Intelligent monitoring method and system for full biological film degradation based on multi-source data acquisition

Abstract

The present invention discloses a method and system for intelligently monitoring the degradation of fully biofilm based on multi-source data acquisition, relating to the field of artificial intelligence technology. The method comprises: first, acquiring multi-source monitoring data for fully biofilm through multispectral sensing and image acquisition; extracting and fusing multimodal features for scenes containing film appearance images and multi-source sensing data of target films to generate target film degradation features; utilizing a pre-trained model's basic feature extractor to extract multi-resolution hierarchical film visual features, embedding these features with the target film degradation features to obtain cross-modal comprehensive monitoring features; and completing degradation status monitoring based on these features. This method overcomes the bottleneck of multi-source data fusion and full-resolution feature collaboration, accurately correlating visual cues with multi-source sensing quantitative information, improving the accuracy of degradation stage determination and regional analysis, and providing technical support for the refined management of fully biofilm.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and system for intelligent monitoring of full biological mulch degradation based on multi-source data acquisition. Background Art

[0002] Biofilm mulch is a key technology for green agricultural development, but traditional methods for monitoring its degradation status suffer from inefficiency, insufficient spatiotemporal precision, and inadequate multi-source data fusion. Existing technologies lack a targeted feature embedding mechanism for multimodal data, leading to ambiguous determination of degradation stages. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for intelligent monitoring of biofilm degradation based on multi-source data collection.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligently monitoring biofilm degradation based on multi-source data collection, comprising:

[0005] Obtain multi-source monitoring data for all biofilms;

[0006] If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensing data, a feature extraction operation is performed on the ground film appearance image and the target ground film multi-source sensing data to obtain a target ground film degradation feature, where the target ground film degradation feature is a fusion of features of the ground film appearance image and features of the target ground film multi-source sensing data;

[0007] Performing feature extraction of different resolutions on the mulch appearance image using a pre-trained basic feature extractor of a target mulch degradation monitoring model to obtain mulch visual features at multiple resolution levels, and performing feature embedding operations on the target mulch degradation features and the mulch visual features at multiple resolution levels to obtain cross-modal comprehensive monitoring features;

[0008] The degradation status of the full biological mulch is monitored according to the cross-modal comprehensive monitoring characteristics to obtain a degradation status monitoring result of the full biological mulch.

[0009] In a possible implementation, if the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensing data, performing a feature extraction operation on the ground film appearance image and the target ground film multi-source sensing data to obtain target ground film degradation features includes:

[0010] If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensor data, and the sensor signal carrier type of the target ground film multi-source sensor data includes a sensor vector atlas type, then selecting the atlas modal processing sub-model of the target ground film degradation monitoring model to perform feature extraction operations on the ground film appearance image and the target ground film multi-source sensor data, respectively, to obtain ground film appearance atlas features and sensor vector atlas features, respectively;

[0011] Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature;

[0012] The target mulch film degradation feature is obtained by performing feature space remapping on the cross-atlas fusion intermediate feature.

[0013] In a possible embodiment, the target mulch film degradation monitoring model further includes at least one cross-modal feature mapping network; and performing feature space remapping on the cross-atlas fusion intermediate features to obtain the target mulch film degradation features includes:

[0014] Loading the cross-atlas fusion intermediate features into the first feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain initial remapping features;

[0015] Performing feature interaction mapping processing on the initial remapped features according to a feature interaction operator to obtain cross-modal intermediate features;

[0016] The cross-modal intermediate features are loaded into the second feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain the target mulch film degradation features.

[0017] In one possible implementation, the atlas modal processing submodel includes a first feature space remapping unit and a second feature space remapping unit, and the atlas modal processing submodel further includes a cross-convolution correlation component; the atlas modal processing submodel for selecting the target mulch film degradation monitoring model performs feature extraction operations on the mulch film appearance image and the target mulch film multi-source sensor data, respectively, to obtain mulch film appearance atlas features and sensor vector atlas features, respectively, including:

[0018] Loading the ground film appearance image and the target ground film multi-source sensor data into the first feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain ground film appearance atlas features and sensor vector atlas features;

[0019] Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature, including:

[0020] Loading the ground film appearance map features and the sensor vector map features into the cross convolution correlation component to calculate the modal correlation between the ground film appearance map features and the sensor vector map features, and generating a dynamic fusion coefficient according to the modal correlation;

[0021] Performing spatial adaptive weighting on the ground film appearance map features and the sensor vector map features according to the dynamic fusion coefficient;

[0022] The coupled features are obtained by performing a feature coupling operation on the surface map features of the ground film and the sensor vector map features after spatial adaptive weighting;

[0023] The coupling features are loaded into the second feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain the cross-atlas fusion intermediate features.

[0024] In one possible implementation, the method further includes:

[0025] If the multi-source monitoring data includes a ground film appearance image, the ground film appearance image is loaded into the first feature space remapping unit of the single-source monitoring sub-model of the target ground film degradation monitoring model for feature processing;

[0026] Loading the features output by the first feature space remapping unit of the single source monitoring sub-model into the second feature space remapping unit of the single source monitoring sub-model for feature processing to obtain single source degradation features;

[0027] The target mulch film degradation characteristics are obtained by performing feature space remapping on the single-source degradation characteristics.

[0028] In one possible embodiment, the basic feature extractor includes a plurality of basic feature extraction components connected in sequence, and the target mulch film degradation monitoring model further includes a plurality of cross-modal feature mapping networks connected in sequence; the number of the cross-modal feature mapping networks is the same as the number of the basic feature extraction components; the basic feature extractor of the pre-trained target mulch film degradation monitoring model performs feature extraction of different resolutions on the mulch film appearance image to obtain mulch film visual features at multiple resolution levels, and performs feature embedding operation on the target mulch film degradation features and the mulch film visual features at multiple resolution levels to obtain cross-modal comprehensive monitoring features, including:

[0029] Loading the film appearance image into a first basic feature extraction component to perform feature extraction at a first resolution level to obtain film visual features at a first resolution level, and performing feature embedding processing on the film visual features at the first resolution level and the target film degradation features to obtain first cross-modal comprehensive monitoring features;

[0030] Loading the first cross-modal integrated monitoring feature into a next basic feature extraction component to perform feature extraction at a second resolution level to obtain ground film visual features at the second resolution level;

[0031] performing feature space remapping on the target mulch film degradation feature through a next cross-modal feature mapping network connected to the first cross-modal feature mapping network to obtain a deep mulch film degradation feature corresponding to the mulch film visual feature at the second resolution level; the target mulch film degradation feature is a feature output by the first cross-modal feature mapping network;

[0032] generating a first fusion coefficient of the ground film visual feature at the second resolution level according to the ground film visual feature at the second resolution level;

[0033] generating a second fusion coefficient corresponding to the deep film degradation characteristic according to the first fusion coefficient;

[0034] linearly superimposing the ground film visual feature at the second resolution level and the deep ground film degradation feature according to the first fusion coefficient and the second fusion coefficient to obtain a second cross-modal comprehensive monitoring feature;

[0035] The second cross-modal comprehensive monitoring feature is loaded into the next basic feature extraction component, and the feature embedding operation is re-executed until the last basic feature extraction component outputs the target scale ground film visual feature, and the cross-modal comprehensive monitoring feature is obtained based on the target scale ground film visual feature and the target ground film degradation feature.

[0036] In a possible implementation, monitoring the degradation status of the full biological mulch according to the cross-modal integrated monitoring feature to obtain the degradation status monitoring result of the full biological mulch includes:

[0037] Loading the cross-modal integrated monitoring features into the degradation analysis network of the target mulch film degradation monitoring model to perform degradation status monitoring;

[0038] The degradation stages and degradation area heat maps of the full biological mulch film output by the degradation analysis network are obtained, and the degradation status monitoring results are obtained according to the degradation stages and the degradation area heat maps.

[0039] In a possible implementation, obtaining multi-source monitoring data of the entire biofilm mulch includes:

[0040] In multiple spectral bands, the surface interface area of the entire biofilm is imaged separately by multispectral sensing nodes to obtain multiple multispectral images, where each multispectral image corresponds to a spectral band;

[0041] Obtaining a reflectance dataset corresponding to each multispectral image, and calculating a degradation deformation vector of each pixel based on the reflectance dataset corresponding to each multispectral image to generate a degradation deformation topology map corresponding to the surface interface area of the full biological mulch film, and determining the degradation deformation topology map as the multi-source sensing data of the target mulch film;

[0042] A ground film appearance image corresponding to the surface interface area is selected from a plurality of multispectral images, and multi-source monitoring data corresponding to the full biological ground film surface interface area is generated based on the ground film appearance image and the multi-source sensing data of the target ground film.

[0043] In a possible implementation, the target mulch film degradation monitoring model is obtained by the following method, including:

[0044] Acquire a multi-source monitoring data instance including a ground film appearance image instance and a ground film multi-source sensing data instance of the ground film monitoring sample, and full degradation cycle annotation data associated with the ground film monitoring sample;

[0045] Loading the multi-source monitoring data instance into the multi-source monitoring sub-model of the initial model, and loading the ground film appearance image instance into the single-source monitoring sub-model of the initial model, to obtain the corresponding ground film degradation feature instance;

[0046] Performing feature processing of different resolutions on the ground film appearance image instance by the basic feature extractor of the initial model to obtain ground film visual feature instances at multiple resolution levels, and performing feature embedding operation on the ground film degradation feature instance and the ground film visual feature instances at multiple resolution levels to obtain a cross-modal comprehensive monitoring feature instance;

[0047] Performing degradation state monitoring on the cross-modal integrated monitoring feature instance through the degradation analysis network of the initial model to obtain a degradation state monitoring result instance;

[0048] Error parameters are generated according to the degradation status monitoring result instance and the error of the full degradation cycle annotation data, and the network parameters of the initial model are optimized according to the error parameters to obtain the target mulch film degradation monitoring model.

[0049] In a second aspect, the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0050] Compared with existing technologies, the present invention provides the following beneficial effects: Using the disclosed method and system for intelligent monitoring of full-scale biofilm degradation based on multi-source data acquisition, multi-source monitoring data for full-scale biofilm is acquired through multispectral sensing and image acquisition; for scenes containing mulch appearance images and multi-source sensing data of target mulch, multimodal features are extracted and fused to generate target mulch degradation features; multi-resolution hierarchical mulch visual features are extracted using the basic feature extractor of the pre-trained model, and these features are embedded with the target mulch degradation features to obtain cross-modal comprehensive monitoring features; degradation status monitoring is completed based on these features. This method overcomes the bottleneck of multi-source data fusion and full-resolution feature collaboration, accurately associates visual cues with multi-source sensing quantitative information, improves the accuracy of degradation stage determination and regional analysis, and provides technical support for the refined management of full-scale biofilm. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0052] Figure 1 A schematic flow chart of the steps of a method for intelligent monitoring of biofilm degradation based on multi-source data acquisition provided by an embodiment of the present invention;

[0053] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0055] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0056] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of a method for intelligent monitoring of full-biological mulch degradation based on multi-source data collection provided by an embodiment of the present disclosure. The method for intelligent monitoring of full-biological mulch degradation based on multi-source data collection is introduced in detail below.

[0057] Step S201, obtaining multi-source monitoring data of the entire biological mulch;

[0058] Step S202: If the multi-source monitoring data includes a mulch film appearance image and multi-source sensing data of a target mulch film, a feature extraction operation is performed on the mulch film appearance image and the multi-source sensing data of the target mulch film to obtain a target mulch film degradation feature, where the target mulch film degradation feature is a fusion of features of the mulch film appearance image and features of the multi-source sensing data of the target mulch film.

[0059] Step S203: extracting features of the mulch appearance image at different resolutions using a pre-trained basic feature extractor of the target mulch degradation monitoring model to obtain mulch visual features at multiple resolution levels, and embedding the target mulch degradation features with the mulch visual features at multiple resolution levels to obtain cross-modal comprehensive monitoring features.

[0060] Step S204 , monitoring the degradation status of the full biological mulch film according to the cross-modal comprehensive monitoring characteristics to obtain a monitoring result of the degradation status of the full biological mulch film.

[0061] In an exemplary embodiment of the present invention, in an agricultural production scenario, the server, acting as the core computing node of an intelligent monitoring system for biofilm degradation, first acquires multi-source monitoring data on the biofilm when conducting intelligent monitoring of its degradation status. For example, a contiguous field of planted crops covered with biofilm is deployed with multispectral sensing nodes (e.g., drones equipped with multispectral cameras or multispectral imaging equipment fixed to the field). The server establishes communication connections with these sensing nodes via an Internet of Things (IoT) protocol. The server controls multispectral sensor nodes to image the surface interface of the biofilm in multiple spectral bands (such as the red, green, and blue channels of the visible light band and the near-infrared band), acquiring multiple multispectral images corresponding to different spectral bands. Simultaneously, the server obtains a reflectance dataset for each multispectral image (calculated by the sensor node's spectral analysis module based on the principle of ground object reflectance spectrum, reflecting the spectral reflectance characteristics of the film and the surrounding surface). Based on the reflectance dataset, the server calculates the degradation deformation vector for each pixel. By comparing the reflectance differences and texture changes of pixels at different times or bands, a vector space model of the degradation degree is constructed, generating a degradation deformation topology map of the biofilm surface interface (i.e., the multi-source sensor data of the target film). Furthermore, the server selects images from the multispectral images that clearly demarcate the film and soil and have a recognizable appearance as film appearance images. The server integrates the film appearance images with the multi-source sensor data of the target film to form multi-source monitoring data for the biofilm surface interface area of the farmland, completing data acquisition.

[0062] After the server confirms that the multi-source monitoring data includes the ground film appearance image and the target ground film multi-source sensor data, it executes the step of performing feature extraction operations on the ground film appearance image and the target ground film multi-source sensor data to obtain the target ground film degradation characteristics. Assuming that the sensor signal carrier of the target ground film multi-source sensor data is a sensor vector atlas type (such as a vector atlas converted from a degradation deformation topology map), the server calls the atlas modal processing sub-model of the pre-trained target ground film degradation monitoring model: First, the first feature space remapping unit of the atlas modal processing sub-model (composed of convolution, pooling, and normalization layers) performs feature extraction on the ground film appearance image and the target ground film multi-source sensor data respectively to obtain the ground film appearance atlas features (covering spatial features such as ground film texture, color, shape, etc.) and sensor vector atlas features (reflecting the spatial distribution and vector association of degradation deformation); then, the two types of atlas features are input into the cross-convolution correlation component, which calculates the modal association between the two (such as feature dimension similarity) and generates a dynamic fusion coefficient (if the ground film appearance color is large, the atlas will be used as the model). The color lightening is strongly correlated with the increase in the amplitude of the sensor deformation vector, and the corresponding regional fusion coefficient has a high weight). The two types of atlas features are spatially adaptively weighted based on the dynamic fusion coefficient, and then the coupled features are obtained through feature coupling operations (such as element-by-element multiplication, channel cascade and convolution); then, the coupled features are input into the second feature space remapping unit to extract the features and obtain cross-atlas fusion intermediate features; the cross-atlas fusion intermediate features are then passed through the first feature space remapping unit of the cross-modal feature mapping network (the fully connected layer adjusts the dimension) to obtain the initial remapping features, and the cross-modal intermediate features are obtained through feature interaction operators (such as the attention module focuses on key degradation features). Finally, the remapping is completed by the second feature space remapping unit to generate the target film degradation features that integrate the film appearance and multi-source sensor features. If the multi-source monitoring data only contains the appearance image of the ground film (simple monitoring scenario), the server calls the single-source monitoring sub-model: the ground film appearance image is first processed by the first feature space remapping unit, and the output features are then processed by the second feature space remapping unit to obtain the single-source degradation features. Finally, the single-source degradation feature space is remapped (linear transformation layer) to obtain the target ground film degradation features, ensuring the effectiveness of feature generation under different acquisition scenarios.

[0063] After obtaining the target film degradation feature, the server performs feature extraction at different resolutions on the film appearance image using the pre-trained basic feature extractor of the target film degradation monitoring model to obtain film visual features at multiple resolutions. The server then performs a feature embedding operation on the target film degradation feature and the multi-resolution film visual features to obtain a cross-modal integrated monitoring feature. The basic feature extractor of the target film degradation monitoring model consists of multiple sequentially connected basic feature extraction components (e.g., a ResNet-like residual block structure) and is equipped with an equal number of cross-modal feature mapping networks. Taking a model with three basic feature extraction components and three cross-modal feature mapping networks as an example, the server first inputs the film appearance image into the first basic feature extraction component. Through convolution kernel adjustment and step-size control downsampling, the server extracts film visual features at the first resolution level (preserving the low-resolution features of the film's overall outline and large-scale texture). Simultaneously, the target film degradation feature and the first-resolution visual features are input into the feature embedding module of the first cross-modal feature mapping network. Correlations between channels and spatial dimensions are calculated to generate fusion weights. After weighted fusion, the convolution operation is performed to obtain the first cross-modal integrated monitoring feature. Next, the first cross-modal integrated monitoring feature is fed into the next basic feature extraction component, which further downsamples and extracts second-resolution visual features (focusing on local film details, such as minor damage). The target film degradation feature output by the first cross-modal network is fed into the next cross-modal network, where it undergoes dimensionality adjustment using a multi-layer perceptron (MLP) to produce a deep film degradation feature that matches the second-resolution visual feature. The server generates a first fusion coefficient based on the local saliency of the second-resolution visual feature (e.g., calculating edge density using the Sobel operator and determining key areas using attention weights). The server then generates a second fusion coefficient for the deep film degradation feature based on its distribution. The two features are then linearly superimposed to obtain the second cross-modal integrated monitoring feature. This process is cyclical: the second cross-modal feature is input into the next basic component, and feature extraction, cross-modal remapping, fusion coefficient calculation and superposition are repeated until the last basic component outputs the target-scale visual features of the ground film (pixel-level depiction of molecular-level degradation traces), which are finally fused with the target ground film degradation features after processing by the last cross-modal network to obtain the cross-modal comprehensive monitoring features. This feature integrates the multi-resolution visual information of the ground film appearance from macro to micro and the degradation quantitative characteristics of multi-source sensing, providing comprehensive support for subsequent analysis.

[0064] Finally, the server monitors the degradation status of the biofilm based on the cross-modal integrated monitoring features, obtaining degradation status monitoring results. The server inputs the cross-modal integrated monitoring features into a degradation analysis network (composed of a fully connected layer, a convolutional attention layer, and a classification and regression branch). First, global feature pooling extracts global statistical information, while the convolutional attention layer focuses on key degradation areas. The classification branch uses a softmax classification based on the "undegraded / early / mid-stage / late-stage" labels to output the degradation stage. The regression branch uses a fully convolutional network to generate a heat map of the degradation areas (with color depth corresponding to the degree of degradation, red areas showing heavy degradation and blue areas showing light degradation). The server integrates the degradation stages with the heat map to generate monitoring results (e.g., "The biofilm is in the mid-stage of degradation, with significant degradation at the edges and locally damaged areas, and slower degradation in the middle areas") and pushes them to farmland managers' terminals (mobile apps, smart screens) via the IoT platform, providing a basis for agricultural decision-making.

[0065] During the model training phase, the server also serves as the core: it collects multi-source monitoring data instances containing mulch monitoring plots (mulch appearance images, multi-source sensor data) and full degradation cycle annotated data (manual sampling to detect mulch molecular weight, damage rate annotated stages and areas); inputs multi-source data into the initial model multi-source monitoring sub-model, and inputs images into the single-source monitoring sub-model to obtain mulch degradation feature instances; the basic feature extractor processes the image into multi-resolution visual feature instances, and the two are embedded in each other to generate cross-modal comprehensive monitoring feature instances; the degradation analysis network outputs the monitoring result instance, and the server calculates its error with the annotated data (classification cross entropy, heat map mean square error), back-propagates to optimize the initial model parameters (convolution kernel weights, fully connected layer bias), iterates until the loss converges, and obtains the pre-trained target mulch degradation monitoring model to ensure the accuracy and robustness of actual monitoring.

[0066] In summary, this method uses the server as the execution body, integrates multi-dimensional information through multi-source data collection, and uses modal fusion, multi-resolution feature extraction and cross-modal embedding technology to achieve accurate and intelligent monitoring of the degradation status of all-biological mulch, providing an efficient technical solution for the scientific management of all-biological mulch in agricultural production. It can be widely used in farmland, orchards and other scenarios, and promote the development of agriculture towards intelligence and ecology.

[0067] In an embodiment of the present invention, if the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensing data, a feature extraction operation is performed on the ground film appearance image and the target ground film multi-source sensing data to obtain target ground film degradation characteristics, which can be implemented through the following example.

[0068] If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensor data, and the sensor signal carrier type of the target ground film multi-source sensor data includes a sensor vector atlas type, then selecting the atlas modal processing sub-model of the target ground film degradation monitoring model to perform feature extraction operations on the ground film appearance image and the target ground film multi-source sensor data, respectively, to obtain ground film appearance atlas features and sensor vector atlas features, respectively;

[0069] Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature;

[0070] The target mulch film degradation feature is obtained by performing feature space remapping on the cross-atlas fusion intermediate feature.

[0071] In an embodiment of the present invention, for example, in a scenario where a fully biofilmed mulch is monitored in a certain experimental field, the server first confirms that the multi-source monitoring data includes an image of the mulch's appearance (showing the mulch's color, texture, and damaged outline) captured by a high-definition camera in the field and a degradation deformation topology map generated based on multispectral reflectance calculations (i.e., multi-source sensing data of the target mulch, whose sensing signal carrier is a sensing vector atlas type, with each pixel vector representing the degree of degradation deformation at that location). The server then selects the atlas modal processing sub-model of the pre-trained target mulch degradation monitoring model and inputs the mulch's appearance image and degradation deformation topology map into the sub-model's first feature space remapping unit (composed of a convolutional layer, a batch normalization layer, and an activation function). For the mulch appearance image, the first feature space remapping unit uses a 3×3 convolution kernel to extract spatial features such as mulch color gradients and damaged edges. After pooling and dimensionality reduction, the mulch appearance map features are obtained. For the degradation deformation topology map, this unit uses a 1×1 convolution based on the direction and amplitude of the deformation vectors of each pixel in the vector map to capture spatial correlation patterns and generate sensor vector map features. After feature extraction, the server inputs the mulch appearance map features and sensor vector map features into the cross-convolution correlation component of the spectral modal processing sub-model. This component first calculates the modal correlation between the two features in the channel and spatial dimensions (for example, the degree of overlap between lighter color areas in the mulch appearance image and increased vector amplitude areas in the degradation deformation topology map). Based on this correlation, a fusion coefficient matrix is dynamically generated. Subsequently, the two spectral features are spatially adaptively weighted according to the fusion coefficients (higher correlation areas are given higher weights). Feature coupling is then performed through channel-by-channel convolution and feature concatenation to obtain preliminary coupled features. The preliminary coupled features are fed into the second feature space remapping unit (composed of depthwise separable convolution and fully connected layers) of the spectral modal processing sub-model. After dimensionality adjustment and feature extraction, cross-atlas fusion intermediate features are output. Finally, the server loads the cross-atlas fusion intermediate features into the cross-modal feature mapping network of the target mulch film degradation monitoring model. The first feature space remapping unit (multi-layer perceptron structure) of this network first performs dimensionality conversion on the intermediate features to obtain initial remapped features. It then uses a feature interaction operator (attention mechanism module) to focus on key features of mulch film degradation (such as appearance spots and deformation vector mutation points corresponding to molecular chain breaks), performs feature interaction mapping on the initial remapped features, and generates cross-modal intermediate features. The cross-modal intermediate features are then fed into the second feature space remapping unit (composed of linear transformation layers and regularization layers) to complete the final feature space remapping, obtaining the target mulch film degradation features that fuse the visual features of the mulch film appearance with the deformation features of multi-source sensors, providing core feature support for subsequent degradation status monitoring.

[0072] In an embodiment of the present invention, the target mulch film degradation monitoring model further includes at least one cross-modal feature mapping network; the feature space remapping of the cross-atlas fusion intermediate features to obtain the target mulch film degradation features can be implemented through the following examples.

[0073] Loading the cross-atlas fusion intermediate features into the first feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain initial remapping features;

[0074] Performing feature interaction mapping processing on the initial remapped features according to a feature interaction operator to obtain cross-modal intermediate features;

[0075] The cross-modal intermediate features are loaded into the second feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain the target mulch film degradation features.

[0076] In an embodiment of the present invention, for example, in a scenario where the degradation of a fully biological mulch film in an orchard is monitored, the server has obtained a cross-atlas fusion intermediate feature (the feature fuses the color texture features of the mulch film appearance image and the degradation deformation vector features of the multi-source sensor data). The server calls the cross-modal feature mapping network of the target mulch film degradation monitoring model, and first loads the cross-atlas fusion intermediate feature into the first feature space remapping unit. This unit is composed of a multi-layer perceptron (MLP) and a batch normalization layer. The MLP adjusts the feature dimension through linear transformation (such as mapping 256-dimensional features to 512 dimensions), and the batch normalization layer stabilizes the numerical distribution, thereby obtaining the initial remapping feature. This feature not only retains the fusion information of the original cross-atlas feature, but also expands the feature expression dimension to adapt to subsequent interactive operations. Next, the server calls the feature interaction operator (a module based on the attention mechanism) to process the initial remapping feature. The attention mechanism module first calculates the attention weights for each feature dimension (for example, areas with appearance spots corresponding to breakage in the film's molecular chains and areas with sudden changes in the deformation vector are assigned high weights due to their strong correlation with degradation). Based on these weights, the initial remapped features are then weighted and aggregated, and nonlinearly transformed, to highlight the correlations between key degradation features. This generates cross-modal intermediate features, which strengthen the interaction between visual cues from the film's appearance and quantitative cues from multi-source sensors. For example, the visual feature of "partial whitening of the film" is deeply bound to the sensory feature of "sudden changes in reflectivity in that area." Finally, the server loads the cross-modal intermediate features into the second feature space remapping unit of the cross-modal feature mapping network. This unit consists of a convolutional layer and a dropout layer. The convolutional layer further refines features using a 1×1 convolution kernel (e.g., filtering noise and enhancing discriminative features of the degradation stage). The dropout layer prevents overfitting, ultimately outputting the target film degradation features. This feature accurately integrates the spatial morphological information of the ground film appearance and the quantitative degradation information of multi-source sensing, providing a unified and highly correlated feature representation for subsequent multi-resolution visual feature embedding and degradation status analysis, ensuring the effective coordination of different modal information during the monitoring process.

[0077] In an embodiment of the present invention, the atlas modal processing sub-model includes a first feature space remapping unit and a second feature space remapping unit; the atlas modal processing sub-model that selects the target ground film degradation monitoring model performs feature extraction operations on the ground film appearance image and the target ground film multi-source sensor data, respectively, to obtain the ground film appearance atlas features and sensor vector atlas features, respectively, which can be implemented through the following examples.

[0078] Loading the ground film appearance image and the target ground film multi-source sensor data into the first feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain ground film appearance atlas features and sensor vector atlas features;

[0079] Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature, including:

[0080] A feature coupling operation is performed on the film appearance map features and the sensor vector map features to obtain coupled features. The feature coupling adopts the 'channel splicing + 1×1 convolution' operation: the weighted film appearance map features (256 channels × 64 × 64) and the sensor vector map features (256 channels × 64 × 64) are first spliced into 512-channel features according to the channel dimension, and then compressed to 256 channels using a 1×1 convolution kernel, achieving deep interaction and redundant filtering of cross-modal information.

[0081] The coupling features are loaded into the second feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain the cross-atlas fusion intermediate features.

[0082] In an embodiment of the present invention, for example, in a scenario where the degradation of a contiguous stretch of farmland is monitored for fully biological mulch film, the server first confirms that the multi-source monitoring data includes an image of the mulch film's appearance (clearly showing the mulch film's surface color distribution, damaged texture, and boundary morphology with the soil) captured by a fixed-point high-definition camera in the field, as well as multi-source sensing data of the target mulch film generated based on multispectral reflectance analysis (i.e., a degradation deformation topology map, which records the degree of deformation caused by molecular chain breakage in each region of the mulch film in the form of a vector map). The server invokes the atlas modal processing submodel of the target mulch film degradation monitoring model and synchronously loads the mulch film's appearance image and the degradation deformation topology map into the submodel's first feature space remapping unit. This unit consists of a series of 3×3 convolutional layers, a maximum pooling layer, and a ReLU activation function. For the mulch appearance image, the convolutional layer captures spatial features such as "local white spots" and "edge damage outlines." The pooling layer compresses the dimensions while retaining key structural information, outputting a 256×64×64 dimensional mulch appearance map feature. For the degradation deformation topology map, the convolutional layer analyzes quantitative features in the vector map, such as "deformation vector amplitude gradient" and "regional deformation direction consistency." The pooling layer focuses on areas of significant deformation and outputs dimensionally matching sensor vector map features, thus completing the extraction of the two types of map features. The server then performs feature coupling on the extracted mulch appearance map features and sensor vector map features. During the coupling process, the server first concatenates the two feature types by channel (for example, the 256-channel mulch appearance feature and the 256-channel sensor vector feature are concatenated into 512 channels). A 1×1 convolution kernel is then used to perform cross-channel information interaction (strengthening the association between "white patches" and "high deformation vector regions"), generating a coupled feature with dimensions of 256×64×64. This feature carries information about the association between visual cues of mulch appearance and quantitative cues from multi-source sensors. Finally, the server loads the coupled feature into the second feature space remapping unit of the spectral modal processing submodel. This unit consists of a depthwise separable convolutional layer and a batch normalization layer: the depthwise separable convolutional layer first extracts details of the coupled features channel by channel (such as strengthening the binding relationship between the appearance texture and deformation vector mutation at the break of the molecular chain), and the batch normalization layer stabilizes the feature value distribution. The final output is a cross-atlas fusion intermediate feature with a dimension of 128×32×32. This feature deeply integrates the core information of the appearance morphology of the ground film and multi-source sensor deformation, providing a unified and closely related feature foundation for the subsequent processing of the cross-modal feature mapping network.

[0083] In an embodiment of the present invention, the atlas modal processing sub-model also includes a cross-convolution correlation component; the feature coupling operation performed on the ground film appearance atlas feature and the sensor vector atlas feature to obtain the coupling feature can be implemented through the following example.

[0084] Loading the ground film appearance map features and the sensor vector map features into the cross convolution correlation component to calculate the modal correlation between the ground film appearance map features and the sensor vector map features, and generating a dynamic fusion coefficient according to the modal correlation;

[0085] Performing spatial adaptive weighting on the ground film appearance map features and the sensor vector map features according to the dynamic fusion coefficient;

[0086] The coupling feature is obtained by performing a feature coupling operation on the ground film appearance atlas feature and the sensor vector atlas feature after spatial adaptive weighting.

[0087] In an embodiment of the present invention, for example, in a scenario of monitoring the degradation of a standardized farmland full-biological mulch film, the server has obtained the appearance atlas features of the mulch film that carry the "local white spots and edge damaged textures" of the mulch film through the first feature space remapping unit of the atlas modal processing sub-model, as well as the sensor vector atlas features that record the "deformation vector distribution caused by molecular chain breakage". At this point, the server loads the two types of features into the cross-convolution association component of the atlas modal processing sub-model: this component has a built-in dual-branch convolution kernel (a combination of 3×3 and 1×1), and first extracts local features from the "white patch edges" of the film appearance atlas features and the "deformation vector mutation area" of the sensor vector atlas features, and calculates the modal correlation of each pixel position using the cosine similarity algorithm (for example, the correlation between the white area with the RGB channel mean < the threshold value in the film appearance and the high deformation area with the L2 norm > the threshold value in the sensor vector atlas can reach 0.85); based on the correlation matrix, the component generates a dynamic fusion coefficient matrix through the Softmax function (for pixel positions with high correlation, the fusion coefficient is tilted towards "enhancing the contribution of dual features", for example, at a correlation of 0.85, the film appearance feature coefficient is set to 0.55 and the sensor feature coefficient is set to 0.45). Next, the server adaptively weights the two types of atlas features spatially based on the dynamic fusion coefficient. For each pixel in the mulch appearance atlas, element-by-element multiplication is performed by the corresponding fusion coefficient (e.g., 0.55), strengthening the weighting between "white patches" and "high deformation." Simultaneously, a coefficient multiplication (e.g., 0.45) is applied to the pixels in the sensor vector atlas to ensure that bimodal information in closely related areas is emphasized. After weighting, the texture gradients of the "white patches" in the mulch appearance atlas and the amplitude changes of the "deformation vectors" in the sensor vector atlas are precisely aligned in the spatial dimension. Finally, the server performs feature coupling operations on the two types of features after spatial adaptive weighting: first, the weighted ground film appearance map features (256 channels × 64 × 64 resolution) and the sensor vector map features (256 channels × 64 × 64 resolution) are spliced into 512-channel features according to the channel dimension; then, cross-channel information fusion is performed through a 1 × 1 convolution kernel (the number of channels is compressed to 256), focusing on retaining strong correlation features such as "white patch edge-deformation vector direction consistency" and "damaged texture density-deformation amplitude gradient", and finally generating a coupled feature with a dimension of 256 × 64 × 64. This feature deeply binds the visual clues of the ground film appearance and the multi-source sensor quantitative clues, providing a strong correlation basis for the feature extraction of the subsequent second feature space remapping unit.

[0088] In the embodiments of the present invention, the following implementation modes are also provided.

[0089] If the multi-source monitoring data includes a ground film appearance image, the ground film appearance image is loaded into the first feature space remapping unit of the single-source monitoring sub-model of the target ground film degradation monitoring model for feature processing;

[0090] Loading the features output by the first feature space remapping unit of the single source monitoring sub-model into the second feature space remapping unit of the single source monitoring sub-model for feature processing to obtain single source degradation features;

[0091] The target mulch film degradation characteristics are obtained by performing feature space remapping on the single-source degradation characteristics.

[0092] In an embodiment of the present invention, for example, when monitoring a certain experimental field full of biological mulch, if the multi-source monitoring data only includes images of the mulch appearance collected by a high-definition camera in the field (in the image, the mulch can be seen to be partially white and with slightly damaged textures on the edges), the server loads the mulch appearance image into the first feature space remapping unit of the single-source monitoring sub-model of the target mulch degradation monitoring model. This unit consists of a 3×3 convolutional layer, a maximum pooling layer, and a ReLU activation function. The convolutional layer captures spatial features such as the "color gradient of the whitened area" and "texture direction of the damaged edge" of the mulch. The pooling layer compresses the feature dimension and retains key structural information, outputting an initial feature with a dimension of 256×64×64 (this feature preliminarily extracts visual cues related to degradation in the mulch appearance). Next, the server loads the initial features output by the first feature space remapping unit into the second feature space remapping unit of the single-source monitoring sub-model. This unit uses a residual block structure (consisting of 1×1 convolution, 3×3 convolution, and shortcut connections) to further refine features through multi-scale convolution: 1×1 convolution focuses on the global distribution density of white areas, 3×3 convolution analyzes the local connectivity of damaged textures, and shortcut connections preserve original feature information to prevent gradient vanishing. The final output is a single-source degradation feature with a dimension of 128×32×32. This feature deeply integrates degradation characteristics such as color change and texture damage in the mulch film appearance. Finally, the server performs feature space remapping on the single-source degradation feature. The mapping module consisting of a fully connected layer and a batch normalization layer is called. The fully connected layer flattens the 32×32 resolution features into vectors and adjusts the dimensions (such as mapping from 4096 dimensions to 2048 dimensions). The batch normalization layer stabilizes the feature numerical distribution and generates target mulch degradation features that meet the input requirements of subsequent models. Although this feature is only based on the mulch appearance image, it accurately captures the key clues of degradation in the appearance dimension through two-stage feature remapping and residual extraction, ensuring the effectiveness of monitoring and analysis in single-source data scenarios.

[0093] In an embodiment of the present invention, the basic feature extractor comprises multiple sequentially connected basic feature extraction components. The basic feature extractor, based on a pre-trained target mulch film degradation monitoring model, extracts features at different resolutions from the mulch film appearance image to obtain multi-resolution visual features. The target mulch film degradation features are then interleaved with the multi-resolution visual features to obtain cross-modal integrated monitoring features. This can be implemented as follows. In this embodiment, the basic feature extractor utilizes three layers of residual blocks connected in series, with convolution kernel sizes of 7×7 (stride 2, downsampling ×2), 3×3 (stride 2, downsampling ×2), and 3×3 (stride 2, downsampling ×2), respectively. Multi-scale convolution is used to cover degradation features across the full range of scales, from macroscopic contours (low resolution) to mesoscopic damage (medium resolution) to microscopic spots (high resolution). The number of cross-modal feature mapping networks is the same as for the basic components (three each), ensuring scale alignment between degradation features and visual features across the low-medium-high resolution hierarchy (e.g., low-resolution degradation features match macroscopic visual features, and high-resolution degradation features match microscopic visual features).

[0094] Loading the film appearance image into a first basic feature extraction component to perform feature extraction at a first resolution level to obtain film visual features at a first resolution level, and performing feature embedding processing on the film visual features at the first resolution level and the target film degradation features to obtain first cross-modal comprehensive monitoring features;

[0095] Loading the first cross-modal integrated monitoring feature into the next basic feature extraction component to perform feature extraction at a second resolution level to obtain a ground film visual feature at the second resolution level, and performing a feature embedding operation based on the ground film visual feature at the second resolution level and the target ground film degradation feature to obtain a second cross-modal integrated monitoring feature;

[0096] The second cross-modal comprehensive monitoring feature is loaded into the next basic feature extraction component, and the feature embedding operation is re-executed until the last basic feature extraction component outputs the target scale ground film visual feature, and the cross-modal comprehensive monitoring feature is obtained based on the target scale ground film visual feature and the target ground film degradation feature.

[0097] In an exemplary embodiment of the present invention, in a scenario involving monitoring the degradation of fully biological mulch film on a large-scale farmland, the server invokes the basic feature extractor of a pre-trained target mulch film degradation monitoring model (composed of three sequentially connected basic feature extraction components, each containing a residual convolution layer and a pooling layer). First, the server loads an image of the mulch film's appearance (512×512 resolution, showing the overall spread of the mulch film and the boundary between its edges and the soil) captured by a high-definition camera in the field into the first basic feature extraction component. This component extracts low-resolution features such as the "macroscopic outline of the mulch film" and "large-scale color gradients" using a 3×3 convolution kernel (with a stride of 2). After dimensionality reduction using maximum pooling, it outputs visual features of the mulch film at the first resolution level (256×64×64 dimensions, preserving global clues to the overall degradation trend of the mulch film, such as the color uniformity of the entire mulch film and the direction of large-scale edge damage). At the same time, the server calls upon the previously generated target mulch film degradation features (a high-dimensional vector that integrates multi-source sensor deformation information and appearance features) and feeds the first-resolution visual features and the target mulch film degradation features into a feature embedding module (based on a channel-attention mechanism). The attention module first calculates the correlation between the channels of the two feature types (for example, the cosine similarity between the "lightening of the mulch film RGB channel" channel and the "degradation deformation amplitude statistics" channel), generating a dynamic weight matrix. The visual features and degradation features are then weighted and fused channel by channel according to the weights. After refinement through a 1×1 convolution, the first cross-modal integrated monitoring feature (256×64×64 dimensions, integrating global visual cues of the mulch film with quantitative multi-source degradation information, such as the correlation between areas of abnormal color across the entire mulch film and areas of significant sensor deformation) is obtained. The server then loads the first cross-modal integrated monitoring feature into the next basic feature extraction component. This component focuses on medium-resolution features such as "texture density of local damage to the ground film" and "boundary gradient of white patches" through a combination of 1×1 and 3×3 convolutions (with a step size of 2). After average pooling, it outputs the visual features of the ground film at the second resolution level (dimension 128×32×32, depicting the mesoscopic degradation details of the ground film, such as the distribution of local damage groups and the aggregation morphology of small-scale white patches). The server again calls the target mulch degradation feature and feeds the second-resolution visual features and the target mulch degradation features into a similar feature embedding module. The module first analyzes the spatial correlation between the "damage texture density" channel and the "deformation vector gradient" channel (for example, areas with dense damage correspond to areas with sudden changes in the deformation vector, resulting in a higher correlation value). This generates a regional fusion coefficient. Based on the fusion coefficient, the two types of features are spatially weighted (increasing the contribution of visual texture and sensory deformation features in areas with dense damage). After fusion through a 3×3 convolution, the second cross-modal integrated monitoring feature (128×32×32 dimensions, enhancing the integration of mesoscopic degradation details with multi-source quantitative information) is obtained. For example, localized damage clusters correspond to concentrated areas of molecular chain breakage in sensor data. Finally, the server loads the second cross-modal integrated monitoring feature into the final basic feature extraction component.This component uses depthwise separable convolution (with a stride of 2) to capture high-resolution features such as molecular-level degradation spot edges and gradient changes at micro-damage edges. After adaptive pooling, it outputs target-scale visual features of the mulch film (with dimensions of 64×16×16, accurately depicting microscopic degradation traces, such as the outline of spots formed by individual molecular chain breaks and the edge texture of micron-level damage). The server inputs these target-scale visual features and target mulch film degradation features into the final feature embedding module. This module uses a spatial attention mechanism to focus on the spatial correspondence between molecular spot locations and deformation vector mutation points (e.g., the overlap between molecular spot coordinates and vector mutation coordinates in sensor data is assigned a high attention weight), generating pixel-level fusion weights. The microscopic visual features and multi-source degradation features are weighted pixel by pixel according to the weights. After global pooling and linear transformation, a cross-modal integrated monitoring feature (with dimensions of 256, integrating visual cues from macroscopic outlines to microscopic spots of the mulch film with quantitative degradation information from multi-source sensors, providing full-scale correlation features for subsequent degradation status analysis) is obtained. The entire process uses multi-component step-by-step feature extraction and cross-modal embedding to achieve progressive fusion of mulch degradation information from global trends to microscopic traces, ensuring that subsequent degradation status monitoring can accurately capture the correlation logic of degradation characteristics at various scales.

[0098] In an embodiment of the present invention, the target mulch film degradation monitoring model also includes a plurality of cross-modal feature mapping networks connected in sequence; the number of the cross-modal feature mapping networks is the same as the number of the basic feature extraction components; the feature embedding operation is performed based on the mulch film visual features of the second resolution level and the target mulch film degradation features to obtain the second cross-modal comprehensive monitoring features, which can be implemented through the following examples.

[0099] performing feature space remapping on the target mulch film degradation feature through a next cross-modal feature mapping network connected to the first cross-modal feature mapping network to obtain a deep mulch film degradation feature corresponding to the mulch film visual feature at the second resolution level; the target mulch film degradation feature is a feature output by the first cross-modal feature mapping network;

[0100] The second cross-modal comprehensive monitoring feature is obtained by superimposing the ground film visual feature of the second resolution level and the deep ground film degradation feature.

[0101] In an embodiment of the present invention, for example, in a scenario where the biofilm degradation of a 1,000-acre farmland is monitored, the target film degradation monitoring model is configured with three sequentially connected cross-modal feature mapping networks (the same number as the basic feature extraction components). When the server processes the film visual features at the second resolution level (this feature is output by the second basic feature extraction component, with a dimension of 128×32×32, and carries meso-level degradation details such as the "local distribution of damaged groups" and the "meso-level whitening patch aggregation morphology" of the film), it first calls the next cross-modal feature mapping network connected to the first cross-modal feature mapping network (i.e., the second cross-modal network) to perform feature space remapping on the target film degradation features. Here, the target film degradation feature is the output of the first cross-modal network (a high-dimensional feature with a dimension of 256 that integrates macroscopic visual cues of the film and multi-source sensory deformation). The first feature space remapping unit of the second cross-modal network (composed of a multi-layer perceptron and a batch normalization layer) first flattens the 256-dimensional feature and then uses a linear transformation to map it to a 128×32×32 dimension (matching the dimension of the second-resolution visual feature). The feature interaction operator (a spatial attention-based module) then calculates the spatial correlation between the "mesoscopic damage cluster location" and the "sensor deformation vector concentration area" (e.g., the area where the pixel coordinates of the damage cluster overlap with the coordinates of the deformation vector mutation, with a correlation value of 0.8 or above), generating a spatial weight matrix. The remapped features are then weighted to enhance the associated regions pixel by pixel, resulting in a deep film degradation feature (with a dimension of 128×32×32, deeply binding mesoscopic visual details with multi-source sensory quantitative information, such as the sensor feature of "high deformation vector amplitude gradient" corresponding to the area with "high damage cluster density"). The server then feeds the second-resolution visual features of the ground film (carrying mesoscopic visual cues such as the "edge gradient of the damage cluster" and the "mean color of the white patches") and the deep ground film degradation features into the feature superposition module. This module first aligns the two feature types (both at a resolution of 128 channels × 32 × 32). Then, through element-by-element weighted summation (for highly correlated pixel locations, the weights of both visual and degradation features are set to 0.5; for locations with low correlation, the weights are distributed 0.3:0.7 to balance information), fusing the "visual outline of the damage cluster" with the "quantified distribution of deformation." This ultimately generates the second cross-modal integrated monitoring feature (128 × 32 × 32 dimensions). This feature preserves the spatial morphology of the mesoscopic damage texture of the ground film while enhancing the quantitative degradation cues from multi-source sensing in the corresponding area, providing a mesoscopic cross-modal correlation foundation for the subsequent high-resolution feature extraction of basic components. The entire process ensures the precise binding of visual details and sensor quantification of mulch film degradation information at the mesoscale through step-by-step remapping and feature superposition of cross-modal networks, providing coherent hierarchical support for the fusion of full-scale degradation features and ensuring the continuity and integrity of feature associations in subsequent microscale monitoring.

[0102] In the embodiment of the present invention, the superposition of the ground film visual features at the second resolution level and the deep ground film degradation features to obtain the second cross-modal comprehensive monitoring features can be implemented through the following examples.

[0103] generating a first fusion coefficient of the ground film visual feature at the second resolution level according to the ground film visual feature at the second resolution level;

[0104] The second fusion coefficient corresponding to the deep film degradation feature is generated according to the first fusion coefficient; the fusion coefficient is generated based on the spatial attention mechanism, and the formula is: , where Visual feature pixel value, is the degradation feature pixel value, The attention weight matrix (the weight distribution of the key degradation areas is learned through pre-training) ensures that the bimodal feature weights of the key degradation areas such as broken edges and molecular spots are automatically strengthened

[0105] The second cross-modal comprehensive monitoring feature is obtained by linearly superimposing the ground film visual feature at the second resolution level and the deep ground film degradation feature according to the first fusion coefficient and the second fusion coefficient.

[0106] In an embodiment of the present invention, for example, in a scenario of monitoring the full biological mulch film degradation of a demonstration farmland, when the server processes the mulch film visual features of the second resolution level (output by the second basic feature extraction component, with a dimension of 128×32×32, carrying visual clues such as "mesoscopic damage group distribution density" and "white spot color mean gradient") and deep mulch film degradation features (after remapping by the second cross-modal feature mapping network, with a dimension of 128×32×32, containing sensor quantitative information such as "deformation vector amplitude gradient of the damage group corresponding area" and "molecular chain break density associated with white spot"), it first generates a first fusion coefficient based on the mulch film visual features of the second resolution level. The server invokes a spatial attention-based coefficient generation module to perform a saliency calculation on the "damage cluster edge pixels" (with large gradient changes and high texture complexity) and "white patch core pixels" (with RGB mean values below a threshold) in the visual features. Damage cluster edge pixels, due to their significant degradation signature, are assigned a first fusion coefficient of 0.8; soil background pixels, due to their weak association with degradation, are assigned a coefficient of 0.2. This ultimately generates a 32×32 first fusion coefficient matrix (each pixel corresponds to a fusion weight for the visual feature) that matches the resolution of the visual features. Next, the server generates a second fusion coefficient corresponding to the deep film degradation feature based on the first fusion coefficient. Using the feature association mapping module, the server analyzes the spatial overlap of "damage cluster edge" regions with high first fusion coefficients with "deformation vector abrupt changes" (where the L2 norm of the deformation vector exceeds the threshold) in the deep film degradation feature. A second fusion coefficient of 0.8 is assigned to these overlapping regions to ensure that the weights of the two features match in the key degradation zone. A coefficient of 0.2 is assigned to "soil background" regions with low first fusion coefficients, corresponding to "deformation stable regions" in the deep film degradation feature. This results in a second fusion coefficient matrix that is consistent with the spatial distribution of the first fusion coefficients. Finally, the server linearly superimposes the two features based on the first and second fusion coefficients. For each pixel, an element-by-element calculation is performed: "second-resolution visual features × first fusion coefficient + deep mulch degradation features × second fusion coefficient." For example, for pixels at the edge of a damaged cluster, the visual features (depicting the morphology of the damaged texture) and the degradation features (depicting the quantification of deformation) are each superimposed with a weight of 0.8 to strengthen the "damage texture-deformation degree" association. For soil background pixels, the visual features (soil texture) and the degradation features (deformation stability information) are each weighted 0.2 to mitigate interference. This ultimately generates a second cross-modal integrated monitoring feature with a dimension of 128 × 32 × 32. At the mesoscale, this feature preserves the visual morphological details of mulch damage clusters and white patches while deeply integrating quantitative degradation information from multi-source sensors in the corresponding area, providing a hierarchical cross-modal correlation foundation for subsequent high-resolution feature extraction.

[0107] In the embodiment of the present invention, the degradation status monitoring of the full biological mulch according to the cross-modal integrated monitoring feature to obtain the degradation status monitoring result of the full biological mulch can be performed through the following examples.

[0108] Loading the cross-modal integrated monitoring features into the degradation analysis network of the target mulch film degradation monitoring model to perform degradation status monitoring;

[0109] The degradation stages and degradation area heat maps of the full biological mulch film output by the degradation analysis network are obtained, and the degradation status monitoring results are obtained according to the degradation stages and the degradation area heat maps.

[0110] In an embodiment of the present invention, for example, in the scenario of monitoring the degradation of a fully biological mulch film in a certain experimental field, the server has generated a cross-modal comprehensive monitoring feature (with a dimension of 256, carrying full-scale degradation-related clues) that integrates the macroscopic contours, mesoscopic damage details, microscopic molecular-level spots, and multi-source sensor deformation information of the mulch film through steps such as multi-source data fusion and cross-modal feature embedding. The server loads this feature into the degradation analysis network of the target mulch film degradation monitoring model (the network consists of a classification branch and a regression branch in parallel: the classification branch contains a global average pooling layer and a softmax classifier for determining the degradation stage; the regression branch consists of a transposed convolution layer and an attention upsampling module for generating a degradation area heat map). In the classification branch, the global average pooling layer first compresses the 256-dimensional cross-modal comprehensive monitoring feature into a 1×1×256 global statistical vector, which aggregates the statistical laws of the degradation characteristics of the mulch film from macro to micro; then, the softmax classifier performs probability calculations on the vector based on the "undegraded / early / mid-stage / final stage" label mapping relationship learned in the pre-training stage. If the output probability of the "mid-term" category reaches 0.85, the fully biofilm mulch is considered to be in the mid-stage of degradation. In the regression branch, the transposed convolutional layer first upsamples the low-dimensional features to a resolution that matches the original mulch image (e.g., 512×512), restoring spatial details. The attention upsampling module then focuses on key degradation regions, such as "spot clusters formed by molecular chain breakage" and "edge gradients of mesoscopic damage clusters." By strengthening the feature weights of these regions, a heat map of degradation regions is generated. In the heat map, red pixels correspond to "areas of concentrated molecular degradation and core areas of sudden deformation vector changes" (degradation ≥ 70%), while blue pixels correspond to "soil background areas and stable deformation areas" (degradation ≤ 10%). The color gradient visually illustrates the spatial distribution of degradation. For example, dense red pixels at the edge of the mulch indicate significant degradation, while alternating blue and green pixels in the middle indicate relatively mild degradation. Finally, the server integrates the information of the degradation stage and the heat map of the degradation area to generate the degradation status monitoring results: "The full biological mulch in the experimental field is in the mid-term degradation stage, and the degradation degree of the edge and local damaged cluster areas is more than 70%. The degradation degree of the middle area is mostly maintained at 30%-50%. The area in direct contact with the soil is enriched with microorganisms, and the local degradation degree exceeds 60%." This result is pushed to the farmer's management terminal through the agricultural Internet of Things platform, providing farmers with accurate data support for judging the timing of replacing the mulch film and adjusting the field water and fertilizer management strategy. Throughout the process, the degradation analysis network, with its dual-branch structure of "qualitative classification + quantitative regression", not only realizes the macroscopic judgment of the degradation stage, but also completes the microscopic visualization of the degradation area, so that the monitoring results have both guidance for production decision-making and reference for field operations, which strongly supports the refined management of agricultural production.

[0111] In the embodiment of the present invention, the acquisition of multi-source monitoring data of the entire biological mulch film can be implemented through the following examples.

[0112] In multiple spectral bands, the surface interface area of the full biological mulch is imaged separately by multispectral sensing nodes to obtain multiple multispectral images, where each multispectral image corresponds to a certain spectral band; the four bands of red (650nm), green (550nm), blue (450nm), and near infrared (850nm) are selected because the reflectivity of full biological materials such as PBAT in the mulch decreases linearly with the breakage of the molecular chain in the near infrared band (fitting coefficient 0.95), the visible light band can intuitively present color changes and damaged textures, and the four-band combination takes into account both the quantitative degradation degree and visual morphological information;

[0113] Obtain the reflectivity data set corresponding to each multispectral image, and calculate the degradation deformation vector of each pixel based on the reflectivity data set corresponding to each multispectral image to generate a degradation deformation topology map corresponding to the surface interface area of the full biological mulch, and determine the degradation deformation topology map as the multi-source sensing data of the target mulch; the reflectivity is inverted based on the Lambert-Beer law, and the formula is: , combined with the calibration data of the spectroradiometer (such as ASD FieldSpec 4), correct errors such as atmospheric scattering and sensor noise; the degradation deformation vector includes the reflectivity change of the red / near infrared band ( ), gray-level co-occurrence matrix texture entropy (H), time series gradient ( ), the dimension is [ ,H, ,...],Enhancing degradation sensitivity through multi-feature fusion.

[0114] A ground film appearance image corresponding to the surface interface area is selected from a plurality of multispectral images, and multi-source monitoring data corresponding to the full biological ground film surface interface area is generated based on the ground film appearance image and the multi-source sensing data of the target ground film.

[0115] In an exemplary embodiment of the present invention, in a scenario where the degradation of fully biofilm mulch is monitored within a smart greenhouse, a server, acting as the data collection and processing hub, first dispatches multispectral sensing nodes deployed within the greenhouse (e.g., a tracked mobile sensing vehicle equipped with red, green, blue, and near-infrared cameras) to image the surface interface of the fully biofilm mulch in multiple spectral bands. Following a pre-set cruising path, the sensing vehicle sequentially switches between the red (650nm), green (550nm), blue (450nm), and near-infrared (850nm) bands, capturing the mulch area covering the crop rows. This generates four multispectral images corresponding to different spectral bands. In the red band image, the mulch exhibits a light red texture due to the molecular structure of organic matter, while the soil appears dark red. In the near-infrared band image, the reflectivity of undegraded mulch is high (bright white), while the reflectivity of initially degraded areas is reduced (light gray) due to molecular chain breakage, providing spectral dimension difference information for subsequent analysis. After receiving image data uploaded by the multispectral sensing nodes via the Internet of Things protocol, the server obtains a reflectivity dataset for each multispectral image. The sensor node has a built-in spectroradiometer. Based on the Lambert-Beer law and a database of ground feature spectra, it calculates the reflectance of each pixel in the red, green, blue, and near-infrared bands (for example, the reflectance of a pixel in the red band is 0.62, and the reflectance of a pixel in the near-infrared band is 0.85). The reflectance matrix is then packaged into a dataset and pushed to the server. The server invokes the degradation deformation vector calculation module to perform spatiotemporal comparison and feature interpretation on the reflectance datasets for each band. The server uses a difference algorithm to calculate the reflectance change (ΔR = R_t - R_{t-7}) for the same pixel over consecutive monitoring periods (for example, a 7-day interval). Texture features such as texture entropy and contrast are extracted using the gray-level co-occurrence matrix. The reflectance change and texture features are then integrated into a degradation deformation vector (for example, if ΔR_red = -0.15 and texture entropy = 1.8 for a pixel, the corresponding vector dimension is [-0.15, 1.8, ...]). The server arranges the degradation deformation vectors of all pixels according to spatial coordinates, generating a topological map of the degradation deformation of the mulch-surface interface (the color depth of each pixel maps the deformation vector amplitude, with red representing significant deformation and blue representing minimal deformation). This topological map serves as the multi-source sensing data for the target mulch. Finally, the server selects an image of the mulch's appearance from multiple multispectral images. Because true color images (RGB images), composed of red, green, and blue bands, can visually demonstrate mulch color variations (such as white patches), damaged contours (such as torn edges), and the soil boundary, the server selects this RGB image as the mulch's appearance image. The server then aligns the mulch's appearance image with the degradation deformation topological map by timestamp and spatial coordinates, encapsulating it into multi-source monitoring data containing both visual and quantitative sensing information. This provides a multi-dimensional input foundation for subsequent mulch degradation feature extraction and status monitoring.During the entire data acquisition process, the server realizes multi-dimensional collection of mulch film degradation information from spectral differences to spatial topology through band imaging, precise reflectivity interpretation and deformation vector space modeling of multi-spectral sensing nodes, laying a solid data foundation for intelligent monitoring of full-biological mulch film degradation.

[0116] In an embodiment of the present invention, the target mulch film degradation monitoring model is obtained by the following methods, including:

[0117] Acquire a multi-source monitoring data instance including a ground film appearance image instance and a ground film multi-source sensing data instance of the ground film monitoring sample, and full degradation cycle annotation data associated with the ground film monitoring sample;

[0118] Loading the multi-source monitoring data instance into the multi-source monitoring sub-model of the initial model, and loading the ground film appearance image instance into the single-source monitoring sub-model of the initial model, to obtain the corresponding ground film degradation feature instance;

[0119] Performing feature processing of different resolutions on the ground film appearance image instance by the basic feature extractor of the initial model to obtain ground film visual feature instances at multiple resolution levels, and performing feature embedding operation on the ground film degradation feature instance and the ground film visual feature instances at multiple resolution levels to obtain a cross-modal comprehensive monitoring feature instance;

[0120] Performing degradation state monitoring on the cross-modal integrated monitoring feature instance through the degradation analysis network of the initial model to obtain a degradation state monitoring result instance;

[0121] Error parameters are generated according to the degradation status monitoring result instance and the error of the full degradation cycle annotation data, and the network parameters of the initial model are optimized according to the error parameters to obtain the target mulch film degradation monitoring model.

[0122] In an embodiment of the present invention, exemplarily, in a model training scenario of an agricultural experimental field, the server acts as a training hub and first obtains training data: from a database of 20 1m×1m ground film monitoring plots arranged in the experimental field, extracts multi-source monitoring data instances of each plot (including ground film appearance image instances taken by an RGB camera, such as the image of plot #5 showing that the edge of the ground film has become lighter in color and has a locally damaged texture; and ground film multi-source sensing data instances generated by a multispectral sensing node, i.e., a degradation deformation topology map, which depicts the spatial distribution of deformation caused by molecular chain breakage in the form of a map after calculating the deformation vector of each pixel through multispectral reflectance); and at the same time retrieves the full degradation cycle annotation data of manual weekly sampling and testing (such as plot #5, which is labeled "mid-term degradation, molecular weight reduction of 60% in the edge area and breakage rate of 35%" after laboratory testing, corresponding to the "mid-term" stage label and the regional thermal label of "red edge, light in the middle"). Next, the server inputs the sub-model to extract features: The multi-source monitoring data instance is loaded into the multi-source monitoring sub-model (atlas modality processing sub-model) of the initial model. Using the first feature space remapping unit (composed of convolution and pooling layers), it extracts mulch film appearance atlas features (such as the texture gradient and color mean at the damaged edge of sample #5) and sensor vector atlas features (the amplitude and directional correlation of the corresponding regional deformation vector). Simultaneously, the mulch film appearance image instance is loaded into the single-source monitoring sub-model. After a two-stage feature remapping process (the first remapping unit extracts basic visual features, and the second remapping unit refines degradation-related clues), the output is a single-source degradation feature (such as the color distribution entropy of the damaged cluster in sample #5). The output features of the two sub-models are processed through spatial remapping and interaction operators in the cross-modal feature mapping network to generate a mulch film degradation feature instance (the features of sample #5 are deeply bound to strong correlations such as "edge damage texture" and "deformation vector mutation"). Subsequently, the server extracts and embeds basic features: the basic feature extractor of the initial model (composed of three residual components connected in series) is called to process the ground film appearance image instance to generate multi-resolution ground film visual feature instances. The first residual component outputs low-resolution features (capturing the overall color distribution trend and macro-damage outline of sample plot #5) through large convolution kernels and pooling; the second residual component outputs medium-resolution features (focusing on the density of local damage groups and the aggregated morphology of white patches) through multi-scale convolution; the last residual component outputs high-resolution features (analyzing the edge gradient of molecular-level degradation spots and the texture details of micro-damage) through depthwise separable convolution. The server inputs the mulch degradation feature instances and the visual feature instances of each resolution into the feature embedding module, calculates the modal correlation through the attention mechanism (such as low-resolution features are associated with global deformation trends, and high-resolution features are associated with microscopic spots and sensor mutation points), performs feature coupling after dynamic weighting, and generates cross-modal comprehensive monitoring feature instances (integrating the degradation correlation clues of sample plot #5 from macro to micro, such as "overall color becomes lighter-global deformation amplitude increases" and "molecular spots-local deformation vector mutation").Next, the server performs degradation analysis and error calculation: The cross-modal integrated monitoring feature instances are fed into the degradation analysis network of the initial model. The classification branch uses global pooling and a softmax function to output a degradation stage prediction for quadrat #5 (e.g., "mid-term"). The regression branch uses transposed convolution and attention upsampling to generate a degradation region heatmap (with 30% red pixels along the edges). The server compares the labeled data for the entire degradation cycle (quadrat #5 is actually "mid-term" with 35% red pixels along the edges), calculates the classification cross-entropy loss (error value 0.12) and the mean squared error of the heatmap (error value 0.05), and integrates these to generate an error parameter. Finally, the server optimizes the model: Backpropagation is performed based on the error parameter to adjust network parameters at each layer of the initial model. For example, the weights of the cross-convolution correlation component in the multi-source monitoring submodel are used to strengthen the modal correlation between "damage texture" and "deformation vector." The residual block convolution kernel parameters of the basic feature extractor are optimized to improve the accuracy of micro-degradation feature extraction. The bias of the fully connected layer of the degradation analysis network is adjusted to reduce the error between the stage prediction and heatmap generation. After 50 rounds of iterative training, the model loss converged (classification loss < 0.05, heat map loss < 0.02), and finally a pre-trained target mulch film degradation monitoring model was obtained, ensuring the precise fusion of multimodal features and accurate determination of degradation status during actual monitoring.

[0123] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent monitoring method for biofilm degradation based on multi-source data collection. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0124] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. A method for intelligent monitoring of biofilm degradation based on multi-source data acquisition, characterized in that: include: Obtain multi-source monitoring data for all biofilms; If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensing data, a feature extraction operation is performed on the ground film appearance image and the target ground film multi-source sensing data to obtain a target ground film degradation feature, where the target ground film degradation feature is a fusion of features of the ground film appearance image and features of the target ground film multi-source sensing data; The basic feature extractor of the pre-trained target film degradation monitoring model is used to perform feature extraction of different resolutions on the film appearance image to obtain film visual features at multiple resolution levels, and the target film degradation features and the film visual features at multiple resolution levels are subjected to feature embedding operations to obtain cross-modal comprehensive monitoring features, including: loading the first cross-modal comprehensive monitoring features into the next basic feature extraction component to perform feature extraction at the second resolution level to obtain film visual features at the second resolution level; performing feature space remapping on the target film degradation features through the next cross-modal feature mapping network connected to the first cross-modal feature mapping network to obtain deep film degradation features corresponding to the film visual features at the second resolution level; the target film degradation features are features output by the first cross-modal feature mapping network; generating a first fusion coefficient of the film visual features at the second resolution level according to the film visual features at the second resolution level; generating a second fusion coefficient corresponding to the deep film degradation features according to the first fusion coefficient; and and the second fusion coefficient to linearly superimpose the second resolution level ground film visual features and the deep ground film degradation features to obtain the second cross-modal comprehensive monitoring features; the second cross-modal comprehensive monitoring features are loaded into the next basic feature extraction component, and the feature interlocking operation is re-executed until the last basic feature extraction component outputs the ground film visual features of the target scale, and the cross-modal comprehensive monitoring features are obtained according to the ground film visual features of the target scale and the target ground film degradation features; the basic feature extractor includes a plurality of basic feature extraction components connected in sequence, and the target ground film degradation monitoring model also includes a plurality of cross-modal feature mapping networks connected in sequence; the number of the cross-modal feature mapping networks is the same as the number of the basic feature extraction components; the ground film appearance image is loaded into the first basic feature extraction component to perform feature extraction of the first resolution level to obtain the ground film visual features of the first resolution level, and the ground film visual features of the first resolution level and the target ground film degradation features are subjected to feature interlocking operation to obtain the first cross-modal comprehensive monitoring features; The degradation status of the full biological mulch is monitored according to the cross-modal comprehensive monitoring characteristics to obtain a degradation status monitoring result of the full biological mulch.

2. The method according to claim 1, characterized in that If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensing data, performing a feature extraction operation on the ground film appearance image and the target ground film multi-source sensing data to obtain target ground film degradation features, including: If the multi-source monitoring data includes a ground film appearance image and target ground film multi-source sensor data, and the sensor signal carrier type of the target ground film multi-source sensor data includes a sensor vector atlas type, then selecting the atlas modal processing sub-model of the target ground film degradation monitoring model to perform feature extraction operations on the ground film appearance image and the target ground film multi-source sensor data, respectively, to obtain ground film appearance atlas features and sensor vector atlas features, respectively; Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature; The target mulch film degradation feature is obtained by performing feature space remapping on the cross-atlas fusion intermediate feature.

3. The method according to claim 2, characterized in that The target mulch film degradation monitoring model further includes at least one cross-modal feature mapping network; the step of performing feature space remapping on the cross-atlas fusion intermediate features to obtain the target mulch film degradation features includes: Loading the cross-atlas fusion intermediate features into the first feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain initial remapping features; Performing feature interaction mapping processing on the initial remapped features according to a feature interaction operator to obtain cross-modal intermediate features; The cross-modal intermediate features are loaded into the second feature space remapping unit of the cross-modal feature mapping network to perform feature space remapping to obtain the target mulch film degradation features.

4. The method according to claim 2, characterized in that The atlas modal processing sub-model includes a first feature space remapping unit and a second feature space remapping unit, and the atlas modal processing sub-model also includes a cross-convolution correlation component; the atlas modal processing sub-model of the target mulch film degradation monitoring model performs feature extraction operations on the mulch film appearance image and the target mulch film multi-source sensor data, respectively, to obtain mulch film appearance atlas features and sensor vector atlas features, respectively, including: Loading the ground film appearance image and the target ground film multi-source sensor data into the first feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain ground film appearance atlas features and sensor vector atlas features; Performing a feature embedding operation on the ground film appearance atlas feature and the sensor vector atlas feature to obtain a cross-atlas fusion intermediate feature, including: Loading the ground film appearance map features and the sensor vector map features into the cross convolution correlation component to calculate the modal correlation between the ground film appearance map features and the sensor vector map features, and generating a dynamic fusion coefficient according to the modal correlation; Performing spatial adaptive weighting on the ground film appearance map features and the sensor vector map features according to the dynamic fusion coefficient; The coupled features are obtained by performing a feature coupling operation on the surface map features of the ground film and the sensor vector map features after spatial adaptive weighting; The coupling features are loaded into the second feature space remapping unit of the atlas modal processing sub-model to perform feature extraction operations to obtain the cross-atlas fusion intermediate features.

5. The method according to claim 1, wherein The method further comprises: If the multi-source monitoring data includes a ground film appearance image, the ground film appearance image is loaded into the first feature space remapping unit of the single-source monitoring sub-model of the target ground film degradation monitoring model for feature processing; Loading the features output by the first feature space remapping unit of the single source monitoring sub-model into the second feature space remapping unit of the single source monitoring sub-model for feature processing to obtain single source degradation features; The target mulch film degradation characteristics are obtained by performing feature space remapping on the single-source degradation characteristics.

6. The method according to claim 1, characterized in that The step of monitoring the degradation status of the full biological mulch according to the cross-modal comprehensive monitoring characteristics to obtain the degradation status monitoring result of the full biological mulch includes: Loading the cross-modal integrated monitoring features into the degradation analysis network of the target mulch film degradation monitoring model to perform degradation status monitoring; The degradation stages and degradation area heat maps of the full biological mulch film output by the degradation analysis network are obtained, and the degradation status monitoring results are obtained according to the degradation stages and the degradation area heat maps.

7. The method according to claim 1, characterized in that The multi-source monitoring data of the full biological mulch film is obtained, including: In multiple spectral bands, the surface interface area of the entire biofilm is imaged separately by multispectral sensing nodes to obtain multiple multispectral images, where each multispectral image corresponds to a spectral band; Obtaining a reflectance dataset corresponding to each multispectral image, and calculating a degradation deformation vector of each pixel based on the reflectance dataset corresponding to each multispectral image to generate a degradation deformation topology map corresponding to the surface interface area of the full biological mulch film, and determining the degradation deformation topology map as the multi-source sensing data of the target mulch film; A ground film appearance image corresponding to the surface interface area is selected from a plurality of multispectral images, and multi-source monitoring data corresponding to the full biological ground film surface interface area is generated based on the ground film appearance image and the multi-source sensing data of the target ground film.

8. The method according to claim 1, characterized in that The target mulch film degradation monitoring model is obtained by the following methods, including: Acquire a multi-source monitoring data instance including a ground film appearance image instance and a ground film multi-source sensing data instance of the ground film monitoring sample, and full degradation cycle annotation data associated with the ground film monitoring sample; Loading the multi-source monitoring data instance into the multi-source monitoring sub-model of the initial model, and loading the ground film appearance image instance into the single-source monitoring sub-model of the initial model, to obtain the corresponding ground film degradation feature instance; Performing feature processing of different resolutions on the ground film appearance image instance by the basic feature extractor of the initial model to obtain ground film visual feature instances at multiple resolution levels, and performing feature embedding operation on the ground film degradation feature instance and the ground film visual feature instances at multiple resolution levels to obtain a cross-modal comprehensive monitoring feature instance; Performing degradation state monitoring on the cross-modal integrated monitoring feature instance through the degradation analysis network of the initial model to obtain a degradation state monitoring result instance; Error parameters are generated according to the degradation status monitoring result instance and the error of the full degradation cycle annotation data, and the network parameters of the initial model are optimized according to the error parameters to obtain the target mulch film degradation monitoring model.

9. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 8.

Images