteristics of powdery mildew on cucurbit leaves, comprising the following steps: S1. Collect standardized plant samples, diseased raw materials, calibrated collection equipment, and baseline environmental datasets to obtain multi-time-point, multi-modal raw image datasets and synchronous environmental records.

[0023] S1.1 Select representative cucurbit varieties, including known disease-resistant and disease-susceptible control varieties, and cultivate them to the appropriate leaf stage when the leaves are fully expanded to obtain standardized plant samples.

[0024] Furthermore, in the preparation stage, representative cucurbit varieties such as cucumber, melon, and pumpkin are selected, including known disease-resistant and susceptible control varieties. The selected plant samples are cultivated under controlled environmental conditions until the leaves are fully expanded to the appropriate leaf stage, such as the three- to five-leaf stage or the mature plant stage. This process must ensure that the growth conditions of each plant sample are consistent, including light, temperature, humidity, water and fertilizer management, to eliminate phenotypic differences caused by non-disease factors, thereby obtaining standardized plant samples with uniform physiological state and developmental stage.

[0025] S1.2. Collect powdery mildew pathogens from cucurbits using the standardized plant samples, prepare standard inoculation materials or obtain pathogen raw materials using natural disease conditions.

[0026] Furthermore, using standardized plant samples, the pathogen of powdery mildew in cucurbits is collected or propagated through methods such as live leaf spore brushing or spore suspension spraying. The collected pathogen is then used to prepare standardized inoculum with uniform concentration and stable activity. Alternatively, standardized plant samples are placed in controlled natural environments that induce powdery mildew outbreaks to obtain pathogenic raw materials with clearly defined pathogenicity that can be reused. Obtaining the pathogen from standardized plant samples ensures the biological homology and adaptability between the pathogen and subsequent test samples. Whether preparing standardized inoculum or utilizing controlled natural disease conditions, the starting point and initial intensity of disease occurrence can be controlled or accurately recorded.

[0027] S1.3 Prepare portable multispectral imaging equipment, chlorophyll fluorescence imager, RGB camera and environmental sensors, and perform whiteboard reflectance calibration and baseline testing to obtain standardized plant samples, diseased raw materials, calibrated collection equipment and baseline environmental datasets.

[0028] Furthermore, portable multispectral imaging equipment, a chlorophyll fluorescence imager, an RGB camera, and environmental sensors for temperature, humidity, and light intensity were prepared. Before data acquisition, all imaging equipment underwent reflectivity calibration using a standard white board to eliminate the influence of differences in the equipment's own spectral response and fluctuations in lighting conditions; simultaneously, baseline testing and calibration were performed on the environmental sensors. Environmental parameters at the initial moment of acquisition were recorded, and finally, the calibrated acquisition equipment and baseline environmental dataset containing the initial environmental conditions were obtained.

[0029] S1.4. Based on the standardized plant sample, select target leaves from the upper functional leaves of the plant and mark the fixed measurement position on the target leaves.

[0030] Furthermore, based on standardized plant samples, functional leaves with active physiological functions and minimal shading effects were selected from the upper and middle parts of each plant as target leaves. On these selected target leaves, non-toxic, non-fading markers were used to clearly mark the fixed areas or points for subsequent image and data acquisition, ensuring that the same leaf location was tracked and observed throughout the entire experimental period. Selecting functional leaves from the upper and middle parts best reflects the overall physiological state of the plant and its systematic response to disease. This operation of marking and fixing measurement locations transforms macroscopic plant sample observation into longitudinal tracking of a fixed microscopic leaf area. This avoids spatial heterogeneity errors introduced by collecting data from different parts of the leaf at different time points, ensuring that data such as visible light RGB images, multispectral images, and chlorophyll fluorescence parameters collected at multiple time points strictly originate from the same small area of ​​the leaf.

[0031] S1.5, Based on standardized plant samples with fixed measurement locations, the diseased raw materials, the calibrated acquisition equipment, and the baseline environmental dataset.

[0032] Furthermore, the four core elements—standardized biological samples, standardized pathogens, standardized collection tools, and standardized environmental baselines—were logically aligned and physically prepared in both time and space. This integration ensured that subsequent collection work was conducted within an experimental framework where all variables were clearly defined, measured, and controlled. By correlating plant samples with fixed measurement locations with calibrated equipment, known pathogens, and recorded environmental baselines, any changes in the collected data can be more reliably attributed to the infection process of standardized pathogens at specific locations on standardized plants within a specific environmental context, enhancing the causal inference capability of the experiment. S1.6. At multiple time points before and after inoculation, the calibrated acquisition device is used to simultaneously acquire visible light RGB images, multispectral images, chlorophyll fluorescence parameters, and corresponding environmental data of the target leaves marked with fixed measurement positions, so as to obtain a multi-time point, multi-modal raw image dataset and synchronous environmental records.

[0033] Furthermore, a zero baseline was set before pathogen inoculation, and at multiple key time points after inoculation, calibrated acquisition equipment was used to synchronously collect data from target leaves marked with fixed measurement locations. The data acquisition included: acquiring visible light images using an RGB camera, acquiring spectral images covering specific wavelengths using a portable multispectral imaging device, acquiring fluorescence parameter images such as Fv / Fm using a chlorophyll fluorescence imager, and simultaneously recording real-time temperature, humidity, and light data using environmental sensors. Each acquisition was associated with a timestamp and plant identifier, forming a time-series aligned, multi-point, multimodal raw image dataset and synchronous environmental records.

[0034] S2. Clean and perform preliminary feature extraction on the original image dataset and synchronous environmental records of multiple time points and multiple modalities to obtain a standardized feature dataset. Construct and quantize the composite topological network of the leaf wax layer and villi layer to obtain an enhanced fingerprint dataset.

[0035] S2.1 Perform background removal, leaf region segmentation, and multimodal image registration on the original image dataset with multiple time points and multiple modalities, as well as the visible light RGB images, multispectral images, and chlorophyll fluorescence parameter images in the synchronous environmental records.

[0036] Furthermore, after acquiring the original image dataset and synchronous environmental records from multiple time points and modalities, preprocessing is required for the visible light RGB images, multispectral images, and chlorophyll fluorescence parameter images. The core tasks of preprocessing are to remove background interference, accurately locate the target leaf region, and ensure that image data from different sources are strictly aligned spatially. Background removal typically utilizes color space conversion and thresholding techniques, such as distinguishing leaves from backgrounds like culture media and flowerpots in the HSV color space based on the leaf's hue and saturation characteristics. Leaf region segmentation, building upon background removal, extracts complete and accurate single-leaf masks from the image through edge detection, region growing, or deep learning-based semantic segmentation methods. Multimodal image registration is more challenging because RGB cameras, multispectral imaging devices, and chlorophyll fluorescence imagers have different spatial resolutions, imaging fields of view, and spectral response characteristics. Image registration algorithms such as feature point matching, affine transformation, or perspective transformation are required. Using a high-resolution RGB image as a spatial reference, each band of the multispectral image and the chlorophyll fluorescence parameter image are registered with the RGB reference image respectively. This ensures that the morphological, spectral, and physiological information of the same leaf at the same time from different sensors can be mapped to the same pixel coordinate grid, resulting in a spatially aligned, registered preprocessed image dataset.

[0037] S2.2 Based on the registered preprocessed image dataset, by extracting spatial statistical features describing the texture patterns of the leaf surface and spectral combination indicators reflecting the physiological state of the leaf, a preliminary feature dataset of conventional spatial texture features and basic spectral indices is constructed and obtained.

[0038] Furthermore, based on the registered preprocessed image dataset, features describing leaf condition are extracted from both the spatial and spectral domains. In the spatial domain, conventional spatial texture features are primarily extracted. For example, for the green channel or grayscale image of an RGB image, a series of Haralick texture features based on the gray-level co-occurrence matrix, including contrast, energy, entropy, and correlation, are extracted. These features quantify the spatial distribution and interrelationships of image pixel grayscale, capturing texture changes such as leaf surface roughness, patchiness, and regularity caused by disease. In the spectral domain, a series of basic spectral indices are calculated based on the reflectance of specific bands in multispectral images. For example, the normalized vegetation index (NDI) is calculated using the reflectance of the near-infrared and red bands, the photochemical reflectance index is calculated using the reflectance of the 531nm and 570nm bands, and a modified red-edge normalized vegetation index is calculated using the reflectance of bands near the red edge. These spectral indices are validated combined indicators sensitive to physiological and biochemical parameters such as chlorophyll content, light use efficiency, and water status. All conventional spatial texture features and basic spectral indices extracted from the registered images of all time points and all modalities of a leaf sample are summarized and organized to form a preliminary feature dataset.

[0039] S2.3. Perform noise and outlier removal processing on the preliminary feature dataset of conventional spatial texture features and basic spectral indices to obtain a standardized feature dataset.

[0040] Furthermore, after obtaining the preliminary feature dataset containing conventional spatial texture features and basic spectral indices, data cleaning and standardization are necessary to improve data quality and model training stability. Noise and outlier removal are crucial steps. Outliers may originate from transient interference during image acquisition, sensor noise, or minor errors in the segmentation and registration process. Statistical methods can be used for identification and processing, such as using interquartile range (IQR) or Z-score methods to identify data points that significantly deviate from the overall feature distribution, and then selecting whether to remove, correct, or retain them based on the specific circumstances. Subsequently, data standardization is performed, with common practices including Z-score standardization or Min-Max normalization. Z-score standardization transforms the data for each feature into a distribution with a mean of 0 and a standard deviation of 1; Min-Max normalization linearly scales the data to a specific interval. Through these processes, the influence caused by differences in units and numerical ranges between different features can be eliminated, ensuring that all features have equal importance weights in subsequent model training. After noise and outlier removal and standardization, a standardized feature dataset is obtained.

[0041] S2.4. Near-infrared sensitive band enhancement processing is performed on the multispectral images in the standardized feature dataset to separate the three-dimensional microstructure signal that characterizes the wax crystals and villi distribution on the leaf surface, thus obtaining the microstructure signal of the leaf surface.

[0042] Furthermore, after obtaining the standardized feature dataset, the multispectral images are specifically processed to extract leaf surface microstructure information. Near-infrared sensitive bands in the multispectral images are enhanced. Near-infrared bands are sensitive to scattering of internal leaf structures and surface micro-geometric features. In practice, specific near-infrared bands sensitive to diffuse reflection of the wax layer and multiple scattering effects of villi can be selected. Through methods such as band ratio analysis, principal component analysis, or independent component analysis, signal components related to the distribution density and crystal morphology of leaf epidermal wax crystals, as well as the height and density of epidermal villi, are enhanced. Simultaneously, spectral interference caused by changes in internal biochemical components such as mesophyll tissue and chlorophyll content are suppressed. From the multispectral data, a two-dimensional image signal mainly reflecting the three-dimensional spatial distribution and morphological characteristics of the outermost layer of wax and villi is separated; this signal is the leaf surface microstructure signal.

[0043] Specifically, this method utilizes the unique scattering response mechanism of near-infrared spectroscopy to micrometer-scale surface geometry, rather than traditional spectral absorption characteristics. The smooth crystalline surfaces of the waxy layer and the fibrous structure of the villi produce unique light scattering patterns in the near-infrared band, resulting in subtle differences in reflected signals. Through targeted band enhancement and signal separation processing, such as amplifying the surface scattering contribution through specific band ratios or stripping internal biochemical signals through component analysis, the implicit information about the three-dimensional microstructure of the surface, hidden in the broadband spectral reflectance, is made explicit into an analyzable image. This step successfully transforms spectral data characterizing the internal physiological state of the leaf into a structural map that can indirectly depict the spatial distribution morphology of the outermost physical barrier, the waxy layer and villi.

[0044] S2.5. Perform connected component analysis and network modeling on the microstructure signal of the blade surface to construct a wax-villi composite topology network that reflects the integrity of the wax layer coverage and the connection relationship of villi nodes, and obtain the wax-villi composite topology network structure.

[0045] Furthermore, image segmentation and connected component analysis were performed on the microstructure signals of the leaf surface to identify high-intensity regions representing continuous waxy coverage and localized bright clusters representing villi formation points. High-intensity waxy coverage regions were modeled as surfaces or barrier regions in the network, and villi formation points were modeled as nodes. Based on the spatial adjacency relationships between these nodes and regions, a waxy-villi composite topological network was constructed. The network construction rules can be defined as follows: if villi nodes are spatially adjacent and determined to belong to the same cluster or are connected through microstructure signals, then connecting edges are established between the nodes; continuous waxy coverage regions are represented as low-permeability barrier paths connecting multiple nodes. By obtaining the network's topological properties, such as node degree distribution, clustering coefficient, path length, and the area, perimeter, number, and distribution of pores in the barrier regions, an abstract network model that can quantitatively reflect the integrity of the leaf waxy layer coverage and the connection relationship between villi nodes is finally obtained, namely, the waxy-villi composite topological network structure.

[0046] Specifically, this study combines connected component analysis from image processing with topological modeling from complex network science to abstract and quantify biologically important functional units of leaf surface microstructures. Instead of simply calculating the density or area of ​​wax or villi, it models the relationships between them as a network: villi nodes constitute potential spore anchoring points, while continuous waxy regions form barriers that inhibit spore movement and germination. Through network representation, the physical defense structure of the leaf surface is transformed into a series of computable topological indices. For example, a highly connected cluster of villi nodes surrounded by intact waxy barrier paths may form a local defensive island, making it difficult for pathogenic spores to spread on the leaf surface; conversely, a sparsely connected network with broken barrier paths may provide expansion channels.

[0047] S2.6 By analyzing the integrity of barrier paths and the uniformity of pore distribution in the waxy-velvet composite topological network structure, topologically robust and topologically fragile phenotypic subtypes are identified, and an enhanced fingerprint dataset of topological network barrier features is obtained.

[0048] Furthermore, based on the constructed waxy-villi composite topological network structure, topological analysis was performed to identify different resistance structural phenotypic subtypes. The analysis focused on two core topological properties: the integrity of barrier paths and the uniformity of pore distribution. Barrier path integrity analysis aimed to assess whether there were breaks or weak points in the paths in the network, which consist of continuous waxy regions designed to impede spore movement. This could be measured by calculating indicators such as the size of the largest connected barrier region and the ratio of barriers in the shortest paths between all nodes in the network. Pore distribution uniformity analysis focused on the size, number, and distribution of pore spaces in the network that allow spores to pass through and are not covered by wax or dense villi. This could be assessed by calculating the statistical distribution of pore area and spatial autocorrelation. A topologically robust phenotype may be characterized by highly intact barrier paths and small, uniformly distributed pores, forming a dense defense network; while a topologically fragile phenotype may be characterized by broken barrier paths, large and unevenly distributed pores, and obvious expansion channels. Through comprehensive analysis of the above topological attributes, a topological subtype label (such as tough or fragile) and a corresponding quantized feature vector are assigned to each leaf sample. These features, together with the previously extracted conventional features, constitute the enhanced fingerprint dataset.

[0049] Specifically, by defining and quantifying two topological indicators with clear biophysical significance—barrier path integrity and porosity uniformity—we can directly diagnose the Achilles' heel of the leaf surface microstructure defense system. For example, incomplete barrier paths indicate a defect in the continuity of the waxy layer, which may allow spores to spread rapidly along a specific direction; uneven porosity distribution indicates the existence of large, localized weak areas, facilitating spore colonization. By identifying topologically robust and vulnerable subtypes, we can reveal why and how their structural defenses are susceptible to disease. This subtype differentiation based on topological mechanisms provides breeders with the ability to screen varieties with specific superior structural traits and plant protection personnel with the ability to develop precise agronomic measures targeting different failure modes. S3. Perform cross-modal generative causal alignment processing on the enhanced fingerprint dataset and early physiological fingerprints to obtain the mechanism-level phenotypic feature vector after fusing topological barrier features and cross-modal alignment.

[0050] S3.1 Extract physiological response fingerprints from early chlorophyll fluorescence parameters associated with the enhanced fingerprint dataset, identify early physiological abnormal signals including the dynamic change pattern of maximum photochemical efficiency, and obtain early physiological fingerprints.

[0051] Furthermore, based on the enhanced fingerprint dataset, the focus is on chlorophyll fluorescence parameter image data correlated with the early stages of disease development. From these time-series images, feature sequences reflecting the dynamic changes in photochemical efficiency of the photosynthetic system are extracted to form a physiological response fingerprint. For marked leaf regions at fixed locations, the changes in key fluorescence parameters over time, such as maximum photochemical efficiency (Fv / Fm), actual photochemical efficiency (ΦPSII), and non-photochemical quenching (NPQ), are tracked and extracted. By analyzing the morphological characteristics of these curves, such as the starting point, rate of decline, and recovery trend of Fv / Fm, or the rapid rise and relaxation process of NPQ, early abnormal signals deviating from the dynamic pattern of healthy leaves are identified. The set of feature patterns distilled from the time-series fluorescence parameters, characterizing the early stress response of the photosynthetic apparatus to pathogen infection, constitutes the early physiological fingerprint.

[0052] Specifically, this emphasizes the patterns of parameter evolution over time, such as whether Fv / Fm decreases slowly and linearly or experiences a sudden, step-like drop, and whether NPQ initially rises and then falls or remains consistently high. Dynamic patterns contain rich information about the timing, intensity, and recovery capacity of plant defense systems, providing more sensitive and specific early disease indicators than static values. By extracting dynamic patterns as early physiological fingerprints, we effectively capture the traces left at the photosynthetic function level of the early molecular and cellular event chain from pathogen recognition to the initiation of physiological defenses. This provides a highly informative physiological causal characterization with a clear temporal sequence for establishing causal relationships between subsequent and later-appearing visible structural phenotypes.

[0053] S3.2 Construct a causal transfer network based on interpretable generative adversarial mechanisms.

[0054] The generator encodes early physiological fingerprints as potential causal variables and generates structure-aware simulated visible phenotypic sequences based on the leaf surface structure constraints revealed by the topological network barrier features in the enhanced fingerprint dataset.

[0055] Furthermore, to establish a mechanistic link between early physiological responses and later visible phenotypes, a causal transfer network based on the interpretable generative adversarial approach is constructed. The core of this network is a generator that takes early physiological fingerprints as input. The generator first maps the early physiological fingerprints to a low-dimensional latent space through an encoder. The variables in this latent space are expected to capture the underlying causal factors driving phenotypic development, and the waxy-villi composite topological network structural features extracted from the enhanced fingerprint dataset are used as conditional inputs. This conditional input explicitly provides specific structural constraints on the leaf surface. Based on a comprehensive representation that integrates early physiological information and leaf structural conditional information, the decoder generates a simulated, temporal sequence of visible phenotypic development, such as simulating the evolution of visible features like lesion area expansion and color changes, resulting in a structure-aware simulated visible phenotypic sequence.

[0056] Specifically, early physiological fingerprints are encoded as latent representations of causes, while topological network structural features are used as invariant scenario constraints. The generator is asked to answer a counterfactual question: given the structural defense context of a specific leaf, what is the most likely outcome of the observed early physiological stress patterns, leading to the development of visible symptoms? For example, the same early photosynthetic efficiency decline pattern might generate only small and limited simulated lesions on a topologically robust leaf due to structural barriers, while on a topologically fragile leaf, it might generate rapidly expanding simulated lesions. By forcing the generator to operate under given structural conditions, the model essentially learns the physical laws governing how physiological drivers and structural constraints determine phenotypic outcomes.

[0057] S3.3 The discriminator distinguishes between generated sequences and real phenotypic sequences based on adversarial training, obtains the causal rationality score of the generated sequence, and obtains the trained generative causal transfer model.

[0058] Furthermore, the discriminator and generator in the causal transfer network undergo adversarial training. The discriminator's input includes real visible phenotypic sequences extracted from multi-timepoint RGB images, and structure-aware simulated visible phenotypic sequences generated by the generator. The discriminator's task is to accurately distinguish whether the source of the input sequence is real or generated. During training, the generator continuously optimizes to generate increasingly realistic simulated sequences to deceive the discriminator, while the discriminator continuously optimizes to improve its discrimination ability. The ultimate goal of adversarial training is not simply to pursue the generator's perfect replication ability, but to make the simulated sequences generated by the generator under given structural and physiological conditions appear highly realistic to the discriminator. When training reaches equilibrium, the discriminator will give a high realism score to the simulated sequences output by the generator, which can be regarded as a causal plausibility score of the generated sequence, because it measures the degree of agreement between the generated effect under known causes and conditions and the effect observed in the real world, thus obtaining the trained generative causal transfer model.

[0059] Specifically, adversarial training is used as an implicit, data-driven tool for verifying causal consistency. The discriminator goes beyond traditional true / false classifiers; it implicitly learns the complex mapping distribution between early physiological and structural features and later visible phenotypes in the real world. To fool the discriminator, the generator must learn to generate sequences that conform to the true mapping distribution. The adversarial training process forces the generator to discover and internalize causal generation mechanisms that conform to biological and physical laws, rather than merely memorizing superficial associations in data. The discriminator's score for the generated sequence reflects the degree to which the sequence conforms to the learned causal mechanisms. High-scoring generated phenotypic sequences not only appear realistic, but more importantly, their cause-and-effect evolution is reasonable and consistent within the discriminator's understanding of the world model.

[0060] S3.4. The generative causal transfer model trained by inputting early physiological fingerprints is used to obtain a simulated visible phenotypic sequence that is compatible with the topological barrier features of the leaf surface. Combined with the causal rationality score given by the discriminator, the causal consistency between the physiological response and the visible phenotypic under specific structural constraints is evaluated to obtain the structure-aware causal alignment features.

[0061] Furthermore, the trained generative causal transfer model is used to analyze new samples. The early physiological fingerprint of the sample to be analyzed is input into the generator of the model, while the sample's topological network barrier features are used as conditional input. The generator outputs a sample-specific, structure-aware, simulated visible phenotypic sequence. A discriminator evaluates this generated sequence and outputs a causal plausibility score. Subsequently, this simulated phenotypic sequence is compared and analyzed with real visible phenotypic sequences extracted from multi-time-point RGB images actually collected from the sample. The comparison focuses not only on the similarity of the final state but also on the consistency of the entire developmental dynamic. Combining the causal plausibility score provided by the discriminator, the coherence and plausibility of the causal chain leading to the observed visible phenotypic development under the constraints of its specific leaf surface structure, based on the sample's actual physiological response pattern, are comprehensively evaluated. The evaluation produces a quantitative or structured feature representation, namely a structure-aware causal alignment feature, which encodes the strength of consistency among cause, condition, and effect.

[0062] Specifically, each sample undergoes a personalized, counterfactual reasoning-based assessment of causal strength. It doesn't simply stitch features together; it compares the model's predictions of how the disease would develop under this structural condition, starting from this early physiological state, with the actual development. A high causal plausibility score and a high degree of agreement between the simulated and actual sequences indicate that the disease development of this sample highly conforms to the mainstream causal mechanism learned by the model—physiologically driven and structurally regulated—and can be categorized as having high response-phenotype consistency. Conversely, a low score or poor agreement may suggest the existence of other unmodeled dominant factors, or that the causal path of this sample belongs to a minority of anomalous patterns. For example, a leaf might exhibit strong early physiological stress, but the final lesion is mild. The simulated prediction might show that the lesion should be milder under a strong structural barrier, while the actual lesion, though mild, is slightly more severe than the simulated prediction. This subtle difference is captured by the structure-aware causal alignment features, potentially indicating the existence of other subtle resistance mechanisms beyond structural defense.

[0063] S3.5. The structure-aware causal alignment features are fused with the topological network barrier features in the enhanced fingerprint dataset, and the physiological-structural coupling relationship represented by the latent causal variables of the generative causal transfer model is introduced to construct the mechanism-level phenotypic feature vector explained by the physiological-structural coupling mechanism, and the mechanism-level phenotypic feature vector after fusing the topological barrier features and cross-modal alignment is obtained.

[0064] Furthermore, information from different analytical paths is deeply integrated to construct a unified mechanism-level representation. Specifically, the causal alignment features perceived by the results are fused with features from the topological network barrier, which directly describes the physical defense state of the leaf. This fusion is not a simple concatenation but introduces a crucial piece of information: latent causal variables generated by the generator of the generative causal transfer model when producing simulated sequences. These latent causal variables are generated during the encoding of early physiological fingerprints and are interpreted as interpretable latent physiological causal factors driving phenotypic development. During fusion, the physiological driving force type and intensity information represented by these latent causal variables are correlated and combined with the structural constraint information represented by the topological network barrier features. For example, synergistic or antagonistic indices between specific types of physiological driving forces and specific modes of structural constraints are derived, constructing a novel feature vector. This feature vector not only simultaneously contains physiological, structural, and alignment information but also, through the bridging of latent causal variables, intrinsically encodes the mechanistic explanation of how different physiological driving forces, under different structural backgrounds, lead to different phenotypic outcomes. Ultimately, this forms a mechanism-level phenotypic feature vector that integrates topological barrier features with cross-modal alignment.

[0065] Specifically, latent variables in machine learning are given explicit biological mechanism explanations and used as a glue to fuse observational features from different modalities into a coherent narrative. By actively introducing and utilizing latent causal variables in generative causal transfer models, causal assumptions are implanted during the feature construction stage. The resulting mechanism-level phenotypic feature vectors have each dimension or combination of dimensions traceable to specific physiological-structural interaction hypotheses. For example, one part of the feature vector might characterize a rapid local necrosis tendency driven by strong oxidative stress combined with an incomplete waxy barrier; another part might characterize a slow atrophy tendency driven by photosynthetic inhibition combined with a highly porosity villous network. This means that the final input to the classification model is no longer isolated feature statistics, but a set of high-order features pre-organized and pre-interpreted according to the hypothesized pathogenic mechanisms.

[0066] S4. Input the mechanism-level phenotypic feature vector into the classification model to obtain the phenotypic classification results and mechanism label report of the basic disease level, refined subtype and mechanism contribution label.

[0067] S4.1. Based on the mechanism-level phenotypic feature vector, train a lightweight ensemble classification model for basic disease level classification and refined subtype classification. Input the mechanism-level phenotypic feature vector into the trained ensemble classification model to obtain the basic disease level and refined subtype classification results.

[0068] Furthermore, a lightweight ensemble classification model is trained using mechanism-level phenotypic feature vectors as training data. This model typically employs ensemble learning algorithms such as random forests or gradient boosting decision trees. During training, the mechanism-level phenotypic feature vectors are used as input features, and the basic disease level and refined subtype, calibrated manually or using high-precision methods, are used as prediction targets to construct a multi-task or multi-output classification model. The basic disease level usually corresponds to the severity level of the disease, while the refined subtype corresponds to mechanism-based subclasses such as physiological barrier-dominated or structural tolerance-dominated types. After the model training is complete, the mechanism-level phenotypic feature vectors of new, unclassified samples are input into this trained ensemble classification model. The model then simultaneously outputs the prediction of the basic disease level and the refined subtype classification results for that sample, completing the initial classification decision.

[0069] S4.2 Based on the contribution distribution of physiological-structural coupling mechanism in the mechanism-level phenotypic feature vector, calculate the influence weight of different mechanisms on the primary classification results of basic disease level and refined subtype, and generate mechanism contribution labels.

[0070] Furthermore, based on the prior definition of mechanism categories, each feature in the mechanism-level phenotypic feature vector is divided into different mechanism category sets. For example, features reflecting specific physiological driving force patterns are classified into the physiological barrier mechanism set, and features reflecting specific topological barrier defects are classified into the structural tolerance mechanism set. Each set corresponds to an index set. The importance coefficient of each feature in the classification decision is extracted from the trained ensemble classification model. This coefficient quantifies the contribution of the feature to the model in distinguishing different categories. Simultaneously, the mechanism-level phenotypic feature vector of the current sample is standardized to obtain the standardized feature value of each feature. This value reflects the significance of the feature deviating from the normal range in the current sample. The influence weight is calculated according to a given expression: for each mechanism category, the sum of the products of the importance coefficients of all features in its feature set and the standardized feature value of the feature is divided by the sum of such products across all mechanism categories. This yields the influence weight of the mechanism category on the current classification result. This weight value is between 0 and 1, representing the relative importance of the action path represented by the mechanism in the classification decision of the current sample. Based on the calculated influence weights of each mechanism, mechanism contribution labels are generated, such as physiological barrier dominance or structure-physiological synergy.

[0071] The weighting expression for the influence of the primary classification results is: in, For the first The influence weight of the primary classification results of this mechanism, For the first The importance coefficient of each feature. For mechanism-level phenotypic feature vectors, For the index of mechanism categories, for, The total number of mechanism categories, For loop variable, For index set.

[0072] S4.3 Integrate the primary classification results of basic disease severity and refined subtypes with mechanism contribution labels to form a phenotypic classification result and mechanism label report.

[0073] Furthermore, after obtaining the primary classification results of the basic disease severity level and refined subtype, as well as the mechanism contribution labels, the information is structurally integrated. These three types of information—the basic disease severity level representing the disease severity, the refined subtype representing the resistance mechanism type, and the relative contribution weights of each mechanism explaining the cause of the subtype (i.e., the mechanism contribution labels)—are associated and organized according to a pre-defined report template. The integration process ensures that the unique identifier of each sample corresponds to these three types of output results, forming a structured data record containing multi-dimensional information. The integrated data record is then output in a standardized format, forming a document that includes both the final classification conclusion and the mechanism explanation—namely, a phenotypic classification result and mechanism label report.

[0074] S5. Verify the phenotypic classification results and mechanism label report, and generate a phenotypic classification report and decision-making suggestions for powdery mildew on cucurbit leaves.

[0075] S5.1. The basic disease level and refined subtype classification results in the phenotypic classification results and mechanism label reports are compared and verified with the manual visual grading results based on traditional RGB images in the same period to obtain the verification comparison results.

[0076] Furthermore, experienced plant disease experts independently assessed the disease severity of leaves with the same markings using visual grading based on standard disease grading atlases from RGB images collected concurrently at multiple time points. This yielded a baseline disease severity level. The baseline disease severity levels and the corresponding baseline disease severity levels in the refined subtype classification results were then compared with the expert visual grading results on a sample-by-sample, time-point-by-time basis. The reliability of this method in disease severity grading was quantitatively evaluated by calculating evaluation indicators such as confusion matrix, overall accuracy, precision, recall, F1 score, and early detection accuracy. Simultaneously, cases with unique disease development timelines or spatial patterns identified in the expert assessments were cross-validated with the refined subtype classification results to assess the rationality of the mechanism subtype classification, forming a verification comparison between quantitative evaluation indicators and qualitative analysis.

[0077] S5.2 Based on the verification and comparison results, generate a visualized and data-driven phenotypic classification report of powdery mildew on cucurbit leaves, including image annotations, phenotypic subtypes, and mechanism contribution labels. Then, combine the physiological barrier mechanisms and structural tolerance mechanisms corresponding to different subtypes to formulate targeted field control or breeding screening decision-making suggestions.

[0078] Furthermore, by integrating the original images, classification results, and explanatory labels, a visualized phenotypic classification report of powdery mildew on cucurbit leaves is generated: a time-series map presented in a timeline format, containing original RGB, multispectral, and fluorescence pseudocolor images, with identified lesion areas, disease severity, and subtype labels superimposed on the images; detailed data presented in tabular or chart format, including quantitative disease indices at each time point, subtype classification, specific contribution weights and trends of each physiological-structural mechanism; and a summary text description. Based on the detailed subtype labels and the underlying mechanistic contributions, targeted decision-making recommendations are formulated. For example, for varieties or plants identified as topologically vulnerable and dominated by waxy barrier defects, breeders are advised to focus on screening for traits such as continuous and densely crystalline wax layers; for areas detected in the field with early characteristics of weakened physiological barriers, plant protection personnel are advised to apply induced resistance or protective fungicides in advance and pay attention to water and fertilizer balance to enhance plant physiological resistance. By linking mechanistic diagnoses with specific agronomic or breeding goals, actionable phenotypic classification reports and decision-making recommendations for powdery mildew on cucurbit leaves can be generated.

[0079] This embodiment also provides a computer device applicable to the classification method of powdery mildew phenotypes on cucurbit leaves, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the classification method of powdery mildew phenotypes on cucurbit leaves as proposed in the above embodiment.

[0080] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0081] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for classifying powdery mildew phenotypes of cucurbit leaves as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0082] In summary, this invention collects standardized plant samples, diseased raw materials, calibrated collection equipment, and baseline environmental datasets, simultaneously acquiring multi-time-point, multi-modal images and environmental data. After cleaning and preliminary feature extraction, a standardized feature dataset is obtained; subsequently, leaf wax content is constructed and quantified. A villous composite topological network is used to obtain an enhanced fingerprint dataset. Then, early physiological fingerprints and topological barrier features are fused through cross-modal generative causal alignment to construct a mechanism-level phenotypic feature vector. This feature vector is input into a classification model, which outputs classification results and reports containing basic disease severity, refined subtypes, and mechanism contribution labels. Finally, based on the validation results, a visual classification report and targeted decision-making suggestions are generated to achieve early, accurate, and mechanism-explained phenotypic classification of powdery mildew in cucurbits.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for classifying the phenotypic characteristics of powdery mildew on cucurbit leaves, characterized in that: include, We collected standardized plant samples, diseased raw materials, calibrated collection equipment, and baseline environmental datasets to obtain multi-time-point, multi-modal raw image datasets and synchronous environmental records. Cleaning and preliminary feature extraction are performed on the original image datasets and synchronous environmental records from multiple time points and multiple modalities to obtain a standardized feature dataset. A composite topological network of leaf wax layer and villi layer is constructed and quantized to obtain an enhanced fingerprint dataset. The enhanced fingerprint dataset and early physiological fingerprints were subjected to cross-modal generative causal alignment to obtain the mechanism-level phenotypic feature vector after fusing topological barrier features and cross-modal alignment. Input the mechanism-level phenotypic feature vector into the classification model to obtain the phenotypic classification results and mechanism label report of the basic disease level, refined subtype and mechanism contribution label; The phenotypic classification results and mechanism label reports were validated to generate a phenotypic classification report and decision-making recommendations for powdery mildew on cucurbit leaves.

2. The method for classifying powdery mildew phenotypes in cucurbit leaves as described in claim 1, characterized in that: The acquisition device and baseline environment dataset include, Representative cucurbit varieties were selected, including known disease-resistant and disease-susceptible control varieties, and cultured until the leaves were fully expanded to obtain standardized plant samples. The pathogen of powdery mildew in cucurbits was collected using the standardized plant samples, and standard inoculation materials were prepared or the pathogen raw materials were obtained using natural disease conditions. Portable multispectral imaging equipment, chlorophyll fluorescence imager, RGB camera and environmental sensors were prepared, and whiteboard reflectance calibration and baseline testing were performed to obtain standardized plant samples, diseased raw materials, calibrated collection equipment and baseline environmental datasets.

3. The method for classifying powdery mildew phenotypes in cucurbit leaves as described in claim 2, characterized in that: The original image dataset and synchronization environment record include, Based on the standardized plant sample, target leaves are selected from the upper functional leaves of the plant, and fixed measurement positions are marked on the target leaves. Based on standardized plant samples with fixed measurement locations, the diseased raw materials, the calibrated acquisition equipment, and baseline environmental datasets; Before and after inoculation, the calibrated acquisition device was used to simultaneously acquire visible light RGB images, multispectral images, chlorophyll fluorescence parameters, and corresponding environmental data of the target leaves marked with fixed measurement positions at multiple time points, resulting in a multi-time point, multi-modal raw image dataset and synchronous environmental records.

4. The method for classifying powdery mildew phenotypes on cucurbit leaves as described in claim 3, characterized in that: The standardized feature dataset includes, Background removal, leaf region segmentation, and multimodal image registration are performed on visible light RGB images, multispectral images, and chlorophyll fluorescence parameter images from multi-time point and multimodal raw image datasets and synchronous environmental records. Based on the registered preprocessed image dataset, a preliminary feature dataset of conventional spatial texture features and basic spectral indices is constructed and obtained by extracting spatial statistical features describing the texture patterns of leaf surfaces and spectral combination indices reflecting the physiological state of leaves. Noise and outlier removal are performed on the preliminary feature dataset of conventional spatial texture features and basic spectral indices to obtain a standardized feature dataset.

5. The method for classifying powdery mildew phenotypes on cucurbit leaves as described in claim 4, characterized in that: The enhanced fingerprint dataset includes, Near-infrared sensitive band enhancement processing was performed on the multispectral images in the standardized feature dataset to separate the three-dimensional microstructure signal characterizing the wax crystals and villi distribution on the leaf surface, thus obtaining the microstructure signal of the leaf surface. Connectivity analysis and network modeling were performed on the microstructure signals of the blade surface to construct a wax-villi composite topology network that reflects the integrity of the wax layer coverage and the connection relationship of villi nodes, thus obtaining the wax-villi composite topology network structure. By analyzing the integrity of barrier paths and the uniformity of pore distribution in the waxy-velvety composite topological network structure, topologically robust and topologically fragile phenotypic subtypes are identified, and an enhanced fingerprint dataset of topological network barrier features is obtained.

6. The method for classifying powdery mildew phenotypes on cucurbit leaves as described in claim 5, characterized in that: The mechanism-level phenotypic feature vector includes, Physiological response fingerprints were extracted from early chlorophyll fluorescence parameters associated with enhanced fingerprint datasets, and early physiological abnormal signals, including dynamic change patterns of maximum photochemical efficiency, were identified to obtain early physiological fingerprints. Construct a causal transfer network based on interpretable generative adversarial mechanisms; The generator encodes early physiological fingerprints as potential causal variables and generates structure-aware simulated visible phenotypic sequences based on the leaf surface structure constraints revealed by the topological network barrier features in the enhanced fingerprint dataset. The discriminator distinguishes between generated sequences and real phenotypic sequences based on adversarial training, obtains the causal rationality score of the generated sequence, and obtains the trained generative causal transfer model. The generative causal transfer model trained by early physiological fingerprint input is used to obtain a simulated visible phenotypic sequence that is compatible with the topological barrier features of the leaf surface. Combined with the causal rationality score given by the discriminator, the causal consistency between physiological response and visible phenotypic under specific structural constraints is evaluated to obtain the structure-aware causal alignment features. The structure-aware causal alignment features are fused with the topological network barrier features in the enhanced fingerprint dataset, and the physiological-structural coupling relationship represented by the latent causal variables of the generative causal transfer model is introduced to construct the mechanism-level phenotypic feature vector explained by the physiological-structural coupling mechanism. The mechanism-level phenotypic feature vector after fusing the topological barrier features and cross-modal alignment is obtained.

7. The method for classifying powdery mildew phenotypes on cucurbit leaves as described in claim 6, characterized in that: The phenotypic classification results and mechanism label reports include, Based on mechanism-level phenotypic feature vectors, a lightweight ensemble classification model for basic disease level classification and refined subtype classification is trained. The mechanism-level phenotypic feature vectors are then input into the trained ensemble classification model to obtain the basic disease level and refined subtype classification results. Based on the contribution distribution of physiological-structural coupling mechanisms in the mechanism-level phenotypic feature vectors, the influence weights of different mechanisms on the primary classification results of basic disease severity and refined subtypes are calculated, and mechanism contribution labels are generated. The primary classification results of basic disease severity and refined subtypes are integrated with mechanism contribution labels to form a phenotypic classification result and mechanism label report.

8. The method for classifying powdery mildew phenotypes in cucurbit leaves as described in claim 7, characterized in that: The report on the phenotypic classification of powdery mildew on cucurbit leaves and the decision-making recommendations include: The basic disease level and refined subtype classification results in the phenotypic classification results and mechanism label reports were compared and verified with the manual visual grading results based on traditional RGB images during the same period to obtain the verification comparison results. Based on the verification and comparison results, a visual and data-driven phenotypic classification report of powdery mildew on cucurbit leaves is generated, which includes image annotations, phenotypic subtypes, and mechanism contribution labels. Then, combined with the physiological barrier mechanism and structural tolerance mechanism characteristics corresponding to different subtypes, targeted field control or breeding screening decision-making suggestions are formulated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for classifying powdery mildew phenotypes of cucurbit leaves according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the phenotypic classification method for powdery mildew on cucurbit leaves as described in any one of claims 1 to 8.

Images