An environment perception data determination method, apparatus, device, and storage medium

Abstract

This application provides a method, apparatus, equipment, and storage medium for determining environmental sensing data, comprising: firstly, selecting a model of facility agriculture environmental control equipment, deploying multiple environmental parameter sensing instruments according to their spatial volume parameters, and recording the relative spatial positions of each instrument; establishing a unified spatiotemporal network coordinate system, mapping the positions of the sensing instruments and product quality sampling monitoring points to network coordinates respectively; constructing a multi-channel spatial input tensor containing instrument coordinates and sensing data, performing three-dimensional spatial diffusion and time accumulation processing through a neural network to generate a dynamic environmental feature field; finally, obtaining the dynamic network coordinates of the sampling points, and extracting the environmental feature vector under the corresponding spatiotemporal coordinates through interpolation sampling, as the environmental sensing data of the monitoring point. Using the above method, the accuracy and effectiveness of facility agriculture environmental control can be effectively guaranteed.

Description

Technical Field

[0001] This application relates to the field of smart agriculture data processing technology, and more specifically, to a method, apparatus, device, and storage medium for determining environmental perception data. Background Technology

[0002] In recent years, environmental parameter sensing equipment has been widely deployed in facility agriculture, and continuous monitoring of the production environment throughout the entire growth period has become a routine preliminary step in environmental regulation. Crop growth models built upon environmental sensing data and product quality data are the core foundation for achieving intelligent and precise management, directly determining the production efficiency and product quality of facility agriculture. During the construction of growth models, spatial matching of environmental and quality data is a prerequisite for ensuring the effectiveness of the model. However, due to deployment costs and operational constraints, the installation locations of environmental sensing instruments and the collection points of product samples often cannot completely coincide. Therefore, corresponding technical means are needed to calculate the environmental sensing data at the sampling sites to achieve spatial dimensional matching between the two types of data.

[0003] Currently, to address the issue of non-overlapping locations, the industry generally adopts spatial interpolation: independent spatial interpolation calculations are performed on the discretely deployed environmental parameter sensing data and the dispersed product quality data to generate continuous interpolation results covering the entire space. Then, the two types of interpolation results are matched one-to-one according to the spatial location to obtain the environmental sensing data corresponding to the sample collection site.

[0004] However, this technical solution has obvious drawbacks: the interpolation operation itself relies on smoothing processing to achieve continuous filling of spatial data, which inevitably weakens and erases the local features and spatial heterogeneity of environmental parameters, resulting in a significant deviation between the environmental data of the sampling sites obtained in the final calculation and the true values. This directly reduces the modeling accuracy of the subsequent product growth model and makes it difficult to support the precise environmental control needs of facility agriculture. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, apparatus, equipment and storage medium for determining environmental sensing data, which can effectively ensure the accuracy and effectiveness of environmental regulation in facility agriculture.

[0006] In a first aspect, embodiments of this application provide a method for determining environmental perception data, the method comprising: Select the model of environmental control equipment for facility agriculture; Based on the spatial volume parameters corresponding to the model, multiple environmental parameter sensing instruments are deployed, and the relative spatial positions of each environmental parameter sensing instrument are recorded. A unified spatiotemporal network coordinate system is established based on the spatial volume parameters, and the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point are respectively mapped to network coordinates. Construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data; The input tensor is processed by a neural network to perform three-dimensional spatial diffusion and temporal accumulation, thereby generating a dynamic environmental feature field based on the spatiotemporal network coordinate system. Obtain the dynamic network coordinates of the product quality sampling and monitoring points; Using an interpolation sampling method, with the dynamic network coordinates as the sampling position, environmental feature vectors under the corresponding spatiotemporal coordinates are extracted from the dynamic environmental feature field, which are used as environmental perception data for the product quality sampling monitoring points.

[0007] Optionally, the deployment of multiple environmental parameter sensing instruments includes: Using the center point of the facility agriculture environmental control equipment as a reference, an environmental parameter sensing instrument is installed every 4 cubic meters. Record the three-dimensional relative spatial position of each environmental parameter sensing instrument relative to the center point.

[0008] Optionally, establishing a unified spatiotemporal network coordinate system and mapping the relative spatial positions and product quality sampling monitoring points to network coordinates includes: The space of the facility agriculture environmental control equipment is divided into a three-dimensional voxel grid. The relative spatial positions of the environmental parameter sensing instruments are mapped to the corresponding voxel network coordinates; After the location of the product quality sampling monitoring points is restored to three-dimensional physical coordinates through image recognition and depth measurement, it is then mapped to network coordinates in the voxel grid.

[0009] Optionally, constructing the multi-channel spatial input tensor includes: Construct a three-dimensional voxel tensor containing a data channel, a mask channel, a coordinate channel, and a time channel; The data channel is used to fill in environmental perception data at the network coordinates corresponding to the instrument. The mask channel is used to identify the network coordinates of the instrument. The coordinate channels are used to represent the normalized spatial coordinates of each voxel. The time channel is used to represent the sampling sequence in the time dimension.

[0010] Optionally, the three-dimensional spatial diffusion and temporal accumulation processing using a neural network includes: A neural network with 3D convolution capability and gated recurrent mechanism is used; The discrete environmental perception data is diffused outwards in a spatial dimension through three-dimensional convolution operations. It also accumulates historical environmental information of each network coordinate position in the time dimension through a gating loop mechanism, and outputs a dynamic environmental feature field containing spatiotemporal information.

[0011] Optionally, the extraction of environmental feature vectors using the interpolation sampling method includes: The trilinear interpolation method is used; When the dynamic network coordinates fall inside the cell of the dynamic environment feature field, the environmental feature vector of that point is calculated by weighted aggregation based on the distance between the coordinates and the adjacent voxel nodes in three axes.

[0012] Optionally, the method further includes: The environmental perception data of the product quality sampling monitoring points obtained by extraction is associated with the product quality data of the monitoring points to construct or optimize the product growth model. After extraction, multiple points are randomly selected to collect real environmental data and compare them with the extraction results for verification until the extraction results meet the preset inspection standards.

[0013] Secondly, embodiments of this application provide an environmental perception data determination device, the device comprising: The model selection module is used to select the model of environmental control equipment for facility agriculture; The instrument deployment module is used to deploy multiple environmental parameter sensing instruments according to the spatial volume parameters corresponding to the model, and record the relative spatial position of each environmental parameter sensing instrument. The coordinate mapping module is used to establish a unified spatiotemporal network coordinate system based on the spatial volume parameters, and to map the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point into network coordinates respectively. The tensor construction module is used to construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data. The feature generation module is used to perform three-dimensional spatial diffusion and temporal accumulation processing on the input tensor using a neural network to generate a dynamic environmental feature field based on the spatiotemporal network coordinate system. The coordinate acquisition module is used to acquire the dynamic network coordinates of the product quality sampling monitoring points; The data extraction module is used to extract environmental feature vectors in the corresponding spatiotemporal coordinates from the dynamic environmental feature field using an interpolation sampling method, with the dynamic network coordinates as the sampling position, as environmental perception data for the product quality sampling monitoring points.

[0014] Optionally, the instrument deployment module is specifically used for: Using the center point of the facility agriculture environmental control equipment as a reference, an environmental parameter sensing instrument is installed every 4 cubic meters. Record the three-dimensional relative spatial position of each environmental parameter sensing instrument relative to the center point.

[0015] Optionally, the coordinate mapping module is specifically used for: The space of the facility agriculture environmental control equipment is divided into a three-dimensional voxel grid. The relative spatial positions of the environmental parameter sensing instruments are mapped to the corresponding voxel network coordinates; After the location of the product quality sampling monitoring points is restored to three-dimensional physical coordinates through image recognition and depth measurement, it is then mapped to network coordinates in the voxel grid.

[0016] Optionally, the tensor construction module is specifically used for: Construct a three-dimensional voxel tensor containing a data channel, a mask channel, a coordinate channel, and a time channel; The data channel is used to fill in environmental perception data at the network coordinates corresponding to the instrument. The mask channel is used to identify the network coordinates of the instrument. The coordinate channels are used to represent the normalized spatial coordinates of each voxel. The time channel is used to represent the sampling sequence in the time dimension.

[0017] Optionally, the feature generation module is specifically used for: A neural network with 3D convolution capability and gated recurrent mechanism is used; The discrete environmental perception data is diffused outwards in a spatial dimension through three-dimensional convolution operations. It also accumulates historical environmental information of each network coordinate position in the time dimension through a gating loop mechanism, and outputs a dynamic environmental feature field containing spatiotemporal information.

[0018] Optionally, the data extraction module is specifically used for: The trilinear interpolation method is used; When the dynamic network coordinates fall inside the cell of the dynamic environment feature field, the environmental feature vector of that point is calculated by weighted aggregation based on the distance between the coordinates and the adjacent voxel nodes in three axes.

[0019] Optionally, the device further includes a verification and optimization module for: The environmental perception data of the product quality sampling monitoring points obtained by extraction is associated with the product quality data of the monitoring points to construct or optimize the product growth model. After extraction, multiple points are randomly selected to collect real environmental data and compare them with the extraction results for verification until the extraction results meet the preset inspection standards.

[0020] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the environmental perception data determination method described in any of the optional embodiments of the first aspect are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the environmental perception data determination method described in any of the optional embodiments of the first aspect.

[0022] The technical solution provided in this application includes, but is not limited to, the following beneficial effects: Selecting the model of the environmental control equipment for facility agriculture can provide a unified physical spatial benchmark for subsequent sensor deployment and coordinate system construction, realize a one-to-one correspondence between the equipment model and spatial alignment parameters, ensure the reusability of the method in the same type of facility scenario, and avoid the cost loss caused by repeated calibration of the same type of equipment.

[0023] By deploying multiple environmental parameter sensing instruments according to the spatial volume parameters corresponding to the model and recording the relative spatial position of each instrument, the uniform distribution of sensing points within the facility space can be ensured, providing a comprehensive original data source for subsequent spatial calculations. Accurate recording of relative positions provides precise location basis for subsequent coordinate mapping, ensuring the spatial traceability of the original sensing data.

[0024] Based on the aforementioned spatial volume parameters, a unified spatiotemporal network coordinate system is established. The relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point are mapped to network coordinates respectively. This allows two types of points belonging to different systems to be included in the same set of spatial coordinate benchmarks, eliminating data deviations caused by the spatial inconsistency between the two types of points from the root, and providing a unified spatial reference for subsequent data alignment processing.

[0025] Constructing a multi-channel spatial input tensor that includes the network coordinates of various environmental parameter sensing instruments and their environmental sensing data can structurally integrate discrete environmental sensing values ​​with corresponding spatial location information, enabling the input data to carry both parameter values ​​and spatial attributes. This provides a standardized input format for subsequent neural network processing, helping the model to effectively identify the spatial distribution characteristics of the data.

[0026] By using a neural network to perform three-dimensional spatial diffusion and temporal accumulation processing on the input tensor, a dynamic environmental feature field based on the spatiotemporal network coordinate system is generated. This can simultaneously extend information in the spatial dimension and accumulate information in the temporal dimension: it expands discrete single-point sensing data into a continuous environmental distribution covering the entire space, solving the problem of missing spatial data caused by sparse sampling points; it can also record the temporal evolution of environmental parameters, reflecting the cumulative effect of the environment on crop growth, and ultimately forming a spatiotemporally integrated dynamic environmental feature field.

[0027] By acquiring the dynamic network coordinates of the product quality sampling monitoring points, the spatial position of the product samples in a unified coordinate system can be located in real time and accurately. This adapts to the dynamic scenarios of changes in position and height during crop growth, providing accurate sampling location basis for subsequent precise extraction of environmental data at the corresponding points.

[0028] By using the interpolation sampling method, with the dynamic network coordinates as the sampling position, the environmental feature vectors under the corresponding spatiotemporal coordinates are extracted from the dynamic environmental feature field as environmental perception data of the product quality sampling monitoring point. This can directly obtain environmental data that perfectly matches the spatiotemporal position of the product quality sampling monitoring point, achieving precise spatial alignment between environmental perception data and product quality data, and ensuring that the output data is highly consistent with the actual growth environment of the crop.

[0029] In summary, this solution effectively solves the technical problem of non-overlapping environmental sensing points and product sampling points in facility agriculture through end-to-end processing including unified spatial benchmarks, structured encoding of sensing data, construction of spatiotemporal feature fields, and precise sampling of points. It significantly improves the accuracy of determining environmental sensing data at non-sensor sites, provides reliable data support for the high-precision construction of crop growth models, and can effectively ensure the accuracy and effectiveness of environmental regulation in facility agriculture.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of an environmental perception data determination method provided in Embodiment 1 of this application is shown; Figure 2 A flowchart of an environmental perception data determination method provided in Embodiment 1 of this application is shown; Figure 3 This illustration shows a schematic diagram of an environmental sensing deployment scenario for a glass greenhouse in facility agriculture, as provided in Embodiment 1 of this application. Figure 4 This diagram illustrates a hierarchical architecture of a spatiotemporal feature field alignment system based on a 3D convolutional recurrent network, as provided in Embodiment 1 of this application. Figure 5 This illustration shows an architecture diagram of a 3D-CoordConv coordinate injection module provided in Embodiment 1 of this application; Figure 6 This diagram illustrates the internal structure of a 3D-ConvGRU provided in Embodiment 1 of this application. Figure 7 This document shows a flowchart illustrating a small-scale spatial alignment technique for agricultural environmental sensing data provided in Embodiment 1 of this application. Figure 8 The flowchart of an environmental parameter sensing instrument deployment method provided in Embodiment 1 of this application is shown; Figure 9 A flowchart of a network coordinate mapping method provided in Embodiment 1 of this application is shown; Figure 10 A flowchart of a neural network processing method provided in Embodiment 1 of this application is shown; Figure 11 A schematic diagram of an environmental feature vector extraction method provided in Embodiment 1 of this application is shown; Figure 12 A flowchart of a data verification method provided in Embodiment 1 of this application is shown; Figure 13 This paper shows a schematic diagram of the structure of an environmental perception data determination device provided in Embodiment 2 of this application; Figure 14 A schematic diagram of the structure of a computer device provided in Embodiment 3 of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] Example 1 This method addresses the need for small-scale spatial alignment between environmental perception data and product quality data in facility agriculture, and is intended to solve the problem of insufficient accuracy of existing spatial alignment methods in agricultural scenarios.

[0035] In recent years, environmental parameter sensing equipment has been rapidly adopted and popularized in facility agriculture. Sensing agricultural production environmental parameters throughout the entire growth cycle has become a routine pre-emptive measure for environmental control. Various product growth models for precise environmental management of agricultural products are constantly emerging, driving the development of intelligent production technologies throughout the entire agricultural production process. The core prerequisite for constructing a high-precision product growth model is accurately determining the real environmental sensing data corresponding to the crop growth site; the spatial matching degree of the data directly determines the final effectiveness of the model.

[0036] Product growth models must be constructed based on environmental sensing data and product quality testing data throughout the product's entire life cycle. Production environment parameters can be obtained through environmental sensing instruments, while product quality testing data can be obtained through quantitative testing using laboratory instruments. Both types of data are guaranteed by strict technical standards and standardized operating procedures in terms of both quantity and quality. However, environmental sensing instruments can only collect environmental data from their own deployment sites and cannot directly measure the environmental parameters at product sampling sites. Therefore, it is necessary to indirectly determine the environmental sensing data corresponding to the sampling sites through specific methods to achieve the matching of the two types of data.

[0037] Because the installation sites of environmental parameter sensing instruments typically do not coincide with the sample collection sites, these two sets of data are prone to spatial discrepancies. Furthermore, environmental sensing data exhibits significant three-dimensional non-uniformity, all of which affect the modeling accuracy of subsequent product growth models and reduce environmental control efficiency. Spatial discrepancies and three-dimensional non-uniformity directly lead to the inability to accurately determine the true environmental sensing data corresponding to the product sampling points. Directly using data from nearby sensors will introduce significant systematic errors.

[0038] Domestic and international methods for handling non-overlapping spatial sampling points mainly fall into three categories: The first is spatial interpolation, which involves spatially interpolating environmental parameter sensing data and product quality data separately, and then matching the interpolation results to their corresponding spatial locations. The second is alignment methods based on graph neural networks, which treat discrete data as nodes in a graph, using spatial distance as edges, and achieving alignment through graph convolution. The third is coordinate encoding and embedding methods, which input physical coordinates as feature inputs, mapping data points to high-dimensional vectors, enabling the model to learn the spatial heterogeneity of different locations. The core objective of all three methods is to determine the environmental sensing data at non-sensor deployment sites through algorithmic calculation, thereby achieving spatial data alignment.

[0039] While existing methods have addressed the spatial data alignment problem to some extent, three shortcomings remain in practical farmland applications: First, a single interpolation algorithm smooths out local environmental features, leading to decreased data accuracy; second, static alignment ignores the cumulative impact of the environment on product quality; and third, when sample points are sparse, traditional convolutional kernels are inefficient in processing such extremely sparse data. These three deficiencies ultimately result in significant discrepancies between the calculated environmental perception data and the actual values, failing to meet the requirements for constructing high-precision growth models.

[0040] In summary, existing spatial data alignment methods generally face problems such as sparse sampling points, imprecise models, and non-uniform 3D environments. Improving the spatial alignment between environmental parameter sensing data and product quality detection data at specific sample sites is a fundamental technical problem that urgently needs to be solved to improve the accuracy of product growth models. In other words, how to accurately determine the environmental sensing data corresponding to any product sampling site in small-scale, sparsely sampled, and 3D non-uniform facility agriculture scenarios is a core technical bottleneck that urgently needs to be overcome in the current field of smart agriculture.

[0041] The purpose of this invention is to address the problems of sparse sampling points, imprecise spatial alignment models, and non-uniform 3D environments in existing small-scale environmental parameter sensing data and product quality data, as well as the spatial alignment requirements in intelligent production technology throughout the agricultural production process. This invention proposes a small-scale spatial alignment method for agricultural environmental sensing data based on 3D-CoordConv (3D CoordinateConvolution) and 3D-ConvGRU (3D Convolutional Gated Recurrent Unit) for 3D space. This method achieves high-precision determination of environmental sensing data at any product sampling point through spatiotemporal joint encoding and feature sampling, fundamentally solving the data matching error caused by spatial non-overlap.

[0042] See Figure 1 As shown, Figure 1 The diagram illustrates a small-scale spatial alignment technology roadmap for agricultural environmental sensing data provided in Embodiment 1 of this application. This diagram fully presents the closed-loop technical architecture of the environmental sensing data determination method of this application, which adopts the logic of "dual-branch parallel processing - end feature alignment - verification backflow optimization", corresponding one-to-one with the core execution steps of the method of this application.

[0043] The process begins with the multi-sensor joint calibration stage, which completes the spatial deployment and position calibration of all environmental sensing instruments; then it enters the stage of constructing a spatiotemporal network coordinate system, which divides the internal space of facility agriculture into a unified three-dimensional voxel grid coordinate system, thereby splitting out two parallel processing branches.

[0044] The left-side environmental data processing branch sequentially executes the following steps: determining the environmental parameters and sensing the instrument grid coordinates, mapping the instrument's physical location to voxel network coordinates; establishing a 6-channel 3D voxel tensor map, fusing environmental data, mask labels, and 3D coordinate information; and using 3D-ConvGRU to complete 3D convolutional diffusion, realizing spatial diffusion and temporal cumulative encoding of discrete data, ultimately generating a dynamic environmental feature field.

[0045] The right-hand product data processing branch: sequentially executes the determination of product grid coordinates, obtains the three-dimensional spatial coordinates of crops through visual inspection and depth perception; the product grid coordinates are associated with quality control data, binding the crop's unique ID, quality indicators and corresponding spatial coordinates.

[0046] The processing results from the two branches are merged into the 3D Grid Sample alignment stage, which extracts the environmental features corresponding to the product locations from the dynamic environmental feature field, and completes the spatial alignment of environmental perception data and product quality data. The alignment result enters the real value verification stage. If the verification fails, the process is returned to the front-end multi-sensor joint calibration stage, and the instrument deployment scheme or model parameters are adjusted before the entire process is re-executed. If the verification passes, the final result is output, the setting parameters in the model are obtained, and a dedicated spatial alignment model adapted to the current facility scenario is formed.

[0047] To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart illustrating the method for determining environmental perception data provided in Embodiment 1 of this application will be used to provide a detailed description of Embodiment 1 of this application.

[0048] See Figure 2 As shown, Figure 2 The flowchart of an environmental perception data determination method provided in Embodiment 1 of this application is shown. The method adopts a technical approach of parallel dual-branch encoding of environmental data and association of product data, end feature alignment, and closed-loop verification optimization. The method includes steps S201 to S207: S201: Select the model of environmental control equipment for facility agriculture.

[0049] Specifically, the selected facility agriculture environmental control equipment is a specific model of arched glass greenhouse. When selecting the model, the installation direction of the greenhouse must be determined simultaneously.

[0050] Specifically, the greenhouse is equipped with environmental control and auxiliary monitoring devices such as fan control modules, atomizing spray nozzles, and precipitation measuring barrels, which are typically used for the cultivation and management of facility crops such as morel mushrooms throughout their entire growth period.

[0051] Specifically, for each type of greenhouse and each erection direction, a set of exclusive small-scale spatial alignment parameters for environmental perception data needs to be constructed. These alignment parameters are the final target of this method.

[0052] See Figure 3 As shown, Figure 3 This diagram illustrates a scenario of environmental sensing deployment in a glass greenhouse for facility agriculture, as provided in Embodiment 1 of this application. The diagram showcases the physical environment of the arched glass greenhouse to which the method of this application is applied, visually presenting the spatial layout of the environmental sensing equipment and the distribution of product sampling points, corresponding to the actual deployment of the physical sensing layer in this application. The diagram clearly indicates that the environmental parameter sensing instruments can collect various agricultural environmental parameters such as air temperature, air humidity, soil moisture, soil temperature, and light intensity. The sensing instruments adopt a vertically layered deployment scheme, with probes at heights of 2.5m, 1.5m, and 0.5m, respectively, combined with a uniform horizontal deployment, conforming to the deployment rule of this application of "one environmental parameter sensing instrument per 4 cubic meters." Sampling points for product quality data collection are marked inside the greenhouse, corresponding to the product quality sampling and monitoring points in this application, which can collect quality indicators such as product height and thickness, providing quality data for subsequent growth model construction. The greenhouse is equipped with fan control modules, misting nozzles, and precipitation measuring buckets, among other environmental control and auxiliary monitoring devices, which together form a complete facility agriculture environmental sensing and control system to ensure that the environmental parameters inside the greenhouse are measurable and adjustable.

[0053] S202: Based on the spatial volume parameters corresponding to the model, deploy multiple environmental parameter sensing instruments and record the relative spatial positions of each environmental parameter sensing instrument.

[0054] Specifically, the instruments are deployed according to a density standard of one environmental parameter sensing instrument per 4 cubic meters. A sampling point is set up every 2 meters in the horizontal direction, and detection probes are set up at three heights of 0.5m, 1.5m and 2.5m above the ground in the vertical direction to achieve uniform coverage of three-dimensional space.

[0055] Specifically, before deployment, the overall dimensions and center point of the selected glass greenhouse are determined. The first environmental parameter sensing instrument is set up at the center position at a height of 0.5 meters above the ground. This instrument covers a space of 4 cubic meters (2m long × 2m wide × 1m high) at the center position.

[0056] Specifically, starting from the instrument at the center, environmental parameter sensing instruments are deployed around the greenhouse at 4-cubic-meter intervals. At the same time, based on the center point of the greenhouse, the three-dimensional relative spatial position of each environmental parameter sensing device is accurately recorded, with position data accurate to the centimeter level.

[0057] Specifically, the environmental parameter sensing instruments can collect monitoring indicators including air temperature, air humidity, soil moisture, soil temperature, light intensity, carbon dioxide concentration, oxygen concentration, and other agricultural environmental parameters, providing multi-dimensional raw data sources for subsequent feature field construction.

[0058] S203: Establish a unified spatiotemporal network coordinate system based on the spatial volume parameters, and map the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point to network coordinates respectively.

[0059] Specifically, the core of establishing a spatiotemporal network coordinate system is to incorporate discrete sensing instrument locations and product monitoring locations into the same three-dimensional spatial coordinate system, thereby solving the data deviation problem caused by the non-overlapping of the two types of locations, while also covering the environmental non-uniformity in the vertical direction.

[0060] Specifically, this method adopts a four-layer architecture to achieve end-to-end processing. The hierarchical relationship from physical data acquisition to final feature output is as follows: Figure 4 As shown, this step follows the physical perception layer and enables the data voxelization layer. First, the complete internal space of the glass greenhouse is divided into a regular three-dimensional voxel grid as a unified spatial coordinate reference.

[0061] Specifically, the physical location of the environmental parameter sensing instruments is fixed, and their three-dimensional relative spatial position can be directly mapped to the network coordinates corresponding to the voxel grid. When deploying the instruments, the environmental sensing instruments will be evenly deployed at different height layers to ensure the coverage of spatial information in the vertical dimension.

[0062] Specifically, the product quality sampling and monitoring points are dynamic points, corresponding to the crop monitoring link of the physical perception layer. The FairMOT target detection algorithm identifies crop samples in the greenhouse in real time and outputs detection boxes. Combined with a depth camera, the product height data is directly obtained. Then, perspective transformation is used to restore the two-dimensional points of the image to three-dimensional physical coordinates, and finally mapped to the network coordinates in the voxel grid.

[0063] See Figure 4 As shown, Figure 4This diagram illustrates a hierarchical architecture of a spatiotemporal feature field alignment system based on a 3D convolutional recurrent network, as provided in Embodiment 1 of this application. The diagram fully demonstrates the four-layer hierarchical architecture of the spatiotemporal feature field alignment system based on a 3D convolutional recurrent network proposed in this application. The system completes the entire chain processing from the acquisition of raw physical data to the final feature alignment output from bottom to top, corresponding one-to-one with the core technical steps of the method in this application.

[0064] First layer: Physical perception layer (illustration of a physical greenhouse scene) This layer serves as the data input source for the system, and is marked with sensor instruments (for user reference). The sensing points cover three height levels vertically: 0.5m, 1.5m, and 2.5m, and horizontally follow a density of 2-meter intervals (one sampling point every 4 cubic meters). The FairMOT target detection algorithm is used to identify Deer Antler Mushrooms, and a detection box (Bounding Box) is output to locate the spatial position of the crop. This corresponds to the basic link of environmental data collection and dynamic identification of product locations in this application.

[0065] Second layer: Data voxelization layer This layer completes the voxelization and coordinate injection of discrete data. Through the 3D-CoordConv coordinate injection mechanism, X-coordinate watermarks, Y-coordinate watermarks, and Z-coordinate watermarks are added to the three-dimensional voxel mesh respectively. The absolute spatial coordinate (x, y, z) information is embedded into the features of each voxel, which corresponds to the processing steps of constructing multi-channel 3D input tensors and injecting spatial coordinate information in this application, enabling the model to perceive spatial heterogeneity.

[0066] Third layer: Core processing layer (3D-ConvGRU encoder) This layer is the core computing unit of the system. It uses a 3D-ConvGRU encoder to encode spatiotemporal features, achieves temporal dimension folding through a recurrent gating mechanism, and realizes spatial diffusion through a 3D convolution operator, ultimately outputting a 4DEnvironmental feature field. (4-Dimensional Environmental Feature Field, where) (representing the environmental feature field at time t), corresponding to the core processing step in this application where the neural network performs three-dimensional spatial diffusion and temporal accumulation to generate a dynamic environmental feature field.

[0067] Fourth layer: Feature alignment and sampling layer (3D mesh sampling) This layer achieves accurate feature extraction of product locations. It adopts the 3D Grid Sample sampling method and extracts the feature vector of the target location from the four-dimensional environmental feature field using a star-shaped sampling logic. Finally, it outputs aligned quality-environment feature pairs, which corresponds to the final step in this application of extracting product location environmental features and completing spatial alignment through trilinear interpolation sampling.

[0068] S204: Construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data.

[0069] Specifically, the constructed multi-channel input is a 6-channel 3D voxel tensor map with the tensor format [Batch, Time, Channels, H, W, D], corresponding to the input requirements of the 3D-CoordConv coordinate injection module. The meaning of each dimension can be found in the tensor dimension annotation.

[0070] Specifically, Batch represents the batch dimension, Time represents the standardized continuous time sampling sequence, Channels represents the channel dimension, and H, W, and D correspond to the height, width, and depth dimensions of the voxel grid, respectively.

[0071] Specifically, this step corresponds to Figure 4 The data voxelization layer uses 3D-CoordConv to inject coordinates, adding X-coordinate watermarks, Y-coordinate watermarks, and Z-coordinate watermarks to the voxel data, so that each voxel unit carries absolute spatial position information.

[0072] Specifically, the complete architecture of the 3D-CoordConv coordinate injection module includes, in sequence, a 3D coordinate mesh generator, an X-coordinate mesh normalization unit, a Y-coordinate mesh normalization unit, a Z-coordinate mesh normalization unit, a channel stitching unit, and a feature stitching and output unit.

[0073] Specifically, the module input is the single-channel voxel feature corresponding to the discrete environment perception data, with a dimension of [missing information]. After coordinate injection and channel concatenation, the output is a multi-channel voxel feature that incorporates spatial coordinate information, with dimensions of [dimension not specified]. .

[0074] Specifically, the six channels each perform different information functions, including one data channel, one mask channel, and three coordinate channels, which together with the time dimension form a complete multi-channel input structure.

[0075] Specifically, the data channel only fills in environmental sensing data such as temperature, humidity, carbon dioxide concentration, and oxygen concentration at the voxel positions corresponding to the sensors, while filling in 0 at the other voxel positions without sensors.

[0076] Specifically, the mask channel is used to identify the sensor distribution. Voxel positions with sensors are filled with 1, and voxel positions without sensors are filled with 0, which helps the model identify valid data points.

[0077] Specifically, the three coordinate channels represent the normalized X, Y, and Z axis coordinates of each voxel in 3D space, enabling the model to learn the spatial heterogeneity of different physical locations in the greenhouse and distinguish the influence weight of different sites on product growth.

[0078] See Figure 5 As shown, Figure 5 This diagram illustrates the architecture of a 3D-CoordConv coordinate injection module provided in Embodiment 1 of this application. The diagram shows the internal processing architecture of the core module 3D-CoordConv (3D CoordinateConvolution), which corresponds to the core step of constructing multi-channel 3D voxel tensors in this application. It is used to inject absolute spatial position information into discrete environment perception data, enabling the model to perceive spatial heterogeneity.

[0079] The module is based on a 3D coordinate mesh generator, and the input includes two paths: one is discrete environment-aware data with dimensions of [missing information]. One path is the single-channel 3D voxel environment data; the other path is the 3D coordinate grid output by the coordinate generator, which is processed by X-coordinate grid normalization, Y-coordinate grid normalization, and Z-coordinate grid normalization respectively to generate normalized coordinate feature maps in three dimensions.

[0080] Discrete environment perception data and three-way normalized coordinate features are input into a channel stitching unit for stitching and fusion along the channel dimension. The stitched features are then input into a feature stitching and output unit, ultimately outputting multi-channel voxel features that fuse spatial location information, with a dimension of [missing information]. .

[0081] In the picture , , These represent the height, width, and depth spatial dimensions of the voxel mesh, respectively, serving as spatial size parameters; they are defined at different levels from the hidden state symbols, distinguished by context and subscript.

[0082] S205: The input tensor is processed by a neural network to perform three-dimensional spatial diffusion and time accumulation to generate a dynamic environmental feature field based on the spatiotemporal network coordinate system.

[0083] Specifically, this step corresponds to Figure 4The core processing layer uses a 3D-ConvGRU encoder as the core computing unit and 3D-CoordConv coordinate injection as the prerequisite. It can realize information diffusion in three-dimensional space and record the environmental accumulation effect in the time dimension, thus solving the defects of traditional methods such as sparse sampling points, imprecise models and uneven three-dimensional environment.

[0084] Specifically, the 3D-ConvGRU unit contains three types of gate control structures: reset gate, update gate, and candidate hidden state gate. The input feature tensor at time t-1... Dimensions The hidden state in the previous moment Dimensions ,in This represents the hidden layer channel dimension.

[0085] Specifically, the internal convolution operator of the 3D-ConvGRU unit adopts a three-dimensional convolution kernel, and the hidden layer dimension is set to the corresponding number of channels; the three-dimensional convolution diffusion is achieved through a stereo scanning window. When the window slides over the sensor point with environmental data, the environmental value of that point is multiplied by the convolution weight and distributed to the neighboring voxels around the window.

[0086] Specifically, the operation employs a loop mechanism to achieve temporal folding, inputting temporal data frame by frame and updating the hidden state. Simultaneously, spatial diffusion is achieved through 3D convolution, ultimately outputting a 4D environmental feature field. .

[0087] Specifically, after a three-dimensional convolution operation, the data from a single sensor point is transformed into a tiny dataset that spreads evenly outward from that point, achieving spatial continuity filling of discrete data and avoiding the problem of smoothing local environmental features by a single interpolation algorithm.

[0088] Specifically, 3D-CoordConv sequentially inputs the Time dimension data of the tensor, and through the hidden state channel of the gated loop mechanism, dynamically and automatically accumulates and records the historical environmental evolution trend of each voxel position, forming a time-accumulated temperature type temporal feature, reflecting the cumulative impact of the environment on product quality.

[0089] Specifically, from the perspective of the overall input and output definition of the model, the model input is the feature vector of spatial points, and the model output is three-dimensional spatial parameters; as long as the data of any one point in the space is known, the corresponding data of all other points in the same space can be deduced.

[0090] Specifically, the model is essentially a function that describes the relationship between spatial points and parameters. Its core function is to estimate and predict spatial parameters. The feature vectors obtained after the scheme completes point-to-point spatial alignment can be directly used to build this dedicated model.

[0091] Specifically, after the calculation is completed, a four-dimensional environmental feature volume based on a three-dimensional voxel grid is output, which is a dynamic environmental feature field containing spatiotemporal information.

[0092] Specifically, this method has two core innovative designs: First, it creatively introduces a three-dimensional voxel network to integrate discretely distributed environmental sensing instruments and product points into a three-dimensional coordinate system, solving the problem of bias in environmental perception data in the vertical direction; Second, it creatively adopts a spatial alignment method that combines 3D-CoordConv and 3D-ConvGRU as the core. By adding three absolute coordinates (X, Y, and Z) to the original data through 3D-CoordConv, the 3D-ConvGRU model can automatically learn the weights of the heterogeneous influence of different physical locations in the greenhouse on product growth.

[0093] See Figure 6 As shown, Figure 6 The diagram shows the internal structure of a 3D-ConvGRU provided in Embodiment 1 of this application. It shows the core processor internal structure of the core computing unit 3D-ConvGRU (3D Convolutional Gated Recurrent Unit) of this application, which corresponds to the core link of the spatiotemporal feature encoding of the neural network in this application. It can simultaneously realize feature diffusion in the spatial dimension and information accumulation in the temporal dimension.

[0094] The unit contains two inputs: one is the hidden state from the previous time step. , dimension One path carries accumulated information about the historical environment; the other path contains the input features of the current moment. , dimension That is, the current frame environment voxel features after coordinate injection.

[0095] The two inputs are processed sequentially through three types of gating structures: First, a reset gate controls the degree of retention of historical information, filtering out redundant historical environmental information; second, an update gate controls the weighting of new information at the current moment, determining the magnitude of state updates; finally, a candidate hidden state gate generates the candidate state for the current moment. All three types of gating are implemented using three-dimensional convolution operators, completing spatial feature diffusion while processing temporal information.

[0096] After the three-gated structure performs the calculations, the unit outputs the hidden state at the current moment. It contains both the environmental spatial characteristics at the current moment and the accumulated environmental evolution information over time, and can be used to construct a dynamic four-dimensional environmental feature field.

[0097] S206: Obtain the dynamic network coordinates of the product quality sampling monitoring points.

[0098] Specifically, a unique ID is assigned to each plant or sample at each product quality sampling and monitoring point. By combining visual inspection with depth perception, the spatial location of the sample is continuously tracked, enabling dynamic monitoring throughout the entire growth period.

[0099] Specifically, the product quality data corresponding to the sampling points includes quality indicators such as product height and thickness, which are bound one-to-one with the product ID, serving as quality data support for the subsequent growth model construction.

[0100] Specifically, product quality data does not require separate spatial alignment, as it is directly linked to product sampling points; the model's basic function is only to transform and output environmental perception data, and with functional expansion, it can also achieve the correlation and alignment of quality data.

[0101] Specifically, the FairMOT target detection algorithm is used to output the pixel center position and detection box of the corresponding ID product in the image in real time. The height data of the product's location is directly collected by the depth camera. After perspective transformation, the image pixel coordinates are restored to the three-dimensional physical coordinates inside the greenhouse. Finally, it is mapped to the dynamic network coordinates corresponding to the voxel grid. The coordinates contain both planar position and height information.

[0102] S207: Using an interpolation sampling method, with the dynamic network coordinates as the sampling position, extract the environmental feature vector under the corresponding spatiotemporal coordinates from the dynamic environmental feature field, and use it as the environmental perception data of the product quality sampling monitoring point.

[0103] Specifically, this step corresponds to Figure 4 The feature alignment and sampling layer in the middle adopts the 3D Grid Sample trilinear interpolation sampling method, and extracts the target point features from the four-dimensional environmental feature field with the star-shaped sampling logic. Finally, it obtains the real alignment environmental features experienced by the product at a specific plane position and a specific height, and outputs the alignment quality-environment feature pair.

[0104] Specifically, before sampling, the physical space coordinates of the product need to be mapped to the coordinate system of the feature body, and a normalization operation needs to be performed to transform the coordinates to the standard sampling interval [-1,1]. The normalization calculation formula is as follows:

[0105] Specifically, in the formula The width coordinates of the product location in the original coordinate system of the feature body; The height coordinates of the product location in the original coordinate system of the feature body; The depth coordinates of the product location in the original coordinate system of the feature body; These are the sampled coordinates after normalization in the width direction to the interval [-1, 1]. These are the sampled coordinates after the height direction has been normalized to the interval [-1, 1]. These are the sampled coordinates after the depth direction has been normalized to the interval [-1, 1]. Define the dimensions of the voxel grid for the glass greenhouse in the width direction; Define the dimensions of the glass greenhouse voxel grid in the height direction; Define the dimensions in the depth direction of the glass greenhouse voxel grid.

[0106] Specifically, when the target sampling point falls inside the feature volume, it is usually located inside the cubic cell surrounded by 8 adjacent voxel nodes. At this time, the feature vector of the point is calculated by trilinear interpolation. The calculation process is divided into two steps.

[0107] Specifically, the first step is to identify the feature values ​​corresponding to the 8 integer coordinate nodes closest to the sampling point; the second step is to perform weighted aggregation, which determines the weights based on the relative distances between the sampling point and these 8 nodes in the X, Y, and Z axes. The closer the distance, the higher the weight. The three consecutive linear interpolation calculations are performed in sequence in the three axes to finally obtain the environmental feature vector of the sampling point.

[0108] Specifically, after feature extraction is completed, the environmental perception data of the corresponding points can be matched with the product quality data one by one, providing accurate training data for the product growth model. The essence of this method is to establish a mapping relationship between spatial points and environmental parameters. Only by knowing the perception data of some points in the space, the corresponding environmental data of all other points in the same space can be deduced.

[0109] See Figure 7 As shown, Figure 7 This diagram illustrates a small-scale spatial alignment technique for agricultural environmental sensing data provided in Embodiment 1 of this application. The diagram fully presents the entire technical route of the spatial alignment method of this application, which adopts a logical architecture of "dual-branch parallel processing - end feature alignment - closed-loop verification optimization", and fully covers all execution steps of the method of this application.

[0110] The process begins with the multi-sensor joint calibration stage, which completes the deployment and spatial positioning calibration of all environmental sensing instruments; then it enters the stage of constructing a spatiotemporal network coordinate system, which divides the overall space of the greenhouse into a unified three-dimensional voxel grid coordinate system, thereby splitting into two parallel processing branches.

[0111] The left side represents the environmental data processing branch: First, the grid coordinates of the environmental parameter sensing instrument are determined, and the physical location of the instrument is mapped to the voxel network coordinates; a 6-channel 3D (3-Dimensional) voxel tensor map is established based on the voxel grid. The tensor contains multiple dimensions such as Batch, Time, Channels, W (Width), H (Height), and D (Depth), carrying environmental data, mask labels, and spatial coordinate information; then, 3D convolutional diffusion is completed through 3D-ConvGRU, and the spatiotemporal features are jointly encoded by combining the time dimension folding mechanism, which corresponds to the core steps of input tensor construction and neural network spatiotemporal processing in this application.

[0112] The right side is the product data processing branch: First, the product grid coordinates are determined, and the image position and quality information of the crop are collected in real time through FairMOT (Fair MultipleObject Tracking) algorithm to complete the association between the product grid coordinates and quality control data, which corresponds to the steps of obtaining product dynamic coordinates and associating quality data in this application.

[0113] The processing results from the two branches are merged into the 3D Grid Sample alignment stage to achieve accurate extraction of environmental features at product locations. Then, the process enters the ground truth verification stage. If the verification fails, the process flows back to the multi-sensor joint calibration stage, and the instrument layout or model parameters are adjusted before the entire process is re-executed. If the verification passes, the final result is output, and the setting parameters in the model are obtained, corresponding to the complete process of feature alignment and closed-loop verification in this application.

[0114] In an optional implementation, see Figure 8 As shown, Figure 8 The flowchart illustrates a method for deploying environmental parameter sensing instruments according to Embodiment 1 of this application, wherein the deployment of multiple environmental parameter sensing instruments includes steps S801-S802: S801: Using the center point of the facility agriculture environmental control equipment as a reference, an environmental parameter sensing instrument is installed every 4 cubic meters.

[0115] Specifically, the center point of the glass greenhouse is first located, and the first environmental parameter sensing instrument is deployed at a height of 0.5 meters above the ground, corresponding to a 4-cubic-meter central area that is 2 meters long, 2 meters wide, and 1 meter high.

[0116] Specifically, starting from the central area, the system expands outward in both horizontal and vertical directions. Horizontally, a point is placed every 2 meters, and vertically, points are placed at heights of 0.5m, 1.5m, and 2.5m. An environmental parameter sensing instrument is placed every 4 cubic meters of space to ensure uniform sensing density throughout the entire space.

[0117] S802: Record the three-dimensional relative spatial position of each environmental parameter sensing instrument relative to the center point.

[0118] Specifically, using the center point of the greenhouse as the origin of the coordinate system, the relative position values ​​of each instrument in the X, Y, and Z axes were recorded. The accuracy of all position data recording reached the centimeter level, providing a precise positional basis for subsequent coordinate mapping and model calculation.

[0119] In an optional implementation, see Figure 9 As shown, Figure 9 The flowchart of a network coordinate mapping method provided in Embodiment 1 of this application is shown. The step of establishing a unified spatiotemporal network coordinate system and mapping the relative spatial positions and the positions of product quality sampling monitoring points to network coordinates includes steps S901-S903: S901: Divide the space of the facility agriculture environmental control equipment into a three-dimensional voxel grid.

[0120] Specifically, according to the preset voxel granularity, the entire interior space of the glass greenhouse is divided into several regular cubic voxel units, forming a three-dimensional voxel grid covering the entire space, which serves as a unified spatial reference system for all points.

[0121] S902: Maps the relative spatial position of each environmental parameter sensing instrument to the corresponding voxel network coordinates.

[0122] Specifically, the environmental parameter sensing instrument is a fixed deployment device whose physical location remains unchanged. It can be directly matched to the network coordinates of the corresponding voxel unit based on the three-dimensional relative spatial position of the instrument. The instrument will be evenly distributed at different height layers to cover the environmental differences in the vertical direction.

[0123] S903: After the location of the product quality sampling monitoring point is restored to three-dimensional physical coordinates through image recognition and depth measurement, it is then mapped to the network coordinates in the voxel grid.

[0124] Specifically, the FairMOT object detection algorithm is used to identify product samples in the image, outputting a unique sample ID, image pixel center coordinates, and detection box; depth information is collected using a depth camera to obtain the sample's height data.

[0125] Specifically, by using the principle of perspective transformation, the two-dimensional pixel coordinates in the image are combined with height data to restore the three-dimensional physical coordinates inside the greenhouse. Then, these physical coordinates are mapped onto the voxel mesh to obtain the corresponding dynamic network coordinates.

[0126] In an optional implementation, constructing the multichannel spatial input tensor includes: Construct a three-dimensional voxel tensor containing a data channel, a mask channel, a coordinate channel, and a time channel.

[0127] Specifically, this tensor is constructed based on the coordinate injection mechanism of 3D-CoordConv. Through multi-channel fusion, environmental data, point markers, spatial coordinates and time series information are integrated into a unified input format, enabling the model to learn data features and spatial heterogeneity simultaneously.

[0128] The data channel is used to fill environmental perception data at the network coordinate position corresponding to the instrument; the mask channel is used to identify the network coordinate position where the instrument exists; the coordinate channel is used to represent the normalized spatial coordinates of each voxel; and the time channel is used to represent the sampling sequence in the time dimension.

[0129] Specifically, the data channel carries various agricultural environmental sensing data such as temperature, humidity, carbon dioxide concentration, and oxygen concentration. Only the voxel positions where sensing instruments are deployed are filled with the corresponding measured values, while the remaining blank voxel positions are filled with the value 0.

[0130] Specifically, the mask channel is a binary channel, with voxel positions where sensing instruments are deployed filled with a value of 1, and voxel positions where no instruments are deployed filled with a value of 0, used to mark the spatial distribution of valid data.

[0131] Specifically, the coordinate channel contains three channels, which correspond to the normalized coordinates of each voxel on the three spatial axes of X, Y and Z. Spatial position information is injected through coordinate watermarking, enabling the model to perceive the absolute spatial position of different voxels and learn the environmental heterogeneity weights at different positions.

[0132] Specifically, the time channel corresponds to the standardized continuous time sampling sequence, giving the tensor temporal properties and supporting the time accumulation operation of the subsequent gated loop unit.

[0133] In an optional implementation, see Figure 10 As shown, Figure 10 The flowchart of a neural network processing method provided in Embodiment 1 of this application is shown, wherein the three-dimensional spatial diffusion and time accumulation processing using a neural network includes steps S1001~S1003: S1001: Employs a neural network with 3D convolution capabilities and a gated recurrent mechanism.

[0134] Specifically, the neural network uses 3D-CoordConv to inject coordinate information and construct input tensors, and uses a 3D-ConvGRU encoder as the core computing unit. It also has the ability to perform three-dimensional spatial convolution operations and temporal gating loop operations, and can simultaneously handle two types of tasks: spatial diffusion and temporal accumulation.

[0135] S1002: Discrete environmental perception data is diffused outwards in spatial dimension through three-dimensional convolution operations.

[0136] Specifically, the 3D-ConvGRU uses a three-dimensional convolution kernel and sets a stereo scanning window. During the operation, the window slides in a three-dimensional voxel grid. When the window passes a sensor voxel with measured data, the environmental value at that point is multiplied by the convolution weights, and the value is distributed to the surrounding voxels covered by the window.

[0137] Specifically, after convolution, the discrete perceptual data at a single point is transformed into continuously distributed data that spreads uniformly into three-dimensional space, thus achieving spatial filling of sparse sampling points and solving the problems of low efficiency of traditional convolution and smoothing of local features by single interpolation when the sampling points are sparse.

[0138] S1003: Accumulates historical environmental information of each network coordinate position in the time dimension through a gated loop mechanism, and outputs a dynamic environmental feature field containing spatiotemporal information.

[0139] Specifically, the model is input into the temporal dimension data of the tensor frame by frame in chronological order. The update of the hidden state is controlled by gating structures such as the reset gate, update gate, and candidate hidden state gate, so as to realize the folding of the temporal dimension, record the historical environmental change process of each voxel position, automatically accumulate the temporal effect of the environment, and form temporal features such as time-accumulated temperature.

[0140] Specifically, after the calculation is completed, a four-dimensional dynamic environmental feature field is output. The feature field uses a three-dimensional voxel grid as a spatial reference and also contains dynamic change information in the time dimension, which completely restores the spatiotemporal distribution law of the environment inside the greenhouse.

[0141] In an optional implementation, see Figure 11 As shown, Figure 11 The diagram illustrates an environmental feature vector extraction method provided in Embodiment 1 of this application, wherein the extraction of environmental feature vectors using an interpolation sampling method includes steps S1101~S1102: S1101: The trilinear interpolation method is used.

[0142] Specifically, the 3D Grid Sample sampling operator is adapted to the continuous position sampling requirements of the three-dimensional voxel feature field. Compared with nearest neighbor sampling and other methods, it has higher sampling accuracy and can accurately restore the environmental features of any spatial point.

[0143] S1102: When the dynamic network coordinates fall inside the cell of the dynamic environment feature field, the environmental feature vector of the point is calculated by weighted aggregation based on the distance between the coordinates and the adjacent voxel nodes in three axes.

[0144] Specifically, first, locate the cubic cell where the sampling point is located, identify the 8 adjacent voxel nodes that make up the cell, and obtain the feature values ​​corresponding to the 8 nodes.

[0145] Specifically, the relative distances between the sampling point and the eight nodes in the X, Y, and Z axes are calculated respectively, and the weights are determined based on the inverse ratio of the distances, with the nodes that are closer to each other having higher weights.

[0146] Specifically, three linear interpolation operations are performed sequentially along the X-axis, Y-axis, and Z-axis, and finally weighted aggregation is performed to obtain the environmental feature vector of the sampling point location, achieving accurate feature extraction at the sub-voxel level and outputting aligned quality-environment feature pairs.

[0147] In an optional implementation, see Figure 12 As shown, Figure 12 A flowchart of a data verification method provided in Embodiment 1 of this application is shown, wherein the method further includes steps S1201~S1202: S1201: The environmental perception data of the extracted product quality sampling monitoring point is correlated with the product quality data of the monitoring point to construct or optimize the product growth model.

[0148] Specifically, a unique ID is assigned to each product sample, and the product's quality inspection data, data collection time, and environmental perception data extracted under the corresponding spatiotemporal coordinates are bound and associated one by one.

[0149] Specifically, the aligned spatiotemporal matching data can be directly used to build or optimize product growth models, solving the problem of insufficient modeling accuracy caused by spatial bias in traditional data and improving the overall efficiency of environmental control.

[0150] S1202: After extraction, multiple points are randomly selected to collect real environmental data and compare them with the extraction results for verification until the extraction results meet the preset inspection standards.

[0151] Specifically, during the verification process, 10 points were randomly selected in the glass greenhouse, and real environmental perception data and product quality data were collected at these points at any time. The measured data were then compared with the calculated data at the corresponding locations and times output by the model.

[0152] Specifically, if the comparison results do not meet the preset model verification standards, the verification is deemed unsuccessful, and the process is returned to the multi-sensor joint calibration stage. After adjusting the instrument deployment scheme or model parameters, the entire process calculation is re-executed.

[0153] Specifically, if the comparison results meet the inspection criteria, the inspection is deemed passed, the setting parameters of the current model are determined, and the spatial alignment parameters of the greenhouse of this model and in this erection direction are constructed.

[0154] This method addresses the challenges of sparse sampling points, imprecise spatial alignment models, and uneven three-dimensional environments in traditional farmland small-scale environmental parameter sensing data and product quality data. It achieves high-precision spatial alignment of small-scale agricultural environmental sensing data, demonstrating significant technical advantages and application value.

[0155] Specifically, this method first uses a three-dimensional voxel network to establish a unified spatiotemporal network coordinate system, then uses 3D-CoordConv and 3D-ConvGRU to realize the spatiotemporal diffusion and accumulation of environmental parameter perception data, then uses FairMOT to realize dynamic monitoring of product quality data with ID, and finally uses 3D Grid Sample to complete the extraction of real alignment environmental features experienced by the product at a specific plane position and a specific height.

[0156] Specifically, compared with existing technologies, this method solves the problems of sparse sampling points and uneven three-dimensional environment, significantly improves the accuracy of spatial alignment model, realizes spatial alignment of small-scale agricultural environment perception data, provides reliable data guarantee for product growth model, and can effectively support intelligent environmental control of agricultural production.

[0157] Example 2 See Figure 13 As shown, Figure 13 This illustration shows a structural schematic diagram of an environmental perception data determination device according to Embodiment 2 of this application, wherein the device includes: Model selection module 1301 is used to select the model of environmental control equipment for facility agriculture; The instrument deployment module 1302 is used to deploy multiple environmental parameter sensing instruments according to the spatial volume parameters corresponding to the model, and record the relative spatial position of each environmental parameter sensing instrument. The coordinate mapping module 1303 is used to establish a unified spatiotemporal network coordinate system based on the spatial volume parameters, and to map the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point to network coordinates respectively. Tensor construction module 1304 is used to construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data. The feature generation module 1305 is used to perform three-dimensional spatial diffusion and time accumulation processing on the input tensor using a neural network to generate a dynamic environmental feature field based on the spatiotemporal network coordinate system. The coordinate acquisition module 1306 is used to acquire the dynamic network coordinates of the product quality sampling monitoring points; The data extraction module 1307 is used to extract environmental feature vectors in the corresponding spatiotemporal coordinates from the dynamic environmental feature field using an interpolation sampling method, with the dynamic network coordinates as the sampling position, as environmental perception data of the product quality sampling monitoring points.

[0158] In an optional implementation, the instrument deployment module is specifically used for: Using the center point of the facility agriculture environmental control equipment as a reference, an environmental parameter sensing instrument is installed every 4 cubic meters. Record the three-dimensional relative spatial position of each environmental parameter sensing instrument relative to the center point.

[0159] Optionally, the coordinate mapping module is specifically used for: The space of the facility agriculture environmental control equipment is divided into a three-dimensional voxel grid. The relative spatial positions of the environmental parameter sensing instruments are mapped to the corresponding voxel network coordinates; After the location of the product quality sampling monitoring points is restored to three-dimensional physical coordinates through image recognition and depth measurement, it is then mapped to network coordinates in the voxel grid.

[0160] Optionally, the tensor construction module is specifically used for: Construct a three-dimensional voxel tensor containing a data channel, a mask channel, a coordinate channel, and a time channel; The data channel is used to fill in environmental perception data at the network coordinates corresponding to the instrument. The mask channel is used to identify the network coordinates of the instrument. The coordinate channels are used to represent the normalized spatial coordinates of each voxel. The time channel is used to represent the sampling sequence in the time dimension.

[0161] Optionally, the feature generation module is specifically used for: A neural network with 3D convolution capability and gated recurrent mechanism is used; The discrete environmental perception data is diffused outwards in a spatial dimension through three-dimensional convolution operations. It also accumulates historical environmental information of each network coordinate position in the time dimension through a gating loop mechanism, and outputs a dynamic environmental feature field containing spatiotemporal information.

[0162] Optionally, the data extraction module is specifically used for: The trilinear interpolation method is used; When the dynamic network coordinates fall inside the cell of the dynamic environment feature field, the environmental feature vector of that point is calculated by weighted aggregation based on the distance between the coordinates and the adjacent voxel nodes in three axes.

[0163] Optionally, the device further includes a verification and optimization module for: The environmental perception data of the product quality sampling monitoring points obtained by extraction is associated with the product quality data of the monitoring points to construct or optimize the product growth model. After extraction, multiple points are randomly selected to collect real environmental data and compare them with the extraction results for verification until the extraction results meet the preset inspection standards.

[0164] Example 3 Based on the same application concept, see [link / reference] Figure 14 As shown, Figure 14 This illustration shows a structural schematic diagram of a computer device provided in Embodiment 3 of this application, wherein, as shown... Figure 14 As shown, the computer device 1400 provided in Embodiment 3 of this application includes: The computer device 1400 includes a processor 1401, a memory 1402, and a bus 1403. The memory 1402 stores machine-readable instructions executable by the processor 1401. When the computer device 1400 is running, the processor 1401 communicates with the memory 1402 via the bus 1403. The machine-readable instructions are executed by the processor 1401 to perform the steps of the environmental perception data determination method shown in Embodiment 1 above.

[0165] Example 4 Based on the same concept, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the environmental perception data determination method described in any of the above embodiments.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0167] The computer program product for determining environmental perception data provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0168] The environmental perception data determination device provided in this application embodiment can be specific hardware on a device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0169] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. 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 the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0172] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0174] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for determining environmental perception data, characterized in that, include: Select the model of environmental control equipment for facility agriculture; Based on the spatial volume parameters corresponding to the model, multiple environmental parameter sensing instruments are deployed, and the relative spatial positions of each environmental parameter sensing instrument are recorded. A unified spatiotemporal network coordinate system is established based on the spatial volume parameters, and the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point are respectively mapped to network coordinates. Construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data; The input tensor is processed by a neural network to perform three-dimensional spatial diffusion and temporal accumulation, thereby generating a dynamic environmental feature field based on the spatiotemporal network coordinate system. Obtain the dynamic network coordinates of the product quality sampling and monitoring points; Using an interpolation sampling method, with the dynamic network coordinates as the sampling position, environmental feature vectors under the corresponding spatiotemporal coordinates are extracted from the dynamic environmental feature field, which are used as environmental perception data for the product quality sampling monitoring points.

2. The method according to claim 1, characterized in that, The deployment of multiple environmental parameter sensing instruments includes: Using the center point of the facility agriculture environmental control equipment as a reference, an environmental parameter sensing instrument is installed every 4 cubic meters. Record the three-dimensional relative spatial position of each environmental parameter sensing instrument relative to the center point.

3. The method according to claim 1, characterized in that, The establishment of a unified spatiotemporal network coordinate system, and the mapping of the relative spatial positions and the positions of product quality sampling monitoring points to network coordinates, includes: The space of the facility agriculture environmental control equipment is divided into a three-dimensional voxel grid. The relative spatial positions of the environmental parameter sensing instruments are mapped to the corresponding voxel network coordinates; After the location of the product quality sampling monitoring points is restored to three-dimensional physical coordinates through image recognition and depth measurement, it is then mapped to network coordinates in the voxel grid.

4. The method according to claim 1, characterized in that, The construction of the multi-channel spatial input tensor includes: Construct a three-dimensional voxel tensor containing a data channel, a mask channel, a coordinate channel, and a time channel; The data channel is used to fill in environmental perception data at the network coordinates corresponding to the instrument. The mask channel is used to identify the network coordinates of the instrument. The coordinate channels are used to represent the normalized spatial coordinates of each voxel. The time channel is used to represent the sampling sequence in the time dimension.

5. The method according to claim 1, characterized in that, The process of using neural networks for three-dimensional spatial diffusion and temporal accumulation includes: A neural network with 3D convolution capability and gated recurrent mechanism is used; The discrete environmental perception data is diffused outwards in a spatial dimension through three-dimensional convolution operations. It also accumulates historical environmental information of each network coordinate position in the time dimension through a gating loop mechanism, and outputs a dynamic environmental feature field containing spatiotemporal information.

6. The method according to claim 1, characterized in that, The method of extracting environmental feature vectors using interpolation sampling includes: The trilinear interpolation method is used; When the dynamic network coordinates fall inside the cell of the dynamic environment feature field, the environmental feature vector of that point is calculated by weighted aggregation based on the distance between the coordinates and the adjacent voxel nodes in three axes.

7. The method according to claim 1, characterized in that, The method further includes: The environmental perception data of the product quality sampling monitoring points obtained by extraction is associated with the product quality data of the monitoring points to construct or optimize the product growth model. After extraction, multiple points are randomly selected to collect real environmental data and compare them with the extraction results for verification until the extraction results meet the preset inspection standards.

8. An environmental perception data determination device, characterized in that, The device includes: The model selection module is used to select the model of environmental control equipment for facility agriculture; The instrument deployment module is used to deploy multiple environmental parameter sensing instruments according to the spatial volume parameters corresponding to the model, and record the relative spatial position of each environmental parameter sensing instrument. The coordinate mapping module is used to establish a unified spatiotemporal network coordinate system based on the spatial volume parameters, and to map the relative spatial position of the environmental parameter sensing instrument and the position of the product quality sampling monitoring point into network coordinates respectively. The tensor construction module is used to construct a multi-channel spatial input tensor containing the network coordinates of the sensing instruments for each environmental parameter and their environmental sensing data. The feature generation module is used to perform three-dimensional spatial diffusion and temporal accumulation processing on the input tensor using a neural network to generate a dynamic environmental feature field based on the spatiotemporal network coordinate system. The coordinate acquisition module is used to acquire the dynamic network coordinates of the product quality sampling monitoring points; The data extraction module is used to extract environmental feature vectors in the corresponding spatiotemporal coordinates from the dynamic environmental feature field using an interpolation sampling method, with the dynamic network coordinates as the sampling position, as environmental perception data for the product quality sampling monitoring points.

9. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the environmental perception data determination method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the environmental perception data determination method as described in any one of claims 1 to 7.

Images