Large-model cotton planting decision-making method and system fusing remote sensing and multi-dimensional data

Abstract

The invention discloses a large-model cotton planting decision-making method and system fusing remote sensing and multi-dimensional data, and relates to the technical field of cotton planting, and the method comprises the steps: collecting the complete-cycle data of cotton planting, carrying out the preprocessing, and building a data set; an LLaMA model is improved by adopting a block extension method, an extension block part is trained, and original LLaMA model parameters are frozen; performing parameter fine adjustment on a dimension reduction matrix and a dimension raising matrix of a Transform block in the improved LLaMA model to obtain a large decision model; establishing a vector knowledge base, retrieving the vector knowledge base according to an original query of a user, and constructing an enhanced context; establishing an intention recognition agent, performing intention recognition on the original query by the intention recognition agent, obtaining multi-dimensional calling data based on the recognized intention, and establishing a comprehensive calling result; and inputting the comprehensive calling result and the enhanced context into a large decision model to obtain decision suggestions. The invention constructs a decision-making assistant capable of guiding cotton production.

Description

Technical Field

[0001] This invention relates to the field of cotton planting technology, and more specifically to a large-scale cotton planting decision-making method and system that integrates remote sensing and multidimensional data. Background Technology

[0002] Cotton is a core economic crop in some regions of my country, but cotton production is highly dependent on professional technical guidance. Currently, there is a significant information gap between government agricultural technicians, cotton experts, and frontline farmers. Farmers struggle to obtain the latest agricultural and policy information in key stages such as planting, pest and disease control, and harvesting, leading to losses in yield and income. Existing technologies suffer from limitations: general-purpose large-scale models are ill-suited to local conditions and lack the necessary data processing capabilities, which, if applied to agricultural practices, directly result in economic losses for farmers; cotton cultivation is highly dependent on real-time weather, and static large-scale models cannot provide decision-making suggestions based on real-time data; farmers' actual needs involve taking photos to check for pests and diseases or inputting data to assess growth, which pure LLM (Limited Range Model) cannot accomplish; existing meteorological, remote sensing, and pest and disease identification tools are typically isolated apps or systems that cannot work collaboratively, lacking an intelligent central hub to connect them.

[0003] Therefore, how to construct a decision-making assistant that can guide cotton production is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a large-scale cotton planting decision-making method and system that integrates remote sensing and multidimensional data to solve the problems existing in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A large-scale model-based cotton planting decision-making method integrating remote sensing and multidimensional data includes: Collect full-cycle data on cotton cultivation, preprocess the full-cycle data, and establish a professional text dataset and a QA dataset; The LLaMA model is improved by using a block expansion method. The expansion block is trained using the professional text dataset, and the parameters of the original LLaMA model are frozen. The parameters of the dimensionality reduction matrix and the dimensionality increase matrix of the Transformer block in the improved LLaMA model are fine-tuned using the QA dataset to obtain a large decision model. Semantic segmentation is performed on the full-cycle data to establish a vector knowledge base; the vector knowledge base is retrieved based on the user's original query to obtain knowledge fragments, and an enhanced context is constructed based on the original query and knowledge fragments. An Agent intent recognition intelligent agent is established, which performs intent recognition on the original query, obtains multi-dimensional call data based on the recognized intent, and establishes a comprehensive call result; The integrated call results and enhanced context are input into the decision-making model to obtain decision recommendations.

[0006] Preferably, the full-cycle data includes literature, technical reports, and professional books related to cotton planting; cotton planting data and physiological indicator data of the research group over the years; cotton planting experience collected from field visits to farmers; cotton Q&A data from the big data platform; and synthetic Q&A data generated by combining self-guidance methods with cotton professional materials.

[0007] Preferably, the preprocessing uses the Simhash algorithm for fuzzy deduplication and combines it with precise matching to clean the data.

[0008] Preferably, all the original Transformer blocks are frozen. Then, after the 16th and 24th layers of the original LLaMA model, two new Transformer blocks are inserted respectively. The last linear layer weight of the new Transformer block is reset to 0, making it an identity block and keeping the initial output unchanged. During the incremental pre-training process, the original model parameters are frozen, and only the new Transformer blocks are trained. The training data uses the professional text dataset.

[0009] A large-scale cotton planting decision-making system integrating remote sensing and multidimensional data includes: The data collection unit collects data from the entire cotton planting cycle, preprocesses the data, and establishes a professional text dataset and a QA dataset. The decision model building unit improves the LLaMA model using a block expansion method. It trains the expansion block part using the professional text dataset and freezes the parameters of the original LLaMA model. It then uses the QA dataset to fine-tune the parameters of the dimensionality reduction and dimensionality increase matrices of the Transformer block in the improved LLaMA model to obtain the large decision model. The semantic construction unit performs semantic segmentation on the full-cycle data and establishes a vector knowledge base; it retrieves knowledge fragments from the vector knowledge base based on the user's original query and constructs an enhanced context based on the original query and knowledge fragments. The intelligent agent construction unit establishes an Agent intent recognition intelligent agent, which performs intent recognition on the original query, obtains multi-dimensional call data based on the recognized intent, and establishes a comprehensive call result; The decision suggestion acquisition unit inputs the comprehensive call result and the enhanced context into the decision big model to obtain decision suggestions.

[0010] As can be seen from the above technical solution, compared with the prior art, this invention discloses a large-scale cotton planting decision-making method and system that integrates remote sensing and multidimensional data. Through cotton data governance and incremental pre-training, the model deeply masters the professional terminology of cotton planting and the management measures at different stages, solving the problem of general models pretending to know everything in the cotton field. The model's answers are based on a reliable cotton knowledge base and can be combined with real-time information, effectively avoiding production accidents that may be caused by model illusions. Through roleplay fine-tuning, the model is no longer a cold machine, but simulates a cotton agricultural expert, providing friendly and approachable answers that are easier for farmers to accept. Through agent technology, macro-level remote sensing (growth, nutrition, water, pest and disease remote sensing), micro-level diagnosis (user-uploaded photo recognition), and real-time data (meteorology) are integrated. As an intelligent hub, LLM integrates multiple single-function ML / DL models and APIs into a powerful, one-stop decision-making system, solving the problem of data silos in traditional solutions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the structure provided by the present invention; Figure 2 This is a schematic diagram illustrating the LoRa training principle provided by the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] This invention discloses a large-scale cotton planting decision-making method that integrates remote sensing and multidimensional data, such as... Figure 1 As shown, the details are as follows: Collect full-cycle data on cotton cultivation, preprocess the full-cycle data, and establish a professional text dataset and a QA dataset; Step 1: Construction and Governance of the Cotton Vertical Domain Dataset Execution entity: Data processing server.

[0015] Description: To address the issue of fragmented and incomplete data sources in the cotton industry, a special data collection and governance program was implemented. (1) Multi-source data collection: Targeted collection of cotton planting lifecycle data, including: Professional knowledge: Literature, technical reports, and professional books related to cotton cultivation.

[0016] Agricultural data: cotton planting data and physiological index data of the research group over the years (such as irrigation, fertilization and chemical control schemes at different times).

[0017] Empirical data: First-hand cotton growing experience collected through field visits to farmers.

[0018] Platform data: Cotton Q&A data from big data platforms such as "Digital Poplar".

[0019] Synthetic data: Synthetic question-and-answer data generated using a self-guided approach combined with cotton-related professional materials.

[0020] (2) Data governance (cleaning and deduplication): For the large amount of duplicate and unstructured text in the cotton data, the Simhash algorithm is used for fuzzy deduplication, and the dataset is cleaned by combining precise matching to ensure that the data used for training is of high quality and consistent.

[0021] (3) Data classification: The treated data is divided into cotton professional text for incremental pre-training and cotton agricultural QA pairs for instruction fine-tuning.

[0022] The LLaMA model is improved by using a block expansion method. The expansion block is trained using the professional text dataset, and the parameters of the original LLaMA model are frozen. The decision-making large model is obtained by using the dimensionality reduction matrix and the dimensionality increase matrix of the Transformer block in the improved LLaMA model using the QA dataset. Step 2: Incremental pre-training based on block expansion (injecting cotton domain knowledge) Execution entity: Model training server.

[0023] Description: To enable the base model (LLaMA) to master cotton terminology and implicit knowledge while avoiding "catastrophic forgetting," the following strategy is used for incremental pre-training: (1) Model structure expansion: The block expansion method is adopted. A new Transformer block is added after the original Transformer block.

[0024] (2) Identity block initialization: By resetting the last linear layer weight of the newly added block to 0 (becoming Zero-Linear), it is initialized as an identity block, keeping the initial output unchanged.

[0025] (3) Cotton knowledge injection: During incremental pre-training, the original model parameters are frozen, and training is performed only on the newly added blocks. The training data uses the cotton-specific text dataset constructed in step 1, and a certain proportion of general high-quality datasets are mixed in to mitigate catastrophic forgetting.

[0026] Semantic segmentation is performed on the full-cycle data to establish a vector knowledge base; the vector knowledge base is retrieved based on the user's original query to obtain knowledge fragments, and an enhanced context is constructed based on the original query and knowledge fragments. Step 3: Fine-tune instructions based on the cotton agricultural expert role (aligning with human preferences) Execution entity: Model training server.

[0027] describe: (1) Data preparation: Use the cotton planting QA dataset after treatment in step 1.

[0028] (2) Roleplay Alignment: Modify the QA data to "roleplay" so that the model simulates a friendly and professional cotton agricultural expert when answering questions, thereby improving the user interaction experience.

[0029] (3) Efficient parameter fine-tuning: Using LORA (Low-Rank Adaptation) or QLORA methods (such as...) Figure 2 As shown, the original large model parameters are fixed, and only the dimensionality reduction (A) and dimensionality increase (B) matrices of the bypass are trained. Through multiple rounds of training using cotton QA data and Roleplay data, the model acquires the ability to follow instructions in the cotton domain. The cotton QA data is used to construct the knowledge boundary, and the Roleplay data is used to construct language style constraints. During fine-tuning, the low-rank adaptation (LoRA) technique is used to map these data features to parameter changes in the dimensionality reduction and dimensionality increase matrices. By continuously penalizing the model's outputs that deviate from expert tone and professional knowledge, the model parameters are guided to converge, ultimately enabling it not only to possess professional knowledge but also to simulate the interaction style of agricultural experts, achieving a high level of instruction compliance.

[0030] Step 4: Cotton knowledge base retrieval based on RAG (alleviating illusion and timeliness issues)

[0031] Execution entity: Assistant system backend server (RAG module).

[0032] Description: Addresses the issues of illusions and information lag in large-scale cotton farming models regarding agricultural activities (such as weather and irrigation): (1) Construction of cotton knowledge base: The cotton QA pairs, books, technical reports, etc. processed in step 1 are stored in the vector knowledge base (Vector DB) after semantic segmentation.

[0033] (2) Query-retrieval: When a user asks a question about cotton planting (such as "Is the current weather suitable for applying defoliant?"), the system first retrieves the most relevant cotton knowledge fragments from the vector knowledge base.

[0034] (3) Context Enhancement Generation: Combine the user's original query and the retrieved knowledge fragments into an "Enhanced Context" and input it into the large model after fine-tuning in step 3.

[0035] (4) Response generation: The model generates accurate answers based on real-time and reliable cotton knowledge base content to ensure the accuracy and timeliness of agricultural guidance.

[0036] An Agent intent recognition intelligent agent is established, which performs intent recognition on the original query, obtains multi-dimensional call data based on the recognized intent, and establishes a comprehensive call result; The integrated call results and enhanced context are input into the decision-making model to obtain decision recommendations.

[0037] Step 5: Agent-based "Sky-Earth-Human" Multi-model Fusion (Extending Decision-Making Capabilities)

[0038] Execution entity: Assistant system backend server (Agent module).

[0039] Description: Based on agent frameworks such as MetaGPT, an intelligent agent is built that can collaboratively invoke external tools to achieve complex agricultural decisions that LLM (Low-Level Machine) cannot handle independently. Specifically, this intelligent agent system includes the following four core functional modules in its logical architecture: The perception and memory module receives the user's original query input and maintains the historical dialogue context. This module can parse the user's natural language commands and extract key information such as time (e.g., 'the next three days'), location (e.g., 'coordinates of a specific plot of land'), and entity (e.g., 'red spider', 'defoliant').

[0040] Cognitive Planning Module: Acting as the 'brain' of the intelligent agent, this module incorporates an intent recognition algorithm based on a large language model. It employs a thought chain reasoning model, breaking down complex user needs into a sequence of executable sub-tasks. For example, the task 'Help me check what's wrong with this land' is broken down into two sub-steps: 'Retrieve remote sensing imagery' and 'Query historical growth records'.

[0041] Tool Registry: This module predefines and encapsulates a standardized set of API interfaces, including: Macro Remote Sensing Interface: This encapsulates the Google Earth Engine (GEE) API for obtaining vegetation indices such as NDVI and EVI; Microscopic diagnostic interface: Encapsulates a pest and disease identification model based on YOLO / ResNet for processing user-uploaded images; Environmental perception interface: Connects to third-party meteorological services to obtain real-time data such as temperature, humidity, and rainfall. The intelligent agent automatically selects and matches the corresponding tool interface based on instructions from the planning module.

[0042] The Execution & Observation module is responsible for executing the tool call code, capturing multidimensional heterogeneous data returned by the API (such as meteorological data in JSON format, remote sensing image results in raster format, and diagnostic reports in text format), and standardizing, cleaning, and aligning these multi-source data to form a comprehensive call result.

[0043] (1) Intent recognition: The Agent (LLM as the brain) analyzes the user query and determines whether the user needs knowledge Q&A, real-time weather, macro remote sensing analysis or micro disease diagnosis.

[0044] (2) Macroscopic remote sensing call (day): Data Acquisition: If a user queries the crop growth of a plot, the Agent automatically calls platform interfaces such as GEE (Google Earth Engine) to obtain the latest remote sensing images of the specified plot.

[0045] Professional Analysis: The agent distributes remote sensing images to multiple internally integrated ML / DL models for growth detection, nutrient detection, pest and disease remote sensing monitoring, and water content analysis.

[0046] (3) Microscopic diagnostic call (location): Tool Invocation: If a user uploads photos of cotton leaves, the Agent invokes an existing deep learning pest and disease identification model and returns specific diagnostic results.

[0047] (4) Real-time data retrieval (by human): Tool Invocation: If a user inquires about the weather, the Agent invokes an external meteorological API to obtain real-time weather information and agricultural advice.

[0048] (5) Multidimensional result fusion: The Agent collects the output results of all tools (RAG, remote sensing model, diagnostic model, meteorological API), and the LLM performs the final fusion to generate a comprehensive, multidimensional agricultural decision-making suggestion (e.g., "Remote sensing shows that the northeast corner of your plot is lacking nitrogen, the leaf photo confirms that it is a spider mite, the weather will be sunny for the next 3 days, and we suggest that you...").

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A large-scale model cotton planting decision-making method integrating remote sensing and multidimensional data, characterized in that, include: Collect full-cycle data on cotton cultivation, preprocess the full-cycle data, and establish a professional text dataset and a QA dataset; The LLaMA model is improved by using a block expansion method. The expansion block is trained using the professional text dataset, and the parameters of the original LLaMA model are frozen. The parameters of the dimensionality reduction matrix and the dimensionality increase matrix of the Transformer block in the improved LLaMA model are fine-tuned using the QA dataset to obtain a large decision model. Semantic segmentation is performed on the full-cycle data to establish a vector knowledge base; the vector knowledge base is retrieved based on the user's original query to obtain knowledge fragments, and an enhanced context is constructed based on the original query and knowledge fragments. An Agent intent recognition intelligent agent is established, which performs intent recognition on the original query, obtains multi-dimensional call data based on the recognized intent, and establishes a comprehensive call result; The integrated call results and enhanced context are input into the decision-making model to obtain decision recommendations.

2. The large-scale cotton planting decision-making method integrating remote sensing and multidimensional data according to claim 1, characterized in that, The full-cycle data includes literature, technical reports, and professional books related to cotton cultivation; cotton cultivation data and physiological indicator data from the research group over the years; cotton cultivation experience collected from field visits to farmers; cotton Q&A data from the big data platform; and synthetic Q&A data generated using self-guided methods combined with cotton professional materials.

3. The large-scale cotton planting decision-making method integrating remote sensing and multidimensional data according to claim 1, characterized in that, The preprocessing uses the Simhash algorithm for fuzzy deduplication and combines it with precise matching to clean the data.

4. The large-scale cotton planting decision-making method integrating remote sensing and multidimensional data according to claim 1, characterized in that, The improvements to the LLaMA model specifically include: freezing all of its original Transformer blocks; then, inserting two new Transformer blocks after layers 16 and 24 of the original LLaMA model, resetting the last linear layer weight of the new Transformer blocks to 0 to initialize them as identity blocks, and keeping the initial output unchanged; during incremental pre-training, freezing the original model parameters and training only the new Transformer blocks, using the aforementioned professional text dataset as the training data.

5. A large-scale cotton planting decision-making system integrating remote sensing and multidimensional data, employing the large-scale cotton planting decision-making method integrating remote sensing and multidimensional data as described in any one of claims 1-4, characterized in that, include: The data collection unit collects data from the entire cotton planting cycle, preprocesses the data, and establishes a professional text dataset and a QA dataset. The decision model building unit improves the LLaMA model using a block expansion method. It trains the expansion block part using the professional text dataset and freezes the parameters of the original LLaMA model. It then uses the QA dataset to fine-tune the parameters of the dimensionality reduction and dimensionality increase matrices of the Transformer block in the improved LLaMA model to obtain the large decision model. The semantic construction unit performs semantic segmentation on the full-cycle data and establishes a vector knowledge base; it retrieves knowledge fragments from the vector knowledge base based on the user's original query and constructs an enhanced context based on the original query and knowledge fragments. The intelligent agent construction unit establishes an Agent intent recognition intelligent agent, which performs intent recognition on the original query, obtains multi-dimensional call data based on the recognized intent, and establishes a comprehensive call result; The decision suggestion acquisition unit inputs the comprehensive call result and the enhanced context into the decision big model to obtain decision suggestions.

Images