Intelligent Strawberry Yield Prediction System Based on RS485 Sensing and UAV Vision

Abstract

This invention discloses an intelligent strawberry yield prediction system based on RS485 sensing and UAV vision, comprising an environmental perception module, a visual acquisition module, and an intelligent analysis module. The environmental perception module collects multi-parameter data within a greenhouse through a sensor array. The visual acquisition module uses a cruising UAV to acquire images of the strawberry area. The intelligent analysis module receives the above data, uses visual recognition to count the number of fruits at different maturity levels and estimates their size, obtaining fruit state characteristics. These characteristics, together with environmental data, form a multi-dimensional feature vector, which is converted into natural language prompts according to a structured template and input into a multi-factor dynamic yield estimation model driven by LLM. The model uses a RAG architecture for inference and ultimately outputs the predicted harvestable strawberry yield, confidence interval, and prediction basis for a preset future time period. This invention improves yield prediction accuracy, provides comprehensive visual recognition coverage, ensures a transparent and interpretable decision-making process, and outputs results that combine predictive values ​​with agronomical evidence.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and artificial intelligence, and in particular to a strawberry yield intelligent prediction system based on RS485 sensing and UAV vision. Background Technology

[0002] Precision management of strawberry cultivation relies on accurate forecasting of fruit yield and ripening cycle. Accurate yield prediction is crucial for developing harvesting, sorting, logistics, and sales plans, directly impacting economic benefits and resource allocation efficiency. However, current strawberry yield forecasting methods depend on growers' experience and periodic manual sampling, which has significant limitations. First, manual statistics are inefficient, consuming substantial manpower, and cannot provide rapid and comprehensive assessments of large-scale planting areas. Second, experience-based judgment is highly subjective, greatly influenced by individual differences in experience, making standardization and large-scale application difficult, resulting in high prediction error rates, typically exceeding 25%. Finally, traditional methods only obtain static data at specific points in time, failing to enable dynamic monitoring and early warning of yield changes. To address the reliance on manual labor for strawberry yield forecasting, automated or semi-automated methods have been used; however, existing automated or semi-automated solutions have also failed to effectively address the aforementioned core pain points, specifically as follows: Environmental monitoring suffers from limited dimensions and insufficient stability: Existing technologies typically employ wireless sensor networks such as ZigBee, LoRa, or Wi-Fi to collect environmental parameters. However, in actual agricultural production scenarios such as greenhouses and polytunnels, the complex structures of metal frames, irrigation pipes, and plastic films create severe signal shielding and multipath effects, leading to significant wireless signal attenuation, unstable data transmission, and high packet loss rates. This makes it difficult to guarantee the continuity and reliability of environmental monitoring data, posing a potential threat to subsequent accurate analysis.

[0003] Visual recognition suffers from blind spots and information gaps: Some studies have attempted to use fixed cameras for image recognition to count fruit quantity or determine ripeness. However, strawberry plants are short and have dense foliage, often obscuring the fruit with leaves and vines. Fixed cameras offer only a single perspective, resulting in numerous blind spots and severely distorted statistical data. Furthermore, existing visual recognition solutions typically only identify binary states such as "whether there is fruit" and "whether the fruit is ripe," failing to accurately measure the physical dimensions of the fruit (such as diameter and volume), which are crucial parameters affecting the weight of a single fruit.

[0004] Yield estimation models are often "black boxes" and disconnected from the growth environment: Current automated forecasting models are mostly based on simple fruit counting, multiplying the count by a fixed average fruit weight to estimate total yield. This method ignores the dynamic changes in fruit weight with the growth environment; for example, soil fertility, moisture, temperature, and light directly affect the fruit's enlargement rate and final size. A simple counting model cannot reflect the crucial agronomic fact that "the same number of fruits have significantly different weights under different environmental conditions." Furthermore, existing prediction models based on deep learning and other algorithms are often complex "black box" models, lacking interpretability in their decision-making processes. When predictions deviate, growers cannot trace the cause to environmental factors, pests and diseases, or problems with the model itself, leading to distrust of the model results and difficulty in translating the predictions into concrete agricultural operational guidance.

[0005] In summary, existing technologies suffer from poor data acquisition reliability, incomplete visual recognition coverage, and a single, uninterpretable dimension for yield estimation, resulting in limited accuracy in strawberry yield prediction. Summary of the Invention

[0006] This invention provides an intelligent strawberry yield prediction system based on RS485 sensing and UAV vision to overcome the above-mentioned technical problems.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A strawberry yield intelligent prediction system based on RS485 sensing and UAV vision includes: an environmental perception module, a visual acquisition module, and an intelligent analysis module. The environmental sensing module includes a sensor array and a microcontroller connected via an RS485 bus, used to collect multi-parameter environmental data inside the greenhouse or shed; The visual acquisition module includes an autonomous cruise drone equipped with a high-definition camera, used to cruise along a preset route to collect image data of the strawberry planting area; The intelligent analysis module is communicatively connected to the environmental perception module and the visual acquisition module. It receives and processes multi-parameter environmental data from the environmental perception module and image data from the visual acquisition module, and performs the following steps to obtain the strawberry yield prediction result: S1: Based on the image data, use a visual recognition model to identify strawberry fruits of different ripeness and count their number. Based on the principle of monocular visual geometry, inversely deduce the physical size of the fruit to obtain the fruit state characteristics. The multi-parameter environmental data is received and processed, and together with the fruit state features, a multi-dimensional feature vector is formed. S2: Construct a multi-factor dynamic yield estimation model based on LLM, fill each feature value in the multi-dimensional feature vector according to the preset structured natural language prompt template, generate input prompt information, and input the input prompt information into the multi-factor dynamic yield estimation model based on LLM. S3: Introducing the RAG architecture, and combining the RAG architecture and the LLM-driven multi-factor yield dynamic estimation model, the model performs inference based on the input prompts, and generates and outputs results containing the predicted yield of strawberries within a preset future time period, confidence intervals, and the basis for the prediction.

[0008] Furthermore, images of strawberry fruits containing ripe, semi-ripe, and unripe fruits were collected and used to train the YOLOv8n model, resulting in a visual recognition model that identifies the three levels of ripeness.

[0009] Furthermore, based on the geometric principles of monocular vision, the physical size of the fruit is deduced, including: Based on the geometric principles of monocular vision, the true physical size of the fruit is calculated using the pixel dimensions of the fruit's outline in the image and the pixel distance to a reference object with a known actual physical length in the image, as shown in the following formula.

[0010] in, For the actual size of the fruit, The average pixel diameter of the fruit outline in the image. Given a known reference physical length, i.e., a known line spacing. The reference physical length corresponds to the pixel distance in the image.

[0011] Furthermore, receiving and processing the multi-parameter environmental data includes: Receive multi-parameter environmental data and obtain the average value of the multi-parameter environmental data over a historical set time period. The multi-parameter environmental data includes air parameter data and soil parameter data.

[0012] Furthermore, an LLM-driven multi-factor dynamic yield estimation model is constructed, including: Constructing a training dataset: Extract environmental parameters, fruit status, actual yield, and expert notes from an agricultural knowledge base containing strawberry growth records, and construct instruction fine-tuning samples. The instruction fine-tuning samples are in the format of {instruction, input, output} triples. Model fine-tuning: The training dataset is input into the LLM model, and the LoRA low-rank adaptive method is used. Based on the LLM model, the basic model parameters are frozen, and only the newly added low-rank adapter parameters are trained, so that the LLM model can predict the yield based on environmental parameters and fruit status. Model evaluation: Evaluate the model output on the reserved test set until the relative error between its predicted yield and the actual yield is less than a preset threshold, and the readability and agricultural rationality of the output text are evaluated by experts and meet the standards.

[0013] Furthermore, combining the RAG architecture and LLM-driven multi-factor yield dynamic estimation model, inference is performed based on input prompts to generate and output results containing the predicted harvestable strawberry yield, confidence interval, and prediction basis for a preset future time period, including: Based on the current multidimensional feature vector, retrieve K pieces of knowledge related to the current environmental parameters and fruit status from the agricultural knowledge base; The retrieved knowledge and historical records are concatenated with input prompts composed of multi-dimensional feature vectors to form the enhanced input context of LLM. The enhanced input context is fed into the LLM-driven multi-factor dynamic yield estimation model for inference, generating output results that include yield forecasts, confidence intervals, and forecasting rationale.

[0014] Furthermore, the agricultural knowledge base is a vector database. It performs retrieval based on the cosine similarity between the multidimensional feature vector and the feature vector recorded in the agricultural knowledge base. The cosine similarity is sorted from high to low, and the knowledge corresponding to the top K cosine similarity is selected as knowledge related to the current environment and the fruit state.

[0015] Beneficial Effects: This invention provides an intelligent strawberry yield prediction system based on RS485 sensing and UAV vision, which has the following beneficial effects: 1. Improved Yield Prediction Accuracy of the Invention: This invention does not simply rely on fruit counting, but innovatively constructs a multi-dimensional feature fusion model. Specifically, the system combines environmental stress factors with visual perception features. Through deep reasoning of feature vectors using LLM, the model can dynamically reflect the impact of environmental conditions on fruit growth and weight, thus achieving a fundamental improvement from "counting" to "weight estimation." 2. This invention exhibits strong data acquisition robustness and reliable transmission even in complex electromagnetic environments: Addressing the strong electromagnetic interference caused by metal structures within greenhouses and polytunnels, this invention abandons common wireless solutions such as ZigBee and Wi-Fi, innovatively employing a wired bus network based on RS485 differential signals. Each sensor node has a unique address code, supporting plug-and-play functionality. This design enables a data packet loss rate of less than 0.1% even with a bus deployment up to 1200 meters long, ensuring continuous, stable, and high-precision transmission of environmental parameters (temperature, humidity, EC, CO2, etc.), laying a reliable data foundation for subsequent accurate analysis. 3. The visual recognition of this invention has comprehensive coverage, effectively overcoming the problem of plant shading.

[0016] To address the issue of fruit shading caused by short strawberry plants and dense foliage, this invention employs an autonomous drone to collect images from an aerial perspective, achieving blind-spot-free, full-coverage scanning of the planting area. Combined with a specially trained YOLOv8n model, the system achieves an overall strawberry fruit recognition rate of 95.2% and can accurately distinguish between ripe, semi-ripe, and unripe states, fundamentally solving the statistical distortion problems caused by the single perspective and severe shading of fixed cameras. 4. The decision-making process is transparent and interpretable, and the output combines predictive values ​​with agronomical evidence: This invention abandons the "black box" neural network model, instead employing a Large Language Model (LLM) for reasoning. The system not only outputs predicted harvestable yields and confidence intervals for a certain number of days ahead, but more importantly, it simultaneously generates natural language descriptions of the decision-making basis. This basis stems from the correlation analysis of real-time data and an agricultural knowledge base, ensuring that the prediction results are not merely untraceable numbers, but rather decision support information with clear agronomic causal logic. This significantly enhances user trust in the system and its practical guidance value. Attached Figure Description

[0017] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 System block diagram of the intelligent strawberry yield prediction system based on RS485 sensing and UAV vision provided by the present invention; Figure 2 The overall system architecture diagram of the intelligent strawberry yield prediction system based on RS485 sensing and UAV vision provided by the present invention; Figure 3 The LLM output flow chart combined with RAG is provided for the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] This embodiment provides an intelligent strawberry yield prediction system based on RS485 sensing and UAV vision, such as... Figure 1 As shown, it includes: an environmental perception module, a visual acquisition module, and an intelligent analysis module; The environmental sensing module includes a sensor array and a microcontroller connected via an RS485 bus, used to collect multi-parameter environmental data inside the greenhouse or shed; Specifically, such as Figure 2 As shown, in the hardware part, this invention uses an RS485 differential signal acquisition sensor, with each 100m... 2 Deploy one RS485 sensor node with a single bus to connect eight types of sensors: Air: Temperature (±0.3℃), Humidity (±2%RH), CO2 (0–2000ppm), Light Intensity (0–200klux); Soil: temperature (±0.5℃), humidity (0–100% VWC), electrical conductivity EC (0–20mS / cm), pH (3–9), nitrogen, phosphorus and potassium ion concentration (mg / kg level); sensor address encoding, supporting plug-and-play and fault self-diagnosis; An air sensor array and a soil sensor array deployed in the greenhouse are connected to a microcontroller via an RS485 differential signal bus. This method ensures that the packet loss rate of multi-parameter environmental data (temperature, humidity, EC, CO2, etc.) is less than 0.1% even in environments with strong electromagnetic interference, and that the data is uploaded in real time. The visual acquisition module includes an autonomous cruise drone equipped with a high-definition camera, used to cruise along a preset route to collect image data of the strawberry growing area; the drone takes off automatically every morning to avoid dew interference; Specifically, in this plan, the drone automatically cruises and takes pictures along a preset route (altitude 1.5m, speed 0.5m / s); like Figure 2 As shown, the software component is the intelligent analysis module, which is mounted on an edge server and communicates with the environmental perception module and the visual acquisition module. It receives and processes multi-parameter environmental data from the environmental perception module and image data from the visual acquisition module, and performs the following steps to obtain the strawberry yield prediction result: S1: Based on the image data, a visual recognition model is used to identify strawberry fruits of different ripeness and count their quantity. Based on the principle of monocular visual geometry, the physical size of the fruits is deduced to obtain the fruit state characteristics: The visual recognition model was developed by collecting images of strawberry fruits, including ripe, semi-ripe, and unripe fruits, and training the YOLOv8n model to obtain a visual recognition model that can identify the three levels of ripeness. The specific training process is as follows: Strawberry images with three categories of recognition were collected. The training set contained 7,960 images, the test set contained 587 images, and 9,243 labeled images were used for fine-tuning, covering complex scenes with different lighting, occlusion, and angles to improve recognition accuracy. Among them, ripe strawberries accounted for ≥80% red, semi-ripe strawberries were red and white, each accounting for 50%, and unripe strawberries were mainly green. Based on the geometric principles of monocular vision, the physical size of the fruit is deduced to obtain its state characteristics, including: Based on the geometric principles of monocular vision, the true physical size of the fruit is calculated using the pixel dimensions of the fruit's outline in the image and the pixel distance to a reference object with a known actual physical length in the image, as shown in the following formula.

[0021] in, The actual size of the fruit is represented by its diameter in this design. The average pixel diameter of the fruit outline in the image. Given a known reference physical length, i.e., a known line spacing. The reference physical length corresponds to the pixel distance in the image; Specifically, the known reference physical length in this invention is the actual distance between the centers of two adjacent rows of strawberries; Receive and process the multi-parameter environmental data, and combine it with the fruit state features to form a multi-dimensional feature vector: Receive multi-parameter environmental data and obtain the average value of the multi-parameter environmental data within a historical set time period; Specifically, feature vector The data consists of three categories: fruit visual characteristics, soil environmental parameters, and air microclimate parameters. All values ​​have been normalized or unit-standardized, as shown in the table below.

[0022] All environmental parameters are taken as the moving average of the past 72 hours to reflect recent growth conditions; if a sensor fails, the corresponding feature is set to the historical average of that parameter and marked as "data interpolation"; The total number of fruits is: N_total = N_ripe + N_half + N_unripe; S2: Construct a multi-factor dynamic yield estimation model based on LLM, fill each feature value in the multi-dimensional feature vector according to the preset structured natural language prompt template, generate input prompt information, and input the input prompt information into the multi-factor dynamic yield estimation model based on LLM. The construction of an LLM-driven multi-factor dynamic yield estimation model includes: Constructing the training dataset: Environmental parameters, fruit status, actual yield, and expert notes are extracted from an agricultural knowledge base containing strawberry growth records to construct instruction fine-tuning samples. The instruction fine-tuning samples are in the format of {instruction, input, output} triples. The data sources in this scheme are as follows: Agricultural knowledge base: Contains 500+ strawberry growth records, each record including: Date, greenhouse ID, variety, planting area; Environmental parameters (temperature, humidity, EC, light intensity, CO2, soil NPK, etc.); Fruit condition (number of mature fruits, number of semi-mature fruits, number of immature fruits, average diameter); Actual output (kg); Expert notes (such as "high EC levels hindered growth", "low temperature delayed ripening").

[0023] Historical Production Report: A daily production forecast report generated by the system, containing the forecast value, confidence interval, and environmental recommendations; Model fine-tuning: The training dataset is input into the LLM model, and the LoRA low-rank adaptive method is used. Based on the LLM model, the basic model parameters are frozen, and only the newly added low-rank adapter parameters are trained, so that the LLM model can predict the yield based on environmental parameters and fruit status. Model evaluation: Evaluate the model output on the reserved test set until the relative error between its predicted yield and the actual yield is less than a preset threshold, and the readability and agricultural rationality of the output text are evaluated by experts and meet the standards.

[0024] LLM selects a general open-source model. This solution does not restrict the specific type. LoRA fine-tuning is a common technique for model fine-tuning. Therefore, the specific process and steps of LoRA fine-tuning are not extended. Those skilled in the art know how to train. In this invention, the preset structured natural language prompt template is as follows: Structured: Organized according to the logic of "fruit condition → soil conditions → climate conditions"; Readability: Uses terminology familiar to agricultural technicians; Guiding: Explicitly require the output format (numerical value + confidence interval + basis); Contextual enhancement: Add prior knowledge of variety and growth period (configurable); S3: Introducing the RAG architecture, and combining the RAG architecture with an LLM-driven multi-factor yield dynamic estimation model, the system infers based on input prompts to generate and output results containing the predicted harvestable strawberry yield for a preset future time period, confidence intervals, and the basis for the prediction. Figure 3As shown, it includes: Based on the current multidimensional feature vector, retrieve K pieces of knowledge related to the current environmental parameters and fruit status from the agricultural knowledge base; The retrieved knowledge and historical records are concatenated with input prompts composed of multi-dimensional feature vectors to form the enhanced input context of LLM. The enhanced input context is fed into the LLM-driven multi-factor dynamic yield estimation model for inference, generating output results that include yield forecasts, confidence intervals, and forecasting basis. If strawberry cluster data is not obtained, an enhanced context is constructed using historical data and retrieved data. Specifically, RAG is a retrieval-enhanced generation. In the inference process of LLM in this invention, RAG is manifested as retrieving agricultural knowledge base, obtaining retrieved knowledge, and then concatenating the retrieved knowledge with the input prompt information input into LLM to obtain enhanced input context. The following is a sample Prompt template (supporting dynamic filling): You are a seasoned strawberry growing expert. Based on the following real-time monitoring data, please estimate the harvestable yield (in kilograms) of strawberries in this area over the next 7 days. Please answer strictly according to the following format: [Production Forecast]: <Value> kg (Confidence Interval: <Lower Limit> – <Upper Limit> kg) [Main Basis]: <No more than 3 key reasons> === Input Data === 1. Fruit condition: - Mature fruits: {N_ripe} - Semi-ripe fruit: {N_half} pieces (expected to ripen in 5–7 days) - Unripe fruit: {N_unripe} - Average single fruit diameter: {D_avg:.1f} cm 2. Soil environment (average of the past 3 days): - Temperature: {T_soil:.1f} ℃ - Humidity: {H_soil:.1f} % VWC - Electrical conductivity (EC): {EC:.2f} mS / cm - Total nitrogen, phosphorus, and potassium: {NPK_sum} mg / kg 3. Air microclimate (average over the past 3 days): - Temperature: {T_air:.1f} ℃ - Humidity: {H_air:.1f} % RH - Light intensity: {Light:.1f} klux - CO2 concentration: {CO2} ppm Note: The current variety is "Hongyan" and is in its peak fruiting period.

[0025] Example (after filling): You are a seasoned strawberry growing expert. Based on the following real-time monitoring data, please estimate the harvestable yield (in kilograms) of strawberries in this area over the next 7 days. Please answer strictly according to the following format: [Production Forecast]: 42.5 kg (Confidence Interval: 38.2–46.8 kg) [Main Basis]: - There are 120 mature fruits and 85 semi-mature fruits, indicating ample potential for conversion; - The soil EC value is high (1.85 mS / cm), which may slightly inhibit swelling; - Recent abundant sunlight (18.3 klux) is conducive to sugar accumulation and coloring.

[0026] === Input Data === 1. Fruit condition: - 120 ripe fruits - Semi-ripe fruit: 85 pieces (expected to ripen in 5–7 days) - Unripe fruit: 60 - Average fruit diameter: 3.2 cm 2. Soil environment (average of the past 3 days): - Temperature: 19.5 ℃ - Humidity: 68.3% VWC - Electrical conductivity (EC): 1.85 mS / cm - Total nitrogen, phosphorus, and potassium: 420 mg / kg 3. Air microclimate (average over the past 3 days): - Temperature: 22.1 ℃ - Humidity: 75.6% RH - Illumination intensity: 18.3 klux - CO2 concentration: 850 ppm Note: The current variety is "Hongyan" and is in its peak fruiting period.

[0027] The system generates a daily "Yield Forecast Report" and pushes it to the grower's app, which includes: Harvestable yield (kg) Optimal harvest window (3–7 days from now) Environmental optimization recommendations (e.g., "reduce EC to 1.2 mS / cm to promote expansion").

[0028] This scheme uses a combination of RAG and LLM methods to predict strawberry yield, which has the following advantages: Avoid "illusions": Constrain LLM outputs through structured inputs to prevent unfounded guessing; Highly explainable: Experts can trace the environmental or physiological causes of yield changes; Flexible expansion: Simply modify the Prompt template to adapt to other berries such as blueberries and tomatoes; Lightweight deployment: The Prompt length is controlled within 512 tokens, and it is compatible with local small models such as Qwen and Llama3-8B.

[0029] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A strawberry yield intelligent estimation system based on RS485 sensing and unmanned aerial vehicle vision, characterized in that, include: Environmental perception module, visual acquisition module, and intelligent analysis module; The environmental sensing module includes a sensor array and a microcontroller connected via an RS485 bus, used to collect multi-parameter environmental data inside the greenhouse or shed; The visual acquisition module includes an autonomous cruise drone equipped with a high-definition camera, used to cruise along a preset route to collect image data of the strawberry planting area; The intelligent analysis module is communicatively connected to the environmental perception module and the visual acquisition module. It receives and processes multi-parameter environmental data from the environmental perception module and image data from the visual acquisition module, and performs the following steps to obtain the strawberry yield prediction result: S1: Based on the image data, use a visual recognition model to identify strawberry fruits of different ripeness and count their number. Based on the principle of monocular visual geometry, inversely deduce the physical size of the fruit to obtain the fruit state characteristics. The multi-parameter environmental data is received and processed, and together with the fruit state features, a multi-dimensional feature vector is formed. S2: Construct an LLM-driven multi-factor dynamic yield estimation model, fill each feature value in the multi-dimensional feature vector according to a preset structured natural language prompt template to generate input prompt information, and input the input prompt information into the LLM-driven multi-factor dynamic yield estimation model. S3: Introducing the RAG architecture, and combining the RAG architecture and the LLM-driven multi-factor yield dynamic estimation model, the model performs inference based on the input prompts, and generates and outputs results containing the predicted yield of strawberries within a preset future time period, confidence intervals, and the basis for the prediction.

2. The intelligent strawberry yield estimation system based on RS485 sensing and UAV vision of claim 1, wherein Images of strawberry fruits, including ripe, semi-ripe, and unripe fruits, were collected and used to train the YOLOv8n model, resulting in a visual recognition model that identifies the three levels of ripeness. 3.The RS485 sensor and unmanned aerial vehicle vision-based strawberry yield intelligent estimation system according to claim 1, characterized in that, Based on the geometric principles of monocular vision, the physical dimensions of the fruit are deduced, including: Based on the geometric principles of monocular vision, the true physical size of the fruit is calculated using the pixel dimensions of the fruit's outline in the image and the pixel distance to a reference object with a known actual physical length in the image, as shown in the following formula. wherein, is the real size of the fruit, is the average pixel diameter of the fruit profile in the image, is the known reference physical length, i.e. the known row distance, is the corresponding pixel distance of the reference physical length in the image.

4. The system for intelligent estimation of strawberry yield based on RS485 sensing and UAV vision according to claim 1, characterized in that, Receiving and processing the multi-parameter environmental data includes: Receive multi-parameter environmental data and obtain the average value of the multi-parameter environmental data over a historical set time period. The multi-parameter environmental data includes air parameter data and soil parameter data.

5. The intelligent strawberry yield prediction system based on RS485 sensing and UAV vision according to claim 1, characterized in that, Constructing an LLM-driven multi-factor dynamic yield estimation model, including: Constructing a training dataset: Extract environmental parameters, fruit status, actual yield, and expert notes from an agricultural knowledge base containing strawberry growth records, and construct instruction fine-tuning samples. The instruction fine-tuning samples are in the format of {instruction, input, output} triples. Model fine-tuning: The training dataset is input into the LLM model, and the LoRA low-rank adaptive method is used. Based on the LLM model, the basic model parameters are frozen, and only the newly added low-rank adapter parameters are trained, so that the LLM model can predict the yield based on environmental parameters and fruit status. Model evaluation: Evaluate the model output on the reserved test set until the relative error between its predicted yield and the actual yield is less than a preset threshold, and the readability and agricultural rationality of the output text are evaluated by experts and meet the standards.

6. The intelligent strawberry yield prediction system based on RS485 sensing and UAV vision according to claim 1, characterized in that, Combining the RAG architecture and LLM-driven multi-factor yield dynamic estimation model, the model infers based on input prompts, generating and outputting results including predicted strawberry harvest yield, confidence intervals, and prediction basis for a preset future time period, including: Based on the current multidimensional feature vector, retrieve K pieces of knowledge related to the current environmental parameters and fruit status from the agricultural knowledge base; The retrieved knowledge and historical records are concatenated with input prompts composed of multi-dimensional feature vectors to form the enhanced input context of LLM. The enhanced input context is fed into the LLM-driven multi-factor dynamic yield estimation model for inference, generating output results that include yield forecasts, confidence intervals, and forecasting rationale.

7. The intelligent strawberry yield prediction system based on RS485 sensing and UAV vision according to claim 6, characterized in that, The agricultural knowledge base is a vector database. It is used to retrieve information based on the cosine similarity between multidimensional feature vectors and feature vectors recorded in the agricultural knowledge base. The cosine similarity is sorted from high to low, and the knowledge corresponding to the top K cosine similarity is selected as knowledge related to the current environment and the fruit state.

Images