A method, system and storage medium for intelligent sensing and data processing of crop-environment multi-source information in farmland

Abstract

The application discloses a kind of farmland crop-environment multi-source information intelligent sensing and data processing method and system, belong to agricultural information technology field.The method includes: constructing multi-protocol compatible module, the transmission protocol compatible adaptation of multi-source heterogeneous data is carried out, and data communication link is established;The multi-source heterogeneous data frame is carried out data analysis preprocessing, and standardization plain code service data is generated;The standardization plain code service data is carried out data quality inspection, and the abnormal data found in inspection is supplemented with missing;After supplementing, multi-source data is matched in time and space, and multi-source data fusion analysis is carried out;Formulate general open data product format, and the data after fusion is generated as standardization data product.The method can solve the technical problems that the data transmission protocol of existing farmland monitoring system is not compatible, multi-source data is difficult to fuse, data format is not unified, data stream transmission efficiency is low and data acquisition channel is single.

Description

Technical Field

[0001] This invention belongs to the field of agricultural information technology, specifically relating to a method and system for intelligent sensing and data processing of multi-source information on farmland crops and the environment, used to realize intelligent sensing and data processing of multi-source information on farmland. Background Technology

[0002] With the rapid development of precision agriculture, real-time and comprehensive monitoring of crop growth status and environmental parameters in farmland has become a core prerequisite for achieving scientific planting and precise regulation. Currently, the field of farmland monitoring has gradually introduced various sensing methods such as satellite remote sensing, UAV aerial surveying, and ground-based IoT sensors, which can periodically or in real-time acquire multi-dimensional information such as crop growth, soil moisture, and meteorological conditions. However, existing technologies still have many shortcomings, which seriously restrict the level of intelligence and standardization in farmland monitoring.

[0003] First, in existing monitoring systems, different sensing devices use different transmission protocols. For example, ground sensors commonly use RS485 and ZigBee protocols, drones commonly use Wi-Fi and LoRa protocols, and satellite data transmission uses TCP / IP protocols. This lack of protocol uniformity makes it difficult for data from different devices to communicate efficiently, creating "data silos" and increasing the difficulty of data integration.

[0004] Secondly, the data acquisition channels are relatively limited, mainly relying on self-owned sensing equipment for collection. There is a lack of interfaces to connect with publicly available data resources on the Internet, making it difficult to supplement regional-scale meteorological and soil background data, resulting in insufficient completeness of monitoring data.

[0005] Furthermore, the multi-source data acquired by satellites, drones, and ground sensors are biased in terms of time and space, and lack an effective spatiotemporal registration mechanism, making it difficult to accurately integrate the multi-source data and restricting the ability to comprehensively reflect the true state of farmland crops and the environment.

[0006] At the same time, there is currently no unified product format standard for precision irrigation-related observation data. The data structures and encoding methods output by different systems vary greatly, which restricts cross-platform data sharing and reuse and reduces data utilization efficiency.

[0007] Furthermore, existing data stream transmission methods mostly employ a single transmission path, resulting in high transmission latency, high packet loss rate, and poor security, making it difficult to meet the real-time data transmission requirements of farmland monitoring. Regarding edge computing, existing systems mostly adopt a centralized processing architecture, requiring all raw data to be aggregated to a central server before processing. This places heavy pressure on the transmission links, leads to poor timeliness, and lacks technical means for data preprocessing and lightweight transmission at the edge.

[0008] To address the aforementioned technical challenges, there is an urgent need for a method and system for intelligent sensing and data processing of multi-source information on farmland crops and the environment that can achieve multi-protocol compatibility, multi-source data fusion, standardized data formats, efficient data stream transmission, and abundant data acquisition channels, so as to provide technical support for the development of precision agriculture. Summary of the Invention

[0009] This invention provides a method and system for intelligent sensing and data processing of multi-source information on farmland crops and the environment, in order to solve the problems mentioned in the background art.

[0010] In a first aspect, embodiments of the present invention provide a method for intelligent sensing and data processing of multi-source information on farmland crops and the environment. The method includes the following steps: Step 1, constructing a multi-protocol compatible module to adapt the transmission protocols of multi-source heterogeneous data shared from ground-based IoT sensor arrays, UAV monitoring equipment, and satellite remote sensing observation data centers, and establishing a data communication link; Step 2, performing data parsing preprocessing on the multi-source heterogeneous data frames transmitted through the data communication link to generate standardized plaintext business data; Step 3, performing data quality checks on the standardized plaintext business data and supplementing any missing abnormal data found during the checks; Step 4, using the location coordinates and acquisition time of the ground-based IoT sensors as a benchmark, performing spatiotemporal matching on the supplemented multi-source data, and realizing multi-source data fusion analysis through a multi-source data fusion algorithm; Step 5, formulating a universal and open format for precision irrigation observation data products, and generating standardized data products from the fused data.

[0011] In one possible implementation, the multi-protocol compatibility module in step one includes a multi-protocol communication module and a main controller. The main controller runs a protocol adaptation engine to automatically identify the protocol channel of the multi-protocol communication module, automatically match communication parameters and load the corresponding protocol driver. When there is no matching item in the driver library, a general driver template is loaded to enter the downgrade adaptation mode, and data is packaged and submitted through a unified data reading interface.

[0012] In one possible implementation, the data parsing preprocessing in step two includes sequentially performing protocol identification, frame structure parsing, and semantic extraction on multi-source heterogeneous data frames; wherein, the protocol identification is used to determine the protocol category of the data frame, the frame structure parsing is used to mark the field type of each byte and infer the field boundaries, and the semantic extraction is used to establish the mapping relationship between protocol fields and business concepts.

[0013] In one possible implementation, the protocol identification includes: extracting multi-dimensional feature vectors of data frames to train a multi-class random forest model to determine the protocol category, or constructing a byte-level N-gram probability model to calculate the log-likelihood value for matching known protocols; when the protocol category of a data frame cannot be determined, density clustering is performed based on the feature vectors to discover unknown protocols.

[0014] In one possible implementation, the frame structure parsing includes: modeling the data frame parsing as a sequence labeling task; extracting byte-level context features through a bidirectional long short-term memory network and modeling and labeling transition constraints through a conditional random field layer; labeling the field type of each byte in the data frame; inferring field boundaries by combining the mutual information between adjacent bytes; and distinguishing between fixed fields and data fields by analyzing the variation patterns of bytes at the same position in multiple frames of data.

[0015] In one possible implementation, the semantic extraction includes: constructing a knowledge graph between protocol fields and physical quantities and business concepts, and using graph reasoning to automatically discover semantic mapping paths for new protocol fields; migrating the semantic classifiers of known protocols to new protocols through domain adaptation; and constructing positive and negative sample pairs for data reported by the same device through different protocols, so that semantically identical fields are aligned in the embedding space.

[0016] In one possible implementation, prior to the data quality check described in step three, publicly available farmland data resources from an external data platform are obtained via an internet data acquisition interface using an API. This acquired data is then incorporated as part of the standardized plaintext business data. The data quality check in step three focuses on three dimensions: data integrity, timeliness, and format content, and identifies and eliminates abnormal data to form a standardized quality check code. This standardized quality check code includes information on data type, inspection dimensions, anomaly level, and processing result.

[0017] In one possible implementation, the data quality check in step three revolves around three dimensions: data integrity, timeliness, and format content, and investigates abnormal data to form a standardized quality check code. The standardized quality check code includes data type, inspection dimension, anomaly level, and processing result information.

[0018] In one possible implementation, the missing data completion in step three adaptively selects a completion strategy based on the length of the missing data: for a small number of missing data, interpolation is used for completion; for long-term missing data, a temporal completion algorithm based on the Transformer self-attention mechanism is used. The algorithm captures temporal dependencies from the bidirectional context through a two-stage self-attention architecture to generate initial completion values, and then fuses the completion results with the original observations in a weighted combination manner. The completion data output by the temporal completion algorithm is accompanied by a confidence label.

[0019] In one possible implementation, the spatiotemporal matching in step four includes spatial registration and temporal registration. The spatial registration converts multi-source heterogeneous data into the CGCS2000 coordinate system and uses a seven-parameter Bursa model for coordinate transformation. The temporal registration uses a bidirectional timescale exchange algorithm to estimate clock offset and propagation delay, and performs linear clock drift compensation by fitting the drift rate using least squares.

[0020] In one possible implementation, the multi-source data fusion algorithm in step four includes: using Kriging interpolation for spatial completion; using Kalman filtering, probability density matching, or optimal interpolation for numerical fusion, wherein the probability density matching method is used when the station observation data does not meet the preset density conditions, and the optimal interpolation method is used when the station observation data meets the preset density conditions; using DS evidence theory to assess the uncertainty of the fusion result; and setting a data fusion time window to cope with data delay.

[0021] In one possible implementation, the precision irrigation observation data product format described in step five covers at least one of crop parameters, soil parameters, meteorological parameters, and irrigation parameters, and supports custom extended fields.

[0022] In one possible implementation, the multi-protocol compatibility module adapts the transmission protocols of multi-source heterogeneous data and transmits it to edge computing nodes deployed in the core monitoring area of ​​farmland via short-range wireless local area network or long-range low-power wireless method.

[0023] In one possible implementation, steps two through five are completed in the edge computing node to generate the standardized data product.

[0024] In one possible implementation, the standardized data product is transmitted using a hybrid network of fifth-generation mobile communication and optical fiber, with real-time data in the standardized data product transmitted via fifth-generation mobile communication and large-capacity data in the standardized data product transmitted via optical fiber.

[0025] Secondly, embodiments of the present invention provide a system for intelligent sensing and data processing of multi-source information on farmland crops and the environment, characterized by comprising: a multi-protocol compatibility module, used to adapt and establish data communication links for multi-source heterogeneous data shared from ground-based IoT sensor arrays, UAV monitoring equipment, and satellite remote sensing observation data centers; a data parsing and preprocessing module, used to perform data parsing and preprocessing on multi-source heterogeneous data frames transmitted through the data communication links to generate standardized plaintext business data; a data quality control module, used to perform data quality checks on the standardized plaintext business data and supplement missing data found during the checks; a spatiotemporal registration and fusion module, used to perform spatiotemporal matching on the supplemented multi-source data based on the location coordinates and acquisition time of the ground-based IoT sensors, and to achieve multi-source data fusion analysis through a multi-source data fusion algorithm; and a data standardization module, used to formulate a universal and open format for precision irrigation observation data products and generate standardized data products from the fused data.

[0026] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for intelligent sensing and data processing of multi-source information on farmland crops and the environment.

[0027] The present invention provides a method for intelligent sensing and data processing of multi-source information on farmland crops and the environment, which has the following advantages compared with the prior art: First, by constructing a multi-protocol compatible module and a protocol adaptive engine, the system achieves automatic identification and intelligent adaptation of transmission protocols for various heterogeneous sensing devices, solving the problem of unified access to multi-source heterogeneous data in farmland and improving the compatibility and scalability of the sensing system.

[0028] Second, by adopting a three-stage progressive data parsing preprocessing process of protocol identification, frame structure parsing, and semantic extraction, combined with intelligent technologies such as bidirectional long short-term memory networks, conditional random fields, knowledge graph reasoning, and domain adaptation, automatic parsing and semantic understanding of unknown and private protocol data frames are achieved, improving the automation level and parsing accuracy of the data preprocessing process.

[0029] Third, by using the seven-parameter Bursa model coordinate transformation and bidirectional time-scale exchange algorithm for clock synchronization in spatiotemporal matching, combined with multi-source data fusion algorithms such as Kriging interpolation, Kalman filtering, probability density matching, optimal interpolation, and DS evidence theory, accurate registration and deep fusion of multi-source heterogeneous data in spatial and temporal dimensions are achieved, improving the spatial continuity and data reliability of farmland environmental information products.

[0030] Fourth, by using local preprocessing at edge computing nodes and a hierarchical data stream transmission architecture, the amount of transmitted data is compressed to 10% to 20% of the original amount, reducing network transmission load and transmission latency, and improving the system's real-time response capability and transmission efficiency.

[0031] Fifth, a universal and open format for precision irrigation observation data products has been developed, covering four major categories of data: crop parameters, soil parameters, meteorological parameters, and irrigation parameters, and supporting custom extensions. It adopts common formats such as GeoTIFF and JSON for output, facilitating the sharing and application of data products across different agricultural information platforms. Attached Figure Description

[0032] Figure 1 A flowchart of a method for intelligent sensing and data processing of multi-source information on farmland crops and the environment provided in an embodiment of the present invention; Figure 2 A flowchart of data parsing and preprocessing provided for embodiments of the present invention; Figure 3 This is a schematic diagram of a three-dimensional monitoring network integrating space, air, and ground, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of a multi-source intelligent sensing and data processing system for farmland crops and the environment provided in an embodiment of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] like Figure 1 As shown, a first aspect of the present invention provides a method for intelligent sensing and data processing of multi-source information on farmland crops and the environment, comprising: Step 1: Build a multi-protocol compatible module and establish a data communication link.

[0035] This embodiment first constructs a multi-protocol compatible module at the data access end to adapt the transmission protocols of multi-source heterogeneous data shared by ground IoT sensor arrays, UAV monitoring equipment and satellite remote sensing observation data centers, and establish a data communication link.

[0036] The multi-protocol compatibility module pre-configures parsing rules and conversion logic for six protocols: RS485, ZigBee, LoRa, Wi-Fi, TCP / IP, and MQTT. Ground-based IoT sensors use RS485 / 232 and ZigBee protocols for access; drone monitoring equipment uses Wi-Fi for short-range data transmission, switching to LoRa for long-range transmission when outside Wi-Fi coverage; satellite remote sensing data and publicly available cloud platform data use TCP / IP. The multi-protocol compatibility module also reserves protocol extension interfaces; if additional sensor devices using other protocols such as NB-IoT are added later, protocol adaptation can be achieved through custom configuration of parsing rules.

[0037] The multi-protocol compatibility module adopts an architecture of "main controller + multi-protocol communication module + protocol adaptive engine + plug-and-play adaptation layer". The main controller serves as the core management unit of the module, and the multi-protocol communication module integrates multiple protocol communication units. Each protocol communication unit operates independently in partitions, is centrally powered, and uniformly controlled to avoid mutual interference and support hot switching of protocols.

[0038] The protocol adaptive engine automatically identifies the protocol channel of the multi-protocol communication module, automatically matches communication parameters, and loads the corresponding protocol driver. The specific workflow is as follows: When sensors, drones, or satellite data access devices connect to the multi-protocol compatible module, the protocol adaptive engine first listens for device access events through an embedded hot-plug detection mechanism to perceive changes in device topology in real time. Then, it obtains the device type information and supported communication protocol information by reading the device descriptor, registers the device in the device management table, and assigns a logical address. The protocol adaptive engine matches the protocol information in the device descriptor with the driver library, automatically loads the corresponding protocol driver, and automatically configures communication parameters (including baud rate, frequency band, network identifier, and network address). If a fully matching protocol driver exists in the driver library, it is directly loaded and protocol adaptation is completed; if no matching item exists, a general driver template is loaded to enter a degraded adaptation mode, ensuring basic data communication capabilities are maintained even in unknown protocol scenarios. After the driver is loaded, a logical binding is established between the device and the corresponding protocol communication unit. Data is packaged through a unified data reading interface and output in a standardized format to subsequent processing stages.

[0039] Taking soil moisture sensor access as an example: After the sensor is connected via the RS485 interface, the protocol adaptation engine detects the access event and reads the device descriptor, identifying the device as a soil moisture sensor using the RS485 protocol. It then loads the corresponding RS485 driver from the driver library, configures parameters such as the communication baud rate, and establishes logical binding. The data uploaded by the sensor, after protocol adaptation, is packaged through a unified data reading interface and output in a standardized format for subsequent processing.

[0040] Taking drone data access as an example: When the drone is within close range, it transmits the collected data to a multi-protocol compatible module via Wi-Fi. When the drone flies outside the Wi-Fi coverage area, the system automatically switches to the LoRa protocol to continue transmission. After the data arrives, the protocol adaptation engine loads the corresponding driver according to the protocol type, completes protocol adaptation, and outputs the data in a standardized format for subsequent processing.

[0041] Satellite remote sensing data is transmitted directly to the multi-protocol compatible module via TCP / IP protocol. After the protocol adaptation engine identifies the protocol type, it loads the TCP / IP driver to complete the adaptation and outputs the data in a standardized format for subsequent processing.

[0042] Step 2: Perform data parsing and preprocessing on the multi-source heterogeneous data frames to generate standardized plaintext business data.

[0043] The multi-source heterogeneous data frames transmitted via the data communication link described in step one undergo data parsing preprocessing to generate standardized plaintext service data. This data parsing preprocessing is a crucial step connecting the original heterogeneous data with upper-layer service applications; its core lies in parsing transmitted data from different sources and in different formats into unified standard-format plaintext service data. For example... Figure 2 As shown, the data parsing preprocessing includes three processing stages: protocol identification, frame structure parsing, and semantic extraction, sequentially processing multi-source heterogeneous data frames.

[0044] (I) First Stage: Protocol Identification. Protocol identification is used to determine the protocol category of data frames and is the first stage of data parsing preprocessing. This embodiment uses three complementary algorithms to identify protocols from different perspectives: N-Gram is responsible for extracting message byte sequence features and coarse template matching → Random Forest completes accurate classification and identification of known protocols → DBSCAN unsupervisedly mines unknown private protocols and filters abnormal and noisy messages. The three form a closed loop of "supervised accurate identification + sequence feature matching + unsupervised anomaly / new protocol mining", covering the entire scenario of IoT protocol identification: known protocol differentiation, private unknown protocol discovery, malformed / attack / incomplete message filtering, and batch device anomaly clustering alarm.

[0045] Protocol identification random forest classification algorithm: This algorithm classifies protocols from a statistical feature perspective. It extracts multi-dimensional feature vectors from data frames (covering statistical and structural features such as frame length distribution, byte frequency, N-gram distribution, and fixed offset stability), trains a multi-class random forest model, and outputs the probability distribution of each protocol category. The protocol category with the highest probability is selected as the candidate result. When the maximum probability value exceeds a preset recognition threshold, the data frame is determined to belong to the corresponding protocol category; when the maximum probability value does not exceed the threshold, an N-gram language model matching algorithm is used for further discrimination.

[0046] The protocol identification random forest classification algorithm is a core supervised learning algorithm for accurate identification of known IoT protocols. It leverages multi-dimensional features extracted by N-Gram to achieve high-precision multi-class classification of standard protocols such as MQTT, Modbus, and CoAP. Through a multi-decision tree ensemble voting mechanism, it effectively solves problems such as misclassification due to single-template matching and feature confusion, resists minor message defects and noise interference, and boasts fast inference speed and strong generalization ability. It can output classification confidence scores, accurately distinguishing between trusted protocol messages and low-confidence unknown messages, achieving accurate traffic routing for known protocols, and providing precise screening criteria for clustering and mining unknown messages.

[0047] Protocol identification using N-gram language model matching algorithm: This algorithm matches protocols from a sequence probability perspective. For frame data with known protocols, a byte-level N-gram probability model is constructed. For unknown frames, the log-likelihood value under each protocol model is calculated, and the protocol with the highest log-likelihood value is taken as the matching result. When the maximum log-likelihood value exceeds a preset matching threshold, the data frame is determined to belong to the corresponding protocol category. If neither the random forest classification algorithm nor the N-gram language model matching algorithm can determine the protocol category, the DBSCAN density clustering unknown protocol discovery process is initiated.

[0048] As a low-level feature extraction algorithm for protocol identification, this algorithm primarily addresses the challenge of quantifying and modeling binary and custom hybrid messages in the Internet of Things (IoT). By extracting continuous byte segments from the message and statistically analyzing the byte sequence distribution characteristics, the algorithm transforms the raw message into a structured feature vector, achieving a unified representation of multi-protocol messages. It features lightweight design, low latency, and compatibility with proprietary protocol variants. It can quickly perform coarse message matching and preliminary filtering of dirty data, providing high-quality feature input for upper-layer classification and clustering models. This effectively improves the overall protocol identification efficiency and stability, making it suitable for real-time parsing scenarios in edge gateways.

[0049] DBSCAN density clustering discovery algorithm: When neither of the above two algorithms can determine the protocol category of a data frame, density clustering is performed based on the feature vector of the data frame. The clustering parameters include the neighborhood radius ε and the minimum number of samples MinPts. Data frames with density connections in the feature space are grouped into the same cluster. Data frames within the same cluster are likely to belong to the same unknown protocol. After DBSCAN clusters unknown protocol frames into the same cluster, data frames within the cluster can enter the frame structure parsing stage without protocol type labels. Field decomposition is completed through BiLSTM-CRF, mutual information, and consistency analysis. After the unknown protocol clusters accumulate to a certain scale and the protocol type is confirmed by manual review, the protocol type and its feature information can be added to the protocol feature library and driver library to achieve continuous expansion of the protocol library.

[0050] DBSCAN density clustering algorithm: As a core algorithm for unsupervised data mining and anomaly detection, it is specifically designed to handle low-confidence, unknown private protocol messages that supervised algorithms cannot identify. It requires no preset number of classifications, automatically clustering based on feature density connectivity, adapting to arbitrary protocol feature distributions, accurately mining novel private protocol clusters, and naturally isolating noisy messages such as those from device failures, network incompleteness, and network attacks. It can distinguish between single-point, occasional anomalies and batch device cluster failures, supports tiered alarms, and can iteratively update the protocol feature library, achieving a self-optimizing closed loop for protocol recognition capabilities.

[0051] (II) Second Stage: Frame Structure Parsing. Frame structure parsing is used to decompose the data frame into semantic fields and determine field boundaries; it is the second stage of data parsing preprocessing. This stage employs three algorithms to collaboratively complete frame structure parsing from different perspectives: The BiLSTM-CRF sequence labeling algorithm for protocol parsing is the core algorithm in the frame structure parsing stage. It models frame structure parsing as a sequence labeling task, encodes byte sequences from both forward and reverse directions using a bidirectional long short-term memory network, and extracts byte-level context features. It models labeling transition constraints through a conditional random field layer and uses the Viterbi decoding algorithm to solve for the globally optimal labeling sequence, labeling the field type (such as frame header, length, command, data, checksum, etc.) of each byte in the data frame.

[0052] In one embodiment, the BiLSTM-CRF model is first trained in a supervised manner on labeled datasets of known protocols (such as RS485, ZigBee, Wi-Fi, and other commonly used agricultural IoT protocols) to learn the basic patterns of field type labeling; then, it is fine-tuned or directly inferred on new protocol data frames. In another embodiment, for new protocols with no labeled data, pseudo-labeled data can be automatically generated through mutual information field boundary inference and byte consistency analysis, and then the BiLSTM-CRF model is trained in a semi-supervised manner.

[0053] The protocol parsing mutual information field boundary inference algorithm provides field boundary clues for BiLSTM-CRF from the perspective of statistical correlation. Adjacent bytes within the same field have statistical correlation, and the statistical characteristics change abruptly at the field boundary. The local minimum of mutual information is the candidate field boundary. When the mutual information value of adjacent bytes is lower than a preset boundary threshold, that position is determined to be a field boundary.

[0054] Protocol parsing byte consistency analysis algorithm: Provides prior information on field types for BiLSTM-CRF from the perspective of data variation. Collects multiple frames of data from the same protocol and calculates the byte variation coefficient at each offset position: invariant bytes are fixed fields (such as frame headers and length fields), and variable bytes are data fields. When the variation coefficient is close to zero, it is determined to be a fixed field; when the variation coefficient is higher than the upper threshold, it is determined to be a data field; and when it is between the upper and lower thresholds, it is determined to be a counter or sequence number field.

[0055] In one embodiment, the three algorithms work collaboratively using a feature-level fusion approach: the calculation results of mutual information field boundary inference and byte consistency analysis are used as auxiliary feature vectors, concatenated with the byte embedding vectors, and then input into a BiLSTM network, where BiLSTM-CRF completes the core sequence labeling task. Mutual information indicates the field boundary position from a statistical correlation perspective, while byte consistency analysis distinguishes between fixed fields and data fields from a data variation perspective. The fusion of these three algorithms significantly improves the accuracy of frame structure parsing and its adaptability to different protocols.

[0056] In another embodiment, the three algorithms described above can also work collaboratively using a decision-level fusion approach: BiLSTM-CRF, mutual information inference, and byte consistency analysis each independently output field labeling results, and then the outputs of the three are combined through a voting mechanism or a weighted fusion strategy to determine the final field type and boundaries. The advantage of decision-level fusion is that each algorithm runs independently, facilitating modular deployment and fault isolation.

[0057] (III) Third Stage: Semantic Extraction. Semantic extraction is used to establish the mapping relationship between protocol fields and business concepts, and is the third stage of data parsing preprocessing. This stage employs three complementary algorithms to cover the main processing scenarios for different situations of new protocol semantic mapping: Semantic extraction knowledge graph reasoning algorithm: Suitable for scenarios where there are reasonable association paths between new protocol fields and existing knowledge bases. The knowledge graph is organized in a three-layer structure of "protocol field → physical quantity → business concept" and stored in the form of triples. When a new protocol field is encountered, the TransE scoring function is used to search for semantic mapping paths from the new protocol field to existing physical quantities and business concepts in the knowledge graph. The higher the path reasoning score, the more reliable the mapping.

[0058] In one embodiment, the knowledge graph is constructed as follows: First, domain experts define a three-layer ontology structure of "protocol fields → physical quantities → business concepts," including a protocol field layer (such as "0x01 field" and "temperature register"), a physical quantity layer (such as "soil temperature," "air temperature," and "crop canopy temperature"), and a business concept layer (such as "crop growth environment" and "irrigation decision basis"). Then, the correspondence between fields, physical quantities, and business concepts is extracted from the documents and labeled data of known protocols to fill the initial triples. When a new protocol is encountered, the TransE scoring function is used to search the knowledge graph for mapping paths from the new protocol fields to existing physical quantities and business concepts, thus achieving semantic reasoning. In another embodiment, the knowledge graph can also automatically import field semantic definitions and mapping relationships from existing agricultural IoT standards (such as ISO 11783, IEEE 1451, etc.) to accelerate the construction of the knowledge base.

[0059] Semantic extraction transfer learning semantic annotation algorithm: Suitable for scenarios where there is sufficient labeled data for known protocols but a lack of labeled data for new protocols. A field semantic classifier is trained on the source domain (labeled data of known protocols), and domain adaptation is achieved through adversarial training. This makes the feature distribution learned by the feature extractor indistinguishable between the source and target domains (new protocols), thereby transferring the classification ability of known protocols to the new protocol. This algorithm solves the difficulty of supervised learning in the early stages of a new protocol due to a lack of labeled data.

[0060] In one embodiment, transfer learning is implemented as follows: a field semantic classifier is trained on a source domain (such as an RS485 temperature and humidity sensor protocol with a large amount of labeled data) to learn the mapping relationship between field byte patterns and physical quantity semantics; through adversarial training in domain adaptation, a domain discriminator is trained to distinguish whether features come from the source domain or the target domain (new protocol), while a feature extractor is trained to make the domain discriminator unable to distinguish, thereby achieving feature distribution alignment from the source domain to the target domain; after alignment, the semantic classifier trained on the source domain can be directly applied to the semantic annotation of new protocol field fields in the target domain. In another embodiment, an instance-based transfer learning approach can also be used: a subset of samples from the source domain that is most similar to the data distribution of the target domain is selected, and a classifier is directly trained using this subset and applied to the target domain.

[0061] Semantic extraction and contrastive learning field alignment algorithm: Applicable to scenarios where the same device reports data through different protocols. Positive and negative sample pairs are constructed for data reported by the same device through different protocols: fields with the same semantics are positive sample pairs, and fields with different semantics are negative sample pairs. A contrastive learning model is trained to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs in the embedding space, thereby achieving cross-protocol semantic alignment. This algorithm is particularly effective in scenarios where the communication protocol is changed during device upgrades or when the same device supports multiple protocols concurrently.

[0062] In practical deployments, the three semantic extraction algorithms mentioned above can be used independently or in combination. For example, for a new protocol with a structure similar to a known protocol, an initial semantic mapping can be obtained first through knowledge graph reasoning, and then the annotation accuracy can be further optimized through transfer learning. For scenarios where multiple protocols run concurrently on the same device, the rationality of the semantic mapping can be verified by knowledge graph reasoning on the basis of comparative learning alignment.

[0063] Through the progressive processing of the above three stages—protocol identification, frame structure parsing, and semantic extraction—multi-source heterogeneous data frames are parsed into standardized plaintext business data with clear physical meaning, providing a foundation for subsequent data quality checks and fusion analysis.

[0064] In one embodiment, before performing step three, the method for intelligent sensing and data processing of multi-source information on farmland crops and the environment can supplement external reference data through an internet data acquisition interface. The internet data acquisition interface obtains real-time or periodically updated publicly available farmland data resources from smart agriculture cloud platforms, public meteorological data platforms, and soil and crop reanalysis data platforms via API interfaces, and uses the acquired data as part of the standardized plaintext business data. In this embodiment, the internet data acquisition interface uses the HTTPS protocol to connect to the aforementioned external data platforms, and the interface has a built-in API key to achieve legitimate connection. The system presets the data acquisition area to be consistent with the monitoring area. The user selects the data type, including meteorological data (daily average temperature, precipitation, solar radiation), soil background data (soil type, layered soil temperature, layered soil moisture content), and crop reanalysis data. The interface automatically retrieves target data according to the preset area, data type, and time scale, filters invalid data (such as records outside the area or missing data), and automatically downloads the previous day's data every morning. The interface supports incremental data download and full update, allowing users to choose to acquire only updated data or re-acquire the complete dataset according to actual needs, avoiding duplicate acquisition. Furthermore, the interface incorporates a built-in exception handling mechanism, automatically retries and logs responses to data request failures and network interruptions. When the network is interrupted, it supports resume downloading from where it left off, continuing the data download once the network is restored to avoid duplicate downloads and ensure the integrity and reliability of the acquired data. After downloading, the internet data undergoes format conversion and coordinate unification processing to ensure it shares the same spatiotemporal reference as the data collected by the sensing devices.

[0065] By introducing publicly available data resources from external sources through internet data acquisition interfaces, we have achieved integrated and automated collection of multi-source agricultural data. This has enriched the dimensions and coverage of multi-source data, made up for the shortcomings of our own sensing devices in regional-scale data coverage, improved data acquisition efficiency, reduced manual operation costs, and provided a more complete data foundation for subsequent data quality inspection and fusion analysis.

[0066] Step 3: Perform a data quality check on the standardized plaintext business data and supplement any missing data found during the check.

[0067] Perform data quality checks on the standardized plaintext business data generated in step two, and supplement any missing data found during the checks to ensure data reliability.

[0068] (I) Data Quality Inspection. The data quality inspection involves checking the completeness, timeliness, and format of all types of acquired data according to their respective data specifications, removing outliers, and generating a quality check code. The data quality inspection revolves around three core dimensions: data completeness, timeliness, and format. Integrity check: This mainly checks for missing data, including missing key fields, incomplete data records, and insufficient percentage of valid data. For missing data, the type of missing data and the reason for the missing data are marked.

[0069] Timeliness check: Strictly follow the timeliness requirements of observation specifications, check the difference between the data acquisition time and the actual observation time and data transmission time, and eliminate outdated data that exceeds the timeliness threshold and has lost its application value.

[0070] Format and content check: Focus on verifying whether the data format conforms to the preset specifications, whether the field values ​​are within a reasonable range, whether the data units are consistent, and whether the encoding rules are consistent. At the same time, check for abnormal data such as abnormal extreme values ​​and logical contradictions.

[0071] For any abnormal data identified, the anomaly type, location, and cause are recorded in detail, and a standardized quality inspection code is generated simultaneously. This standardized quality inspection code contains key information such as data type, inspection dimension, anomaly level, and processing result, facilitating subsequent data tracing, problem investigation, and data optimization, and providing high-quality data support for subsequent data cleaning, analysis, and application.

[0072] (ii) Missing Data Supplementation. For missing data discovered during quality checks, an adaptive supplementation strategy is selected based on the length of the missing data: This embodiment uses interpolation to fill in a small number of missing values. For short-term time series missing values, weighted linear interpolation or cubic spline interpolation is used, employing effective observation data adjacent to the missing time point for interpolation calculation to recover the missing values. For long-term missing values ​​exceeding 15% of the observation period, the SAITS self-attention temporal completion algorithm is used. Based on the Transformer self-attention mechanism, it utilizes the global dependencies of the time series data to complete the missing values. The temporal completion algorithm adopts a two-stage self-attention architecture. In the first stage, a multi-head self-attention mechanism is used to capture temporal dependencies from the bidirectional context and generate initial completed values. In the second stage, the completed results from the first stage are fused with the original observation values ​​in a weighted combination manner, ensuring that the completed data remains consistent with the original observation information while maintaining temporal consistency. The completed data output by the temporal completion algorithm includes a confidence label, facilitating differentiated processing based on confidence levels during subsequent data fusion analysis.

[0073] Specifically, the input representation of time-series missing data is as follows: Let the original multivariate time series be X = {x1, x2, ..., x...}. T}∈R T×D Where T is the time series length and D is the feature dimension.

[0074] The first-stage completion uses a multi-head self-attention mechanism to encode the data at the observed locations, capture temporal dependencies, and generate initial completion values ​​for missing locations. The completion result output by the first stage is... Represented as:

[0075] Where MSA1(•) represents the first group of masked multi-head self-attention operations, M obs M is a mask matrix for the observation positions (observation positions are 1, missing positions are 0). miss X⊙M is the mask matrix for missing positions (missing positions are 1, observed positions are 0), M is the attention mask matrix (missing positions are set to -∞ to make the attention weights zero, observed positions are set to 0), X⊙M obs This represents the input matrix after retaining the observation location data and setting the missing locations to zero, which serves as the input for the self-attention mechanism.

[0076] The second-stage weighted fusion uses adaptive weights to fuse the first-stage completion results with the re-prediction results of the second group of multi-head self-attention, resulting in the final completion result. Represented as:

[0077] Where MSA2( • ) represents the second group of masked multi-head self-attention operations. The completed matrix output from the first stage is used as the input to the second-stage self-attention mechanism; W comb The adaptive combined weights derived from the attention output are calculated through linear transformation and the Sigmoid activation function, and are used to automatically balance the contribution ratio of the first-stage completion results and the second-stage re-prediction results.

[0078] During model training, the loss function only calculates the prediction error at the observation location. Backpropagation is used to optimize the attention parameters and combined weights so that the completion result is as close as possible to the true value at the observation location.

[0079] The temporal completion algorithm outputs completed data with a confidence label. The confidence score is calculated based on the ratio of the length of consecutive missing segments to the length of available context before and after the missing segment. confidence i = 1-l miss / l context Among them l miss For the length of continuous missing characters, l context This represents the length of the available context before and after the missing segment. The longer the missing segment and the shorter the context, the lower the confidence level. Completed data with a confidence level below a threshold (e.g., 0.3) is marked as "low confidence" for upper-layer application decision-making reference.

[0080] Through the above adaptive missing data imputation strategy, data with different missing patterns are handled appropriately, ensuring the integrity of the dataset and the reliability of subsequent analysis.

[0081] Step 4: Based on the location coordinates and acquisition time of the ground IoT sensor, perform spatiotemporal matching, and realize multi-source data fusion analysis through multi-source data fusion algorithm.

[0082] like Figure 3 As shown, a three-dimensional monitoring network of "space-air-ground" is constructed. This network integrates multi-parameter, multi-scale, and multi-temporal resolution data on crop growth and farmland environment, including data collected and reanalyzed from satellite remote sensing, UAV aerial photography, and ground observation stations. Through spatiotemporal matching, the multi-source data undergoes temporal and spatial calibration to form observational data with a unified spatiotemporal benchmark. Using the location coordinates and acquisition time of ground IoT sensors as a benchmark, the multi-source data supplemented in step three is spatiotemporally matched, and multi-source data fusion analysis is achieved through multi-source data fusion algorithms.

[0083] (i) Spatiotemporal matching. The spatiotemporal matching includes spatial registration and temporal registration.

[0084] The spatial registration process unifies the conversion of multi-source heterogeneous data into the CGCS2000 coordinate system. For data from different sources that may use different coordinate systems, coordinate system identification is first performed. Then, a seven-parameter Bursa-Wolf model is used to achieve accurate conversion from different coordinate systems to the CGCS2000 national geodetic coordinate system. The seven parameters include three translation parameters, three rotation parameters, and one scale parameter, ensuring the consistency of multi-source data in spatial dimensions.

[0085] The time registration employs a bidirectional time-stamp exchange algorithm to estimate clock skew and propagation delay, and compensates for clock drift linearly by fitting the drift rate using least-squares fitting. Different protocols / devices have varying sampling periods (e.g., 1Hz for Wi-Fi sensors, 0.1Hz for LoRa sensors, and 10Hz for RS485 instruments), and each clock exhibits drift. Time registration aligns asynchronous sampled data to a unified time grid, eliminating clock skew and sampling interval differences.

[0086] (1) Clock synchronization: a two-way time stamp exchange algorithm (IEEE 1588 PTP simplified version) is adopted. The master node and the slave node exchange time stamp messages to estimate the clock offset and propagation delay and correct the local clock of the slave node.

[0087] The formulas for calculating clock offset and propagation delay are as follows: Offset = ((t2-t1)-(t4-t3)) / 2 Delay = ((t2-t1)+(t4-t3)) / 2 Where t1 is the master node's sending timestamp, t2 is the slave node's receiving timestamp, t3 is the slave node's sending back timestamp, and t4 is the master node's receiving timestamp.

[0088] The node immediately corrects its local clock based on the offset value obtained from a single synchronization: t sync = t local - Offset Among them, t local t is the reading from the node's local clock. sync This is the clock value after real-time synchronization correction.

[0089] (2) Timestamp correction: After long-term operation, the crystal oscillator drift causes accumulated timestamp deviation. Linear clock drift compensation is performed based on the known synchronization point. The linear drift compensation formula is as follows: t corrected = t raw + ( Δt0+ ρ · ( t raw - t ref )) Where t raw The original timestamp to be corrected is given by Δt0, where Δt0 is the initial offset, ρ is the drift rate (on the order of ppm), and t is the drift rate. ref For reference synchronization time.

[0090] The drift rate ρ is estimated using least squares fitting: ρ=(N∑t i Δ i -∑t i ∑Δ i ) / (N∑t i 2 -(∑t i ) 2 ) Where Δ i = t sync,i - t local,i Let t be the offset of the i-th synchronization. i Let N be the sampling time of the i-th synchronization, and N be the number of synchronizations.

[0091] Instantaneous clock synchronization correction solves the clock deviation at a single synchronization moment, while linear clock drift compensation eliminates the long-term cumulative time deviation caused by crystal oscillator drift. Together, they achieve accurate registration in the time dimension.

[0092] (II) Multi-source data fusion analysis The multi-source data fusion algorithm includes spatial completion, numerical fusion, uncertainty assessment, and timeliness control.

[0093] Kriging interpolation is used for spatial completion. Based on station observation data, Kriging interpolation utilizes a spatial variogram to perform spatial prediction and completion, providing a spatially continuous data field for subsequent numerical fusion. Specifically, a semi-variogram model is constructed based on the spatial distance and orientation between observation points. Known observation points within the neighborhood of the location to be estimated are selected. The optimal weights of each observation point are solved by minimizing the estimation variance, resulting in an optimal linear unbiased estimate of the location to be estimated, thus achieving alignment and completion of spatially heterogeneous data.

[0094] Numerical fusion is performed using Kalman filtering, probability density matching, or optimal interpolation. Kalman filtering achieves dynamic fusion of multi-source data through recursive updates of the state-space model, inputting observations from each source sequentially into the filter and using the noise covariance of each source observation as weights to achieve optimal estimation. The probability density matching method and optimal interpolation are selected based on the density conditions of the station observation data: the probability density matching method is used when the station observation data does not meet the preset density conditions, while the optimal interpolation method is used when the station observation data meets the preset density conditions.

[0095] For the specific implementation of the probability density matching method, firstly, cumulative distribution functions are constructed for both historical station observation data and reanalysis data from the same period. Then, the quantiles of the grid points in the reanalysis data to be corrected are calculated within the cumulative distribution function. Finally, these quantiles are substituted into the inverse function of the cumulative distribution function for the station observation data to obtain the corrected values. The corrected reanalysis data is statistically consistent with the station observations, achieving bias correction without altering the spatial structure of the reanalysis data. In another embodiment, a non-parametric implementation can also be used: the grid points in the reanalysis data are sorted by size to obtain their ranking; the ranking is then converted into quantiles and substituted into the observation quantile function to directly obtain the corrected values.

[0096] For the specific implementation of the optimal interpolation method, the reanalysis data is used as the initial estimated field (background field), and the station observations are used as the observed values. The optimal weights are solved in the sense of minimum variance to minimize the variance of the analysis field error. Specifically, firstly, the background error covariance matrix and the observation error covariance matrix are calculated. The background error is estimated by the statistical difference between the reanalysis data and the station observations, and the observation error is given by the instrument accuracy index. Then, the optimal weights are solved so that the analysis field is the background field plus the optimal weights multiplied by the observation increment (the station observation value minus the interpolation of the background field at the station location), thus achieving the minimum variance estimation.

[0097] Dempster evidence theory is used to assess the uncertainty of the fusion results. Each source observation is transformed into a basic probability assignment in the hypothesis space. The basic probability assignments from multiple sources are fused using Dempster's combination rule, and the credibility evaluation of the fusion results is generated by combining the confidence and support of the multi-source evidence.

[0098] Simultaneously, a data fusion time window is set to address data latency. The fusion time window is used to address inconsistencies in arrival times caused by transmission delays from different data sources, ensuring the timeliness of the fused data.

[0099] Step 5: Develop a universal and open format for precision irrigation observation data products, and generate standardized data products from the merged data.

[0100] The precision irrigation observation data product format is based on XML, clearly defining data encoding rules and field definitions. It covers all the core data required for precision irrigation and supports custom extended fields. The core fields of the data product format include four types of parameters: crop parameters (crop type, growth stage, canopy coverage, crop NDVI, LAI), soil parameters (soil moisture, soil temperature, soil nitrogen and phosphorus content), meteorological parameters (air temperature and humidity, daily maximum temperature, daily minimum temperature, sunlight, precipitation, wind speed), and irrigation parameters (irrigation time, irrigation amount, irrigation method). The data format supports custom extended fields to meet the precision irrigation monitoring needs of different regions and crop types.

[0101] In one embodiment, the data product format is applied to a precision irrigation monitoring scenario for maize farmland. The crop parameters specifically include: crop type as maize, growth period including sowing, tillering, and jointing stages, canopy coverage in percentage, and crop coefficient recorded daily in JSON format during the growth period; the soil parameters specifically include: soil moisture in percentage of volumetric water content, soil temperature in degrees Celsius, and soil type including loam, sandy loam, etc.; the meteorological parameters specifically include: air temperature in degrees Celsius, air humidity in percentage relative humidity, sunshine duration in hours, and precipitation in millimeters; the irrigation parameters specifically include: irrigation time in YYYY-MM-DD HH:MM:SS format, irrigation volume in cubic meters per acre, and irrigation methods including drip irrigation, sprinkler irrigation, etc. Furthermore, the data product format supports extended fields, such as adding fields for pest and disease occurrence in the maize monitoring scenario.

[0102] The data product format is compatible with mainstream data storage formats. Raster image data (such as soil moisture images and crop water consumption images) adopts GeoTIFF format, with soil moisture images having a resolution of no less than 500 meters, preferably within 100 meters; structured data is stored in JSON format for easy database storage and later querying. In one embodiment, the resolution of raster images such as soil moisture and crop water consumption is 100 meters, the resolution of satellite-related raster images is 10 meters, and the resolution of UAV orthophoto images is 0.5 meters; image files are named in the format {imageid}_{date}.tif; the data table structure conforms to relevant specifications, and a unique index of plantid+sdate+edate and a general index of the plant field are established in the maize growth period and crop coefficient data table to facilitate data querying and management. This data format is publicly available and supports cross-platform and cross-system data sharing and reuse, realizing data interoperability between different agricultural information platforms.

[0103] In one implementation, this invention employs a combination of "edge computing + hierarchical transmission." An industrial-grade edge gateway is deployed as an edge computing node in the core monitoring area of ​​farmland. This gateway receives raw monitoring data collected by surrounding ground sensors, raw aerial imagery from UAVs, and grid data on local crop growth and soil moisture transmitted by satellite. The raw data undergoes local preprocessing, and the final data product is generated. Only the final standardized data product is transmitted, reducing data transmission complexity and cost. Specific implementation details are as follows: (a) Edge computing: An industrial-grade edge gateway is deployed as an edge computing node in the core monitoring area of ​​corn farmland. This node connects locally to ground sensor arrays and drones, receiving raw monitoring data collected by ground sensors, raw aerial imagery from drones, and local image fragments of the corn monitoring area transmitted by satellite. The edge computing node has a built-in lightweight data processing module that performs the data parsing and preprocessing, data quality checks, data missing data supplementation, spatiotemporal matching, multi-source data fusion, and format standardization conversion operations described in steps two to five. It directly generates final data products conforming to a common format, including corn canopy coverage reports, soil moisture assessment data, daily meteorological summary data, and irrigation recommendation parameters. Redundant raw data is removed, retaining only the core and valid information. Large-capacity raw data such as satellite and drone imagery are cached locally on the edge computing node (500GB cache capacity) without long-distance transmission, significantly reducing transmission pressure and bandwidth costs.

[0104] (ii) Layered transmission: The transmission is divided into three layers: ① Perception Layer – Edge Node Transmission: Raw data collected by ground sensors is transmitted to a multi-protocol compatible module via RS485 / ZigBee protocol. After being converted into standardized data, Wi-Fi is used for short-distance (≤100m) transmission, and LoRa is used for long-distance (>100m) transmission, achieving low-cost short-distance transmission of raw data to edge computing nodes. Raw images collected by drones are transmitted to edge nodes via LoRa protocol. Satellites only transmit local image fragments of the corn monitoring area to edge nodes, reducing satellite data transmission volume. ② Edge Node – Backbone Network Transmission: The final data products generated by edge computing nodes (the data volume is only about 15% of the raw data) are transmitted using a 5G + fiber optic hybrid network. Real-time data such as corn real-time soil moisture and irrigation early warning are transmitted using 5G, with transmission latency controlled within 15ms and packet loss rate less than 1%. Batch final data products such as daily corn growth reports and weekly soil moisture summaries are transmitted using fiber optics to ensure transmission stability. Compared with transmitting raw data, this reduces transmission bandwidth requirements and traffic costs by more than 70%. Backbone network nodes are equipped with 100GB cache units. When network congestion occurs, the final data product is cached and prioritized for transmission after network recovery. ③ Application Layer Transmission: The data processing layer performs secondary verification and optimization on the final data product transmitted from the edge computing nodes, and then transmits it to the application layer using the AES encryption algorithm. Business data in interface request parameters and returned results are encrypted, and the integrity of the transmitted final data product is verified using the CRC32 checksum algorithm to ensure secure and accurate data transmission. A proactive push mode from lower-level edge nodes is adopted to push the processed final data product to the upper-level precision irrigation management platform, meeting business needs and further simplifying the transmission process.

[0105] Secondly, such as Figure 4As shown, this embodiment of the invention also provides a system for intelligent sensing and data processing of multi-source information on farmland crops and the environment, comprising: The multi-protocol compatibility module is used to adapt the transmission protocols of multi-source heterogeneous data shared from ground IoT sensor arrays, UAV monitoring equipment and satellite remote sensing observation data centers and establish data communication links. The data parsing and preprocessing module is used to perform data parsing and preprocessing on multi-source heterogeneous data frames transmitted through the data communication link to generate standardized plaintext business data. The data quality control module is used to perform data quality checks on the standardized plaintext business data and to supplement any missing abnormal data found during the checks. The spatiotemporal registration and fusion module is used to perform spatiotemporal matching on the supplemented multi-source data based on the location coordinates and acquisition time of the ground IoT sensor, and to realize multi-source data fusion analysis through multi-source data fusion algorithm; The data standardization module is used to develop a universal and open format for precision irrigation observation data products based on the XML language, and to generate standardized data products from the merged data.

[0106] In one embodiment, the system further includes an Internet data acquisition module, which is used to acquire publicly available farmland data resources from an external data platform via an Internet data acquisition interface in the form of an API interface, and to use the acquired data as part of the standardized plaintext business data.

[0107] Furthermore, in a third aspect, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for intelligent sensing and data processing of multi-source information on farmland crops and the environment.

[0108] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent sensing and data processing of multi-source information on farmland crops and the environment, characterized in that, Includes the following steps: Step 1: Construct a multi-protocol compatible module to adapt the transmission protocols of multi-source heterogeneous data shared by ground IoT sensor arrays, UAV monitoring equipment and satellite remote sensing observation data centers, and establish data communication links; Step 2: Perform data parsing and preprocessing on the multi-source heterogeneous data frames transmitted through the data communication link to generate standardized plaintext service data; Step 3: Perform a data quality check on the standardized plaintext business data and supplement any missing abnormal data found during the check; Step 4: Based on the location coordinates and acquisition time of the ground IoT sensors, perform spatiotemporal matching on the supplemented multi-source data, and realize multi-source data fusion analysis through multi-source data fusion algorithm; Step 5: Develop a universal and open format for precision irrigation observation data products, and generate standardized data products from the merged data.

2. The method according to claim 1, characterized in that, The multi-protocol compatibility module mentioned in step one includes a multi-protocol communication module and a main controller. The main controller runs a protocol adaptive engine to automatically identify the protocol channel of the multi-protocol communication module, automatically match communication parameters and load the corresponding protocol driver. When there is no matching item in the driver library, a general driver template is loaded to enter the downgrade adaptation mode, and data is packaged and submitted through a unified data reading interface.

3. The method according to claim 1, characterized in that, The data parsing preprocessing in step two includes sequentially performing protocol identification, frame structure parsing, and semantic extraction on multi-source heterogeneous data frames; wherein, the protocol identification is used to determine the protocol category of the data frame, the frame structure parsing is used to mark the field type of each byte and infer the field boundaries, and the semantic extraction is used to establish the mapping relationship between protocol fields and business concepts.

4. The method according to claim 3, characterized in that, The protocol identification includes: extracting multi-dimensional feature vectors of data frames to train a multi-class random forest model to determine the protocol category, or constructing a byte-level N-gram probability model to calculate the log-likelihood value for matching known protocols; when the protocol category of a data frame cannot be determined, density clustering is performed based on the feature vectors to discover unknown protocols.

5. The method according to claim 3, characterized in that, The frame structure parsing includes: modeling the data frame parsing as a sequence labeling task; extracting byte-level context features through a bidirectional long short-term memory network and modeling and labeling transition constraints through a conditional random field layer; labeling the field type of each byte in the data frame; inferring field boundaries by combining the mutual information between adjacent bytes; and distinguishing fixed fields from data fields by analyzing the variation patterns of bytes at the same position in multiple frames of data.

6. The method according to claim 3, characterized in that, The semantic extraction includes: constructing a knowledge graph between protocol fields and physical quantities and business concepts, and using graph reasoning to automatically discover the semantic mapping path of new protocol fields; migrating the semantic classifier of fields of known protocols to new protocols through domain adaptation; and constructing positive and negative sample pairs for data reported by the same device through different protocols, so that semantically identical fields are aligned in the embedding space.

7. The method according to claim 1, characterized in that, Before the data quality check described in step three, publicly available farmland data resources on the Internet are obtained from an external data platform via an API interface through an Internet data acquisition interface, and the obtained data is used as part of the standardized plaintext business data.

8. The method according to claim 1, characterized in that, The data quality check described in step three focuses on three dimensions: data integrity, timeliness, and format content, and investigates abnormal data to form a standardized quality check code. The standardized quality check code contains information on data type, inspection dimension, anomaly level, and processing result.

9. The method according to claim 1, characterized in that, In step three, the missing data completion strategy is adaptively selected based on the length of the missing data: for a small number of missing data, interpolation is used; for long-term missing data, a temporal completion algorithm based on the Transformer self-attention mechanism is used. The algorithm captures temporal dependencies from the bidirectional context through a two-stage self-attention architecture to generate initial completion values, and then fuses the completion results with the original observations in a weighted combination manner. The completed data output by the temporal completion algorithm is accompanied by confidence labels.

10. The method according to claim 1, characterized in that, The spatiotemporal matching in step four includes spatial registration and temporal registration. Spatial registration converts multi-source heterogeneous data into the CGCS2000 coordinate system and uses a seven-parameter Bursa model for coordinate transformation. Temporal registration uses a two-way timescale exchange algorithm to estimate clock offset and propagation delay, and performs linear clock drift compensation by fitting the drift rate with least squares.

11. The method according to claim 1 or 10, characterized in that, The multi-source data fusion algorithm described in step four includes: spatial completion using Kriging interpolation; numerical fusion using Kalman filtering, probability density matching, or optimal interpolation, wherein the probability density matching method is used when the station observation data does not meet the preset density conditions, and the optimal interpolation method is used when the station observation data meets the preset density conditions; uncertainty assessment of the fusion result is performed using DS evidence theory; and a data fusion time window is set to address data delays.

12. The method according to claim 1, characterized in that, The precision irrigation observation data product format described in step five covers at least one of crop parameters, soil parameters, meteorological parameters, and irrigation parameters, and supports custom extended fields.

13. The method according to claim 1, characterized in that, The multi-protocol compatibility module adapts the transmission protocols of multi-source heterogeneous data and transmits it to edge computing nodes deployed in the core monitoring area of ​​farmland via short-range wireless local area network or long-range low-power wireless method.

14. The method according to claim 13, characterized in that, Steps two through five are completed in the edge computing node to generate the standardized data product.

15. The method according to claim 1 or 14, characterized in that, The standardized data product is transmitted using a hybrid network of fifth-generation mobile communication and optical fiber. Real-time data in the standardized data product is transmitted via fifth-generation mobile communication, while large-capacity data in the standardized data product is transmitted via optical fiber.

16. A system for intelligent sensing and data processing of multi-source information on farmland crops and the environment, characterized in that, include: The multi-protocol compatibility module is used to adapt the transmission protocols of heterogeneous data from multiple sources, including ground-based IoT sensor arrays, UAV monitoring equipment, and satellite remote sensing observation data centers, and to establish data communication links. The data parsing and preprocessing module is used to perform data parsing and preprocessing on multi-source heterogeneous data frames transmitted through the data communication link to generate standardized plaintext business data. The data quality control module is used to perform data quality checks on the standardized plaintext business data and to supplement any missing abnormal data found during the checks. The spatiotemporal registration and fusion module is used to perform spatiotemporal matching on the supplemented multi-source data based on the location coordinates and acquisition time of the ground IoT sensor, and to realize multi-source data fusion analysis through multi-source data fusion algorithm; The data standardization module is used to develop a universal and open format for precision irrigation observation data products, generating standardized data products from the merged data.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for intelligent sensing and data processing of multi-source information on farmland crops and the environment as described in any one of claims 1 to 15.

Images