Method and system for monitoring the cultivation process of banana seedlings using disease-resistant nutrient solution

Abstract

The present invention provides a method and system for monitoring the banana seedling cultivation process using disease-resistant nutrient solution. Based on soil and environmental data from subregions, node weights are calculated to construct a dynamic spatiotemporal graph. When plant physiological data changes exceed a threshold, the spatiotemporal graph is updated. For each subregion, local plant physiological, soil, and environmental data are integrated to form an instance package, and node features are derived through instance-attention weighted aggregation. Based on the node features, the system selectively performs graph convolution on high-risk or uncertain nodes and extracts node temporal features based on historical spatiotemporal graph information to predict the growth anomaly score and specific disease probability of the subregion. If the growth anomaly score of a subregion exceeds a dynamic threshold, multi-instance learning is performed on its instance package to identify key instance patterns indicating early stress. Finally, based on the predicted disease probability and the identified key instance patterns, combined with a preset stress regulation model, adjustment parameters for the disease-resistant nutrient solution are calculated and implemented, achieving precise control.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and more particularly to a method and system for monitoring the cultivation process of banana seedlings using disease-resistant nutrient solution. Background Art

[0002] Banana is an important economic crop, and seedling cultivation plays a crucial role in subsequent growth and yield. However, traditional banana seedling cultivation methods still generally rely on the grower's experience and periodic agricultural operations. This model is particularly inadequate when faced with the complex and changing growing environment and the sophisticated needs of banana seedlings at different growth stages. In particular, the application of disease-resistant nutrient solutions to enhance seedling stress resistance is hindered by the lack of real-time, continuous monitoring of soil conditions, climatic conditions, and plant physiological status, making it difficult to accurately determine the nutrient solution ratio and application timing. Growers often manage banana seedlings based on established procedures or macroscopic observations, failing to dynamically adjust to the actual needs of individual or small-scale banana seedlings. This extensive management approach not only wastes water and fertilizer resources but, more importantly, hinders the full effectiveness of the disease-resistant nutrient solution, resulting in uneven growth uniformity, health, and ultimately disease resistance among banana seedlings. This impacts the efficiency of cultivating high-quality, strong seedlings and the survival rate and yield potential of subsequent field planting.

[0003] Traditional manual inspection and record-keeping methods are not only labor-intensive and inefficient, but also difficult to avoid data bias and lags caused by subjective judgment. For example, when banana seedlings show subtle signs of early stress or disease infestation, manual observation often fails to accurately identify and quantify them immediately, missing the optimal opportunity for intervention. Furthermore, manual monitoring of dynamic changes in key parameters within the cultivation environment, such as light, temperature and humidity, soil pH, and nutrient content, makes it difficult to achieve high-frequency, multi-dimensional data collection and comprehensive analysis. This limited information access means that cultivation decisions lack sufficient data support, making it difficult to respond promptly and effectively to sudden environmental stresses or potential disease risks, let alone optimize nutritional strategies and environmental control measures based on immediate plant feedback. The banana seedling cultivation field urgently needs a systematic solution that integrates sensing technology, data analysis, and intelligent control to overcome the inherent shortcomings of traditional methods, enabling refined and intelligent management of the cultivation process. This will steadily improve the overall quality and disease resistance of banana seedlings and ensure the sustainable development of the banana industry. Summary of the Invention

[0004] To address the technical problem that the banana seedling cultivation process cannot be fully monitored and accurately regulated, the present application proposes a method for monitoring the banana seedling cultivation process using a disease-resistant nutrient solution, which is characterized by comprising:

[0005] Soil data, environmental data, and plant physiological data from a banana seedling cultivation area are collected, the banana seedling cultivation area is divided into multiple sub-areas, edge weights are calculated based on the soil and environmental data, and a spatiotemporal graph is constructed based on the weights and using the sub-areas as nodes. A new spatiotemporal graph is generated when changes in the plant physiological data exceed a preset fluctuation threshold.

[0006] For each sub-region node, an instance package is obtained based on the local plant physiological data and the corresponding soil data and environmental data. The instance package of the sub-region node is used to obtain the characteristics of the weighted aggregation of instance attention within each node. The nodes for graph convolution in the spatiotemporal graph are determined based on the characteristics of the sub-region nodes. The graph convolution is performed on the nodes that need to be convolved. The time series characteristics of the nodes are obtained based on the characteristics of the same sub-region nodes in multiple spatiotemporal graphs. The time series characteristics are used to predict the future growth abnormality score and specific disease probability of each sub-region.

[0007] When the growth anomaly score of a sub-region is greater than an anomaly threshold dynamically adjusted based on the global disease probability, multi-instance learning is performed on the instance package of the sub-region to identify key instance patterns of early stress; based on the disease probability and key instance patterns, combined with a preset regulation model of the stress type, adjustment parameters of the disease-resistant nutrient solution are calculated, and the disease-resistant nutrient solution is adjusted using the adjustment parameters.

[0008] Optionally, calculating edge weights based on soil data and environmental data includes:

[0009] Obtain soil and environmental data for each sub-region;

[0010] The weights of the edges between nodes are calculated based on the similarity of the soil data and environmental data between sub-regions.

[0011] Optionally, determining nodes for performing graph convolution in the spatiotemporal graph according to the characteristics of the sub-region nodes, and performing graph convolution on the nodes requiring graph convolution, includes:

[0012] For each sub-region node, based on its current and historical plant physiological data, soil data, and environmental data, the risk value and confidence level of abnormal growth in the next period are predicted;

[0013] When the growth anomaly risk value of a sub-region node is higher than the preset risk threshold and its corresponding confidence level is lower than the preset confidence level, the sub-region node is marked as a risk node;

[0014] Obtain the previous and next spatiotemporal graphs of the spatiotemporal graph where the risk node is located, take the nodes in the previous and next spatiotemporal graphs that are identical to the risk node as adjacent points of the risk node, and obtain the adjacent points in the spatiotemporal graph where the risk node is located;

[0015] Graph convolution is performed on the risk node using the risk node and its adjacent nodes.

[0016] Optionally, performing multi-instance learning on the instance bag of the sub-region to identify key instance patterns of early stress includes:

[0017] For the instance package of the sub-region where the growth anomaly score exceeds the threshold, each instance feature in the instance package is input into the fully connected layer and then subjected to a nonlinear activation function to obtain the stress contribution score;

[0018] Selecting at least one instance whose coercion contribution score is higher than a preset contribution threshold as a key instance;

[0019] A key instance pattern of early coercion is obtained according to the characteristics of the key instance.

[0020] Optionally, the adjustment parameters of the disease-resistant nutrient solution are calculated based on the disease probability and key instance pattern in combination with a preset stress type regulation model, including:

[0021] Input the predicted specific disease probability and key instance pattern into the preset adjustment model to obtain the adjustment amplitude coefficient;

[0022] The reference concentration is adjusted based on the adjustment amplitude coefficient to obtain an adjusted nutrient solution parameter.

[0023] In addition, the present application proposes a banana seedling cultivation process monitoring system using disease-resistant nutrient solution, comprising:

[0024] a graph construction unit for collecting soil data, environmental data, and plant physiological data from a banana seedling cultivation area, dividing the banana seedling cultivation area into a plurality of sub-areas, calculating edge weights based on the soil data and the environmental data, constructing a spatiotemporal graph based on the weights and using the sub-areas as nodes, and generating a new spatiotemporal graph when a change in the plant physiological data exceeds a preset fluctuation threshold;

[0025] A prediction unit is configured to obtain, for each sub-region node, an instance package based on local plant physiological data and corresponding soil and environmental data; use the instance package of the sub-region node to obtain features of weighted aggregation of instance attention within each node; determine nodes for graph convolution in the spatiotemporal graph based on the features of the sub-region nodes; perform graph convolution on the nodes that require graph convolution; obtain temporal features of the nodes based on features of multiple nodes in the same sub-region in the spatiotemporal graph; and use the temporal features to predict future growth anomaly scores and specific disease probabilities for each sub-region;

[0026] An adjustment unit is configured to perform multi-instance learning on the instance package of the sub-region to identify key instance patterns of early stress when the growth abnormality score of the sub-region is greater than an abnormality threshold dynamically adjusted based on the global disease probability; calculate adjustment parameters of the disease-resistant nutrient solution based on the disease probability and key instance patterns, combined with a preset adjustment model of the stress type, and use the adjustment parameters to adjust the disease-resistant nutrient solution.

[0027] Optionally, calculating edge weights based on soil data and environmental data includes:

[0028] Obtain soil and environmental data for each sub-region;

[0029] The weights of the edges between nodes are calculated based on the similarity of the soil data and environmental data between sub-regions.

[0030] Optionally, determining nodes for performing graph convolution in the spatiotemporal graph according to the characteristics of the sub-region nodes, and performing graph convolution on the nodes requiring graph convolution, includes:

[0031] For each sub-region node, based on its current and historical plant physiological data, soil data, and environmental data, the risk value and confidence level of abnormal growth in the next period are predicted;

[0032] When the growth anomaly risk value of a sub-region node is higher than the preset risk threshold and its corresponding confidence level is lower than the preset confidence level, the sub-region node is marked as a risk node;

[0033] Obtain the previous and next spatiotemporal graphs of the spatiotemporal graph where the risk node is located, take the nodes in the previous and next spatiotemporal graphs that are identical to the risk node as adjacent points of the risk node, and obtain the adjacent points in the spatiotemporal graph where the risk node is located;

[0034] Graph convolution is performed on the risk node using the risk node and its adjacent nodes.

[0035] Optionally, performing multi-instance learning on the instance bag of the sub-region to identify key instance patterns of early stress includes:

[0036] For the instance package of the sub-region where the growth anomaly score exceeds the threshold, each instance feature in the instance package is input into the fully connected layer and then subjected to a nonlinear activation function to obtain the stress contribution score;

[0037] Selecting at least one instance whose coercion contribution score is higher than a preset contribution threshold as a key instance;

[0038] A key instance pattern of early coercion is obtained according to the characteristics of the key instance.

[0039] Optionally, the adjustment parameters of the disease-resistant nutrient solution are calculated based on the disease probability and key instance pattern in combination with a preset stress type regulation model, including:

[0040] Input the predicted specific disease probability and key instance pattern into the preset adjustment model to obtain the adjustment amplitude coefficient;

[0041] The reference concentration is adjusted based on the adjustment amplitude coefficient to obtain an adjusted nutrient solution parameter.

[0042] This application calculates edge weights based on real-time changing soil and environmental data, and generates a new spatiotemporal graph based on significant changes in plant physiological data, so that the model can dynamically capture the mutual influence between sub-regions in the cultivation area and the transformation of key growth nodes, overcoming the defect that traditional methods are difficult to adapt to the spatiotemporal heterogeneity of the environment, and improving the ability to characterize complex growth environments. By aggregating sub-region features through instance attention weighting and selectively performing graph convolution based on node features, the model can focus more on key information, improving the effectiveness and pertinence of feature extraction; at the same time, combining the node information of multiple spatiotemporal graphs to form time series features for prediction, it enhances the early warning capability of growth trends and disease occurrence. In addition, when growth anomalies are monitored, the present invention triggers the execution of multi-instance learning on the instance package of high-risk areas through thresholds, which can effectively identify the key instance patterns that lead to early stress from mixed data, providing a basis for subsequent accurate diagnosis and intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of Example 1;

[0044] Figure 2 Schematic diagram of the sub-region division;

[0045] Figure 3 It is a schematic diagram of the space-time diagram;

[0046] Figure 4 Schematic diagram of risk nodes and their adjacent points. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0048] The terms "first", "second" and corresponding terminology numbers in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances. This is merely a way of distinguishing when describing objects with the same properties in the embodiments of the present application. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, so that a process, method, system, product or apparatus that includes a series of units is not necessarily limited to those units, but may include other units that are not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0049] In addition, in the description of this application, unless otherwise specified, "plurality" means two or more. The term "and / or" or the character " / " in this application is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B, or A / B, can mean: A exists alone, A and B exist at the same time, or B exists alone.

[0050] Specific embodiment, a banana seedling cultivation process monitoring method using disease-resistant nutrient solution, such as Figure 1 Shown, including:

[0051] Step 1: collecting soil data, environmental data, and plant physiological data from a banana seedling cultivation area, dividing the banana seedling cultivation area into multiple sub-areas, calculating edge weights based on the soil data and the environmental data, constructing a spatiotemporal graph based on the weights and using the sub-areas as nodes, and generating a new spatiotemporal graph when a change in the plant physiological data exceeds a preset fluctuation threshold;

[0052] Soil data, environmental data, and plant physiological data from the banana seedling cultivation area are collected. The soil data includes, but is not limited to, soil temperature and humidity, pH value, electrical conductivity, and key nutrient content such as nitrogen, phosphorus, and potassium. The environmental data includes, but is not limited to, air temperature and humidity, light intensity, and carbon dioxide concentration. The plant physiological data includes, but is not limited to, leaf temperature, chlorophyll fluorescence parameters, stem micro-variation, and hyperspectral imaging data.

[0053] Large-scale cultivation areas often have gradient distribution and unevenness of environmental parameters. Simple overall regulation is difficult to meet the differentiated needs of banana seedlings in different locations. By dividing the area into sub-areas, the big problem can be broken down into multiple small problems, making it easier for subsequent models to analyze and predict the status of each sub-area in a targeted manner. Figure 2 A schematic diagram of dividing an area into sub-areas is shown. The sub-area division method includes but is not limited to the spatial clustering results based on geographic grids, sensor layouts, or historical data.

[0054] The nodes of the spatiotemporal graph represent the sub-regions, and the edges represent the mutual influence or correlation strength between the sub-regions. The edge weights are dynamically calculated based on the actual soil data and environmental data collected. For example, if the soil moisture and temperature of two adjacent sub-regions are very similar, the edge weight between them may be high, indicating that their environmental states have strong consistency or mutual influence. Figure 3 A schematic diagram of a spatiotemporal graph is shown. A plant's physiological state is a direct reflection of its response to the environment. When the physiological data of banana seedlings in one or more subregions undergoes significant changes, such as a sudden change in growth rate or abnormal leaf spectra, exceeding the normal physiological fluctuation range, i.e., the preset fluctuation threshold, a new state may have been entered, and the original inter-subregion influence relationship may no longer apply. At this point, a new spatiotemporal graph is generated, and the edge weights of this spatiotemporal graph are recalculated based on the latest soil and environmental data. In another embodiment, the graph's topology is also adjusted based on the changes in correlation indicated by the physiological data. In one embodiment, adjacent subregions are connected by edges. In an alternative embodiment, a spatiotemporal graph is generated at each preset time.

[0055] Step 2: For each sub-region node, an instance package is obtained based on the local plant physiological data and the corresponding soil data and environmental data; the instance package of the sub-region node is used to obtain the characteristics of the weighted aggregation of instance attention in each node, and the nodes for graph convolution in the spatiotemporal graph are determined based on the characteristics of the sub-region nodes. The graph convolution is performed on the nodes that need to be convolved, and the time series characteristics of the nodes are obtained based on the characteristics of the same sub-region nodes in multiple spatiotemporal graphs. The time series characteristics are used to predict the future growth abnormality score and specific disease probability of each sub-region;

[0056] A sub-region may contain multiple banana seedlings. The local plant physiological data and the soil and environmental data closely related to their occurrence are taken as an example. For example, the combination of the spectral data of a leaf and the air temperature, humidity, and light intensity data around it at the same time point is taken as an example. A sub-region node corresponds to a package containing multiple instances. Because the contribution of each instance in the instance package to the overall state of the sub-region may be different, the present invention uses an attention mechanism for weighted aggregation. The attention mechanism can learn which types of instances are more important for judging the health status or potential risks of the current sub-region and assign higher weights to these key instances, thereby obtaining a more representative sub-region node aggregation feature.

[0057] Graph convolutional networks (GCNs) are a method for feature extraction on graph-structured data. They aggregate node features and those of its neighbors. Different nodes have varying importance, and performing graph convolution on all points in the graph may introduce noise from other nodes while leaving reliable nodes unaffected. Key nodes for graph convolution are identified based on the characteristics of sub-region nodes, such as those with a preliminary assessment of anomaly risk, information entropy, or the severity of feature changes. Because spatiotemporal graphs are generated dynamically, they vary at different points in time. For a sub-region at the same geographic location, the features extracted from these continuous or discrete spatiotemporal graph snapshots form a time series. From these time series features, temporal features are extracted that reflect the evolving growth trends of the sub-region. Based on these extracted temporal features, which capture the growth dynamics and spatial correlations of each sub-region, a predictive model is used to output a score for the degree of growth anomaly and the probability of specific diseases occurring in the sub-region over a period of time.

[0058] Step 3: When the growth anomaly score of a sub-region is greater than an anomaly threshold dynamically adjusted based on the global disease probability, multi-instance learning is performed on the instance package of the sub-region to identify key instance patterns of early stress; based on the disease probability and key instance patterns, combined with a preset regulation model of the stress type, the adjustment parameters of the disease-resistant nutrient solution are calculated, and the disease-resistant nutrient solution is adjusted using the adjustment parameters.

[0059] A preliminary growth anomaly score is obtained through spatiotemporal graph and time series analysis. If the anomaly score of a subregion exceeds a dynamically set threshold, it is considered to be at high risk and requires more in-depth diagnosis. The original instance package of the high-risk subregion is further retrieved and the multi-instance learning (MIL) algorithm is executed on it to identify the key instance combination that leads to the early stress state of the subregion.

[0060] After determining the specific disease types that may occur in the sub-area and the key causes of early stress, this information is input into the regulation model of the stress type. In one embodiment, the regulation model is based on expert knowledge or learned through historical data. For example, in the early stages of a certain fungal disease, it may be necessary to increase the supply of potassium and silicon to improve plant resistance. The model will output specific disease-resistant nutrient solution adjustment suggestions based on the input disease probability and stress pattern. The adjustment suggestions include but are not limited to the nutrients or drugs that need to be adjusted, as well as the adjustment direction and amplitude. Based on the calculated adjustment parameters, the formula and application amount of the disease-resistant nutrient solution are actually adjusted by the automatic control unit to accurately fertilize and apply pesticides to the target sub-area.

[0061] The environments of banana seedlings in different locations are not completely independent. For example, if the soil in area A is particularly wet, it may gradually affect the adjacent area B. If area A is well ventilated and the temperature is low, it may also cause the temperature of the adjacent area B to drop slightly. Moreover, the sunlight, humidity, and soil conditions in different locations may not be exactly the same. In an optional embodiment, the edge weights calculated based on soil data and environmental data include:

[0062] Obtain soil and environmental data for each sub-region;

[0063] The weights of the edges between nodes are calculated based on the similarity of the soil data and environmental data between sub-regions.

[0064] For example, in Area A, the soil moisture is 60% and the air temperature is 25°C; in Area B, the soil moisture is 58% and the air temperature is 25.5°C. If the soil in Areas A and B are similarly moist and have similar temperatures, then their soil conditions are very similar. Similarly, if the air temperature, humidity, and sunlight in Areas A and B are similar, then their environmental conditions are very similar. The weights of the edges between nodes are calculated based on the similarity of the soil and environmental data between the sub-areas, for example, through weighted methods.

[0065] Not all sub-regions have the same importance or the same uncertainty at any moment. Some sub-regions may be growing stably, in good health, and have a high confidence in their state predictions; while other sub-regions may show potential signs of abnormal growth, or the model's initial judgment of their state may have greater uncertainty. For those reliable nodes that have been accurately predicted as healthy and with high confidence by the preliminary model, if graph convolution is still performed on them, unnecessary noise may be introduced from their neighboring nodes, which may interfere with the originally clear and reliable node features and even reduce the accuracy of the prediction of these healthy nodes. In an optional embodiment, the nodes for graph convolution in the spatiotemporal graph are determined based on the characteristics of the sub-region nodes, and graph convolution is performed on the nodes that need to be graph convolved, including:

[0066] For each sub-region node, based on its current and historical plant physiological data, soil data, and environmental data, the risk value and confidence level of abnormal growth in the next period are predicted;

[0067] When the growth anomaly risk value of a sub-region node is higher than the preset risk threshold and its corresponding confidence level is lower than the preset confidence level, the sub-region node is marked as a risk node;

[0068] Obtain the previous and next spatiotemporal graphs of the spatiotemporal graph where the risk node is located, take the nodes in the previous and next spatiotemporal graphs that are identical to the risk node as adjacent points of the risk node, and obtain the adjacent points in the spatiotemporal graph where the risk node is located;

[0069] Graph convolution is performed on the risk node using the risk node and its adjacent nodes.

[0070] For each sub-region node, a lightweight prediction model is used to predict the probability of abnormal growth in each sub-region node within the next preset period, using plant physiological data such as leaf spectra and growth rate changes, soil data, and environmental data collected at the current moment and over a period of time. The model then outputs a growth anomaly risk value and a confidence level for the prediction. The risk value ranges from [0, 1], with higher values ​​indicating greater risk. The confidence level also ranges from [0, 1], with higher values ​​indicating greater confidence in the model's risk prediction. If a sub-region node has a high risk but low confidence level, the node is marked as a risk node.

[0071] In the current time-space graph of the risk node, find the adjacent sub-region nodes directly connected to the risk node. In order to introduce the information of the time dimension, obtain the corresponding nodes in the previous time-space graph and the next time-space graph of the risk node. These two nodes are the adjacent points of the risk node in the time dimension. The above spatial adjacent points and temporal adjacent points are used as the adjacent point set for the risk node when performing graph convolution analysis. Figure 4 The risk node and its neighboring nodes are shown. Using the preliminary features of the risk node itself and the features of all determined neighboring nodes, one or more graph convolution operations are performed.

[0072] In an optional embodiment, performing multi-instance learning on the instance bag of the sub-region to identify key instance patterns of early stress includes:

[0073] For the instance package of the sub-region where the growth anomaly score exceeds the threshold, each instance feature in the instance package is input into the fully connected layer and then subjected to a nonlinear activation function to obtain the stress contribution score;

[0074] Selecting at least one instance whose coercion contribution score is higher than a preset contribution threshold as a key instance;

[0075] A key instance pattern of early coercion is obtained according to the characteristics of the key instance.

[0076] Subregions with growth anomaly scores greater than a dynamically adjusted anomaly threshold based on the global disease probability are considered to have a higher overall risk. The instance package for this target subregion is retrieved. An instance package contains multiple instances, each of which is a feature vector representing the combination of local plant physiological data within the subregion with its corresponding environmental data, soil, and other parameters. The feature vector of each instance in the instance package is input into one or more preset fully connected layers. After processing by the fully connected layers, it is passed through a nonlinear activation function such as Sigmoid, ReLU, or Tanh to output a stress contribution score for each instance in the instance package. A higher score indicates a higher likelihood that the instance is directly related to the occurrence of stress. For example, if an instance represents an abnormally elevated temperature in a certain area of ​​the leaf accompanied by a specific spectral reflectance anomaly, a combination that has historically been highly correlated with the early stages of a disease, the instance is likely to receive a higher stress contribution score.

[0077] A contribution threshold is pre-set, derived from experience or statistical analysis. All instances in the target subregion's instance package are traversed, and the coercion contribution score of each instance is compared with the pre-set contribution threshold. If the coercion contribution score of an instance exceeds the threshold, the instance is selected and marked as a key instance. A subregion may have one or more key instances. The original feature vectors of all selected key instances are combined to obtain a key instance pattern. This combination can be performed by methods including, but not limited to, direct concatenation and calculation of the mean vector.

[0078] In an optional embodiment, the adjustment parameters of the disease-resistant nutrient solution are calculated based on the disease probability and key instance pattern in combination with a preset stress type adjustment model, including:

[0079] Input the predicted specific disease probability and key instance pattern into the preset adjustment model to obtain the adjustment amplitude coefficient;

[0080] The reference concentration is adjusted based on the adjustment amplitude coefficient to obtain an adjusted nutrient solution parameter.

[0081] The predicted specific disease probability and key instance patterns are input into a preset adjustment model to obtain an adjustment coefficient. In one embodiment, the adjustment model is a decision tree or neural network model. An exemplary adjustment coefficient is 1.1 or 0.8. In one embodiment, the nutrient solution contains not only nutrients and trace elements, but also pesticides.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some feature data can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0083] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0084] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0085] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Claims

1. A method for monitoring the banana seedling cultivation process using a disease-resistant nutrient solution, characterized in that: include: Soil data, environmental data, and plant physiological data are collected from a banana seedling cultivation area, the banana seedling cultivation area is divided into multiple sub-areas, edge weights are calculated based on the soil and environmental data, and a spatiotemporal graph is constructed based on the weights and with the sub-areas as nodes. A new spatiotemporal graph is generated when changes in the plant physiological data exceed a preset fluctuation threshold. For each sub-region node, an instance package is obtained based on local plant physiological data and corresponding soil data and environmental data. The instance package of the sub-region node is used to obtain the characteristics of the weighted aggregation of instance attention within each node. The nodes for graph convolution in the spatiotemporal graph are determined based on the characteristics of the sub-region nodes. The graph convolution is performed on the nodes that need to be convolved. The time series characteristics of the nodes are obtained based on the characteristics of the same sub-region nodes in multiple spatiotemporal graphs. The time series characteristics are used to predict the future growth anomaly score and specific disease probability of each sub-region. When the growth anomaly score of a sub-region is greater than an anomaly threshold dynamically adjusted based on the global disease probability, multi-instance learning is performed on the instance package of the sub-region to identify key instance patterns of early stress; based on the disease probability and key instance patterns, combined with a preset regulation model for the stress type, adjustment parameters of the disease-resistant nutrient solution are calculated, and the disease-resistant nutrient solution is adjusted using the adjustment parameters; The performing multi-instance learning on the instance bag of the sub-region to identify key instance patterns of early stress includes: For the instance package of the sub-region where the growth anomaly score exceeds the threshold, each instance feature in the instance package is input into the fully connected layer and then subjected to a nonlinear activation function to obtain the stress contribution score; Selecting at least one instance whose coercion contribution score is higher than a preset contribution threshold as a key instance; A key instance pattern of early coercion is obtained according to the characteristics of the key instance.

2. The method according to claim 1, characterized in that The calculation of edge weights based on soil data and environmental data includes: Obtain soil and environmental data for each sub-region; The weights of the edges between nodes are calculated based on the similarity of the soil data and environmental data between sub-regions.

3. The method according to claim 1, characterized in that The determining of nodes for performing graph convolution in the spatiotemporal graph according to the characteristics of the sub-region nodes, and performing graph convolution on the nodes requiring graph convolution, includes: For each sub-region node, based on its current and historical plant physiological data, soil data, and environmental data, the risk value and confidence level of abnormal growth in the next period are predicted; When the risk value of growth anomaly of a sub-region node is higher than the preset risk threshold and its corresponding confidence level is lower than the preset confidence level, the sub-region node is marked as a risk node; Obtain the previous and next spatiotemporal graphs of the spatiotemporal graph where the risk node is located, take the nodes in the previous and next spatiotemporal graphs that are identical to the risk node as adjacent points of the risk node, and obtain the adjacent points in the spatiotemporal graph where the risk node is located; Graph convolution is performed on the risk node using the risk node and its adjacent nodes.

4. The method according to claim 1, wherein The adjustment parameters of the disease-resistant nutrient solution are calculated based on the disease probability and key instance pattern, combined with the preset stress type adjustment model, including: Input the predicted specific disease probability and key instance pattern into the preset adjustment model to obtain the adjustment amplitude coefficient; The reference concentration is adjusted based on the adjustment amplitude coefficient to obtain an adjusted nutrient solution parameter.

5. A banana seedling cultivation process monitoring system using disease-resistant nutrient solution, characterized in that: include: a graph construction unit for collecting soil data, environmental data, and plant physiological data of a banana seedling cultivation area, dividing the banana seedling cultivation area into a plurality of sub-areas, calculating edge weights based on the soil data and the environmental data, constructing a spatiotemporal graph based on the weights and using the sub-areas as nodes, and generating a new spatiotemporal graph when a change in the plant physiological data exceeds a preset fluctuation threshold; A prediction unit is configured to obtain, for each sub-region node, an instance package based on local plant physiological data and corresponding soil and environmental data; use the instance package of the sub-region node to obtain features of weighted aggregation of instance attention within each node; determine nodes for graph convolution in the spatiotemporal graph based on the features of the sub-region nodes; perform graph convolution on the nodes that require graph convolution; obtain temporal features of the nodes based on features of multiple nodes in the same sub-region in the spatiotemporal graph; and use the temporal features to predict future growth anomaly scores and specific disease probabilities for each sub-region; an adjustment unit configured to, when the growth anomaly score of a sub-region is greater than an anomaly threshold dynamically adjusted based on the global disease probability, perform multi-instance learning on the instance package of the sub-region to identify key instance patterns of early stress; calculate adjustment parameters for the disease-resistant nutrient solution based on the disease probability and key instance patterns, in combination with a preset adjustment model for the stress type, and adjust the disease-resistant nutrient solution using the adjustment parameters; The performing multi-instance learning on the instance bag of the sub-region to identify key instance patterns of early stress includes: For the instance package of the sub-region where the growth anomaly score exceeds the threshold, each instance feature in the instance package is input into the fully connected layer and then subjected to a nonlinear activation function to obtain the stress contribution score; Selecting at least one instance whose coercion contribution score is higher than a preset contribution threshold as a key instance; A key instance pattern of early coercion is obtained according to the characteristics of the key instance.

6. The system according to claim 5, characterized in that The calculation of edge weights based on soil data and environmental data includes: Obtain soil and environmental data for each sub-region; The weights of the edges between nodes are calculated based on the similarity of the soil data and environmental data between sub-regions.

7. The system according to claim 5, characterized in that The determining of nodes for performing graph convolution in the spatiotemporal graph according to the characteristics of the sub-region nodes, and performing graph convolution on the nodes requiring graph convolution, includes: For each sub-region node, based on its current and historical plant physiological data, soil data, and environmental data, the risk value and confidence level of abnormal growth in the next period are predicted; When the risk value of growth anomaly of a sub-region node is higher than the preset risk threshold and its corresponding confidence level is lower than the preset confidence level, the sub-region node is marked as a risk node; Obtain the previous and next spatiotemporal graphs of the spatiotemporal graph where the risk node is located, take the nodes in the previous and next spatiotemporal graphs that are identical to the risk node as adjacent points of the risk node, and obtain the adjacent points in the spatiotemporal graph where the risk node is located; Graph convolution is performed on the risk node using the risk node and its adjacent nodes.

8. The system according to claim 5, wherein: The adjustment parameters of the disease-resistant nutrient solution are calculated based on the disease probability and key instance pattern, combined with the preset stress type adjustment model, including: Input the predicted specific disease probability and key instance pattern into the preset adjustment model to obtain the adjustment amplitude coefficient; The reference concentration is adjusted based on the adjustment amplitude coefficient to obtain an adjusted nutrient solution parameter.

Images