Time prior constraint and multi-modal visual fusion based sugarcane developmental stage recognition method

Abstract

The application relates to a sugarcane development period recognition method based on time prior constraint and multimodal vision fusion, which comprises the following steps: step 1, extracting the shooting time and preliminarily positioning the growth period; step 2, judging whether the growth transition period is reached and distributing the processing route; step 3, extracting the direct visual appearance features of the sugarcane image; step 4, starting the multi-path AI model to perform image depth analysis and prediction; and step 5, comprehensively voting and mapping the final result. The application has the beneficial effects of: (1) error time domain truncation and phase locking mechanism; (2) low-label high-generalization recognition ability; and (3) explainable white-box arbitration and spatialized output.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and computer vision, and to a method for image recognition of the entire life cycle development stages of sugarcane that combines phenological prior knowledge, large multimodal semantic model (CLIP), and visual image feature extraction technology. Background Technology

[0002] With the development of smart agriculture, using computer vision to monitor the entire life cycle of crops such as sugarcane has become a research hotspot. However, existing sugarcane development stage identification technologies mainly have the following shortcomings:

[0003] (1) The unreliability of single data-driven approaches: Traditional deep learning models are mostly in a black box state, and are prone to serious illusory misjudgments that deviate from common sense under atypical climatic conditions or cross-seasonal scenarios.

[0004] (2) Poor adaptability to environmental interference: Recognition systems that rely solely on RGB visual features have low robustness when faced with different lighting conditions, shooting distances, and complex field backgrounds.

[0005] (3) Lack of decision interpretability: Existing algorithms often only output the final probability of the development stage, which cannot provide a traceability link that conforms to agricultural logic and is difficult to meet the trust requirements of agricultural production for intelligent decision-making systems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a sugarcane development period identification method based on time prior constraints and multimodal visual fusion, so as to solve the problems of poor model interpretability, easy generation of cross-seasonal illusion misjudgment and insufficient generalization ability in complex environments in the prior art.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for identifying the sugarcane development stage by fusing time prior constraints and multimodal vision, comprising the following steps:

[0008] Step 1: Analyze the time information of the image to be tested and convert it into an ordinal number within the year. Combine the sugarcane baseline development period time interval to calculate the minimum distance between the current date and the phenological boundary. Based on the 20-day threshold and the given formula, determine whether the crop is in a stable period or a boundary period, and generate the corresponding set of legal candidate stages.

[0009] Step 2: For the boundary period image determined in Step 1, use OpenCV computer vision technology combined with color space, Fourier transform, and edge detection technology to extract physical visual features such as vegetation color ratio, texture direction, coverage, and edge density to form low-level visual feature data.

[0010] Step 3: In the boundary period scenario, based on the candidate range defined in Step 1 and the visual features extracted in Step 2, three types of algorithms are run in parallel: K nearest neighbor vector retrieval, CLIP multimodal image-text matching, and visual heuristic rule determination. The time reference results are retained simultaneously, and multiple preliminary recognition results are output.

[0011] Step 4: Using the asymmetric legal candidate stage set generated in Step 1 as the verification standard, filter the output results of the multi-way algorithm in Step 3, remove invalid results that exceed the candidate range, and retain compliant results to enter the voting stage;

[0012] Step 5: Assign fixed weights to the valid results retained in Step 4, perform weighted cumulative voting, and select the developmental stage with the highest cumulative score as the final recognition conclusion for a single image;

[0013] Step 6: Summarize the single-image recognition results of all images from the same monitoring site from Step 5, filter out outliers using the majority voting method, and output the overall dominant developmental period of the site.

[0014] Step 7: Based on the single-map results from Step 5 and the site aggregation results from Step 6, bind the single-map inference logic chain, site results, geographic coordinates, and other information to generate a Shapefile spatial vector file with a complete traceability chain, thus completing the spatialized output of the results.

[0015] In some possible implementations, step 1 specifically involves:

[0016] Step 1.1: Extract the timestamp from the filename of the image to be tested, parse the year, month, and day, and convert them into ordinal numbers within the year. Preset upper and lower boundaries for the five developmental stages of sugarcane: sowing, emergence, tillering, stem elongation, and technological maturity. , and the time intervals for each stage;

[0017] Step 1.2, calculate the year ordinal number of the current date. Minimum distance from the upper and lower boundaries of the baseline developmental period Determine the state and generate a set of valid candidates according to the following formula. :

[0018] ;

[0019] If the minimum distance is greater than 20 days: it is determined to be a stable developmental period, and the valid candidate set only includes the current target developmental period; if the minimum distance is less than or equal to 20 days: it is determined to be a developmental boundary period, and the valid candidate set includes the current target developmental period and adjacent developmental periods.

[0020] Step 1.3: During the stable phase, the current developmental phase result is directly output, and the process is terminated; during the boundary phase, the valid candidate set is retained, and the process proceeds to the next visual analysis step.

[0021] Step 1 constructs time-gating rules based on prior knowledge of sugarcane phenology, using a 20-day threshold to divide developmental states, and rigidly constrains the candidate range through mathematical formulas. The beneficial effects are: zero-error locking during the stable period; no need to call visual algorithms during the stable period, reducing computational consumption; and only retaining adjacent stages as candidates during the boundary period, thus eliminating misjudgments across seasons and stages from the ground up.

[0022] In some possible implementations, step 2 specifically involves:

[0023] Step 2.1, Statistical analysis of pixel proportion based on HSV color space: Define the green range [25, 40, 40]-[95,255, 255], calculate the proportion of green vegetation; extract the region of interest in the image, and statistically analyze the proportion of yellowish-brown pixels to obtain the proportion of withered yellow leaves;

[0024] Step 2.2: Perform a two-dimensional Fourier transform on the image to perform frequency domain analysis, calculate the ratio of spectral energy in the vertical direction to that in the horizontal direction, and obtain the plant verticality / texture direction features;

[0025] Step 2.3: Calculate the vegetation coverage rate by combining the proportion of green vegetation and the proportion of withered leaves, extract edge density features by Canny edge detection, and integrate all features to form a physical visual feature set.

[0026] Step 2 establishes a quantitative feature system corresponding to different growth forms of sugarcane based on the physical properties of the underlying pixels, frequency domain, and edges of the image. The beneficial effects are: using objective physical features as the basis for rule determination compensates for the randomness of pure AI semantic recognition, provides basic constraints for subsequent rule reasoning, and improves the robustness of recognition under complex field lighting and shooting environments.

[0027] In some possible implementations, step 3 specifically involves:

[0028] Step 3.1, K-nearest neighbor algorithm retrieval: The CLIP model is used to extract feature vectors of the image to be tested and the reference image library, and the vector similarity is calculated to complete the nearest neighbor retrieval and output the developmental stage recognition result; the CLIP model is combined to identify the near / far view of the image, and the corresponding view reference image library is dynamically called by the lazy loading mechanism to avoid image library redundancy;

[0029] Step 3.2, CLIP Multimodal Image-Text Matching: Load the preset English text prompts corresponding to different sugarcane development stages and shooting angles, map the image tensor and text features to the same 512-dimensional semantic space, and output the image-text recognition results through zero-shot cosine similarity matching;

[0030] Step 3.3: Based on the physical features such as green ratio, yellow ratio, and frequency domain texture extracted in Step 2, and combined with the preset experience threshold rule engine, the recognition result is obtained through logical reasoning.

[0031] Step 3.4: Use the time inference results from Step 1 as the baseline time result and summarize the preliminary identification results from the four paths.

[0032] Step 3 integrates three heterogeneous algorithms: few-shot KNN retrieval, zero-shot multimodal large-scale model, and traditional machine vision rules, forming a complementary and competitive inference architecture. The benefits are: it eliminates the dependence on large-scale labeled data, achieving high generalization recognition with low labeling; the large-scale model handles high-dimensional semantic understanding, while traditional vision rules handle physical common sense correction, comprehensively improving the accuracy of complex scene recognition.

[0033] In some possible implementations, step 4 specifically involves:

[0034] Step 4.1: Read the list of legal candidate stages for the current boundary period determined in Step 1, and use it as the only valid result range;

[0035] Step 4.2: Verify the recognition results of the three algorithms in Step 3: K-nearest neighbor algorithm retrieval, CLIP, and visual rule. If the result belongs to the valid candidate set, it is retained; if it exceeds the candidate range, it is judged as a logical out-of-bounds error and is directly eliminated, and does not participate in subsequent voting.

[0036] Step 4.3: The time baseline backup result is considered a valid candidate set by default and is directly retained.

[0037] Step 4 uses prior phenological time as a hard logical constraint to filter out abnormal results generated by the AI ​​algorithm that violate agricultural time patterns. The beneficial effects are: further eliminating misjudgments across stages and seasons, ensuring that all decisions involved conform to sugarcane growth patterns, and strengthening the rationality of the identification logic.

[0038] In some possible implementations, step 5 specifically involves:

[0039] Step 5.1: Preset fixed weights for each algorithm module: K-nearest neighbor algorithm retrieval weight 1.5, CLIP multimodal matching weight 1.2, visual heuristic rule weight 1.0, and time baseline minimum weight 0.8;

[0040] Step 5.2: For each type of recognition result that passes the verification, calculate the total score for each developmental stage by accumulating the corresponding weights;

[0041] Step 5.3: Compare the cumulative scores of all candidate developmental stages and determine the stage with the highest score as the final recognition result of a single image.

[0042] Step 5 assigns weights based on the credibility differences of different algorithms and fuses multi-source results through weighted voting to mitigate the shortcomings of a single algorithm. The beneficial effects are: improved recognition stability in ambiguous scenarios such as the boundary period; and transparency in the weighted voting process, achieving a white-box recognition workflow with interpretability.

[0043] In some possible implementations, step 6 specifically involves:

[0044] Step 6.1: Collect the final recognition results of all images from the same site and count the frequency of occurrence of results at each developmental stage;

[0045] Step 6.2: Using the majority voting rule, outlier results with extremely low frequency are removed, and the development period with the highest frequency is selected as the dominant development period of the site.

[0046] Step 6 utilizes the redundancy of results from multiple frames of images from the same site to offset random errors caused by the shooting angle and lighting of a single image through statistical voting. The beneficial effects are: improved overall stability of sugarcane development stage identification at the regional scale, and reduced local misjudgments caused by single-sample interference.

[0047] In some possible implementations, step 7 specifically involves:

[0048] Step 7.1: Organize logical tracing information: record the time window determination results, boundary period markers, original results from four paths: K nearest neighbor algorithm retrieval / CLIP / visual rules / time benchmark, weighted voting count details, and single graph final conclusions to form a complete reasoning chain;

[0049] Step 7.2: Combine the site ID corresponding to the image with the geographic coordinates in the WGS84 coordinate system, and use GeoPandas to construct a spatial data object;

[0050] Step 7.3: Unify the single image recognition results, logical source chain, site dominant development period, and geographic information, and export them as an ESRI Shapefile spatial vector file.

[0051] Step 7 combines the AI's reasoning process and judgment results with geospatial information to standardize the output of agricultural spatial data in a standardized format. The benefits include: the recognition results are fully traceable and verifiable, solving the black-box problem of traditional deep learning; the spatialized files can be directly connected to a smart agriculture GIS platform, supporting regional agricultural spatiotemporal analysis and meeting the credibility requirements of intelligent agricultural decision-making.

[0052] The beneficial effects of this invention are:

[0053] (1) Error time domain truncation and stage locking mechanism

[0054] This invention pioneers a time-gating mechanism that combines zero-error locking during the stable period with an asymmetric candidate set for the boundary period. This is based on the relative number of days between time and the sugarcane phenological window. The surface determination rule is shown in formula (1):

[0055] (1)

[0056] in, The ordinal number of the current date within the year. and These are the upper and lower boundaries of the current baseline developmental stage. This is a valid candidate set. During the stable development period, which accounts for most of the year, when the distance from the boundary is more than 20 days, unnecessary visual interference is bypassed, and the error is directly locked to 0. During the developmental boundary period, the candidate set S generated by the dynamic time bias places the physical stage of the maximum possible error within an adjacent phenological period, thus completely eliminating outrageous misjudgments across seasons and stages from the bottom layer of the algorithm.

[0057] (2) Low-label, high-generalization recognition capability

[0058] This method abandons the single, black-box deep learning model that relies on a large amount of labeled data, and innovatively adopts a multi-source algorithm competition architecture. It integrates the multimodal large-scale model CLIP with dynamic cue word engineering, achieving zero-shot recognition through multimodal image-text matching, while utilizing the KNN algorithm for few-shot vector retrieval from a reference image library, and extracting physically heuristic visual rules for greenness, yellowness, and frequency domain verticality. This method possesses both the high-dimensional semantic generalization capability of a large model and the physical foundation of traditional vision, maintaining extremely high recognition robustness even in the absence of large-scale new site labeled data.

[0059] (3) Explainable white-box arbitration and spatialized output

[0060] This method not only achieves accurate automatic identification but also ensures the high interpretability required for agricultural decision-making. By setting up a weighted arbitration module (e.g., KNN weight 1.5, multimodal 1.2, physical vision 1.0, and time minimum 0.8), each diagnostic process is transformed into a visualized multi-way weighted voting game. The final output baseline GIS spatial vector file (.shp format) fully encapsulates the logical traceability feature string (e.g., Time[month.day]->KNN(period)->MM(period)->Vis(period)->Votes{vote count}) of a single map dimension. This gives the identification results a completely transparent logical white-box characteristic, facilitating manual review and providing a highly reliable data foundation for spatiotemporal big data analysis of smart agriculture platforms. Attached Figure Description

[0061] Figure 1 This is a flowchart of the algorithm of the present invention. Detailed Implementation

[0062] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0063] This invention provides a method for identifying crop development stages based on temporal prior constraints and multimodal visual fusion. The algorithm flow is as follows: Figure 1 As shown, it mainly includes the following steps:

[0064] Step 1: Extract the shooting time and preliminarily determine the growth cycle. By consulting a calendar, the approximate growth range is delineated, mainly consisting of two steps: time information analysis and phenological window delineation. First, the timestamp in the filename of the image to be tested is analyzed and mapped to an ordinal value within the year. Then, based on this ordinal value, the baseline sugarcane development period is determined, and the number of days between the current date and the upper and lower boundaries of the baseline development period is calculated.

[0065] Step Two: Determine if the crop is in a growth transition period and allocate processing routes. Using time-gated triage, determine whether the crop is in a "stable period" or a "boundary period," deciding whether to initiate AI image recognition and generate a valid candidate set. Based on the distance calculated in Step One, determine if the distance is within the baseline developmental period threshold (20 days). If the distance is greater than the threshold, the current date is determined to be in the stable developmental period, skipping visual reasoning and directly outputting the corresponding developmental stage; if the distance is less than or equal to the threshold, it is determined to be in the developmental boundary period, entering the multi-source algorithm competition module. Simultaneously, based on time bias (whether the current time is closer to the previous or next stage), a constrained asymmetric valid candidate stage set is dynamically generated, defining a reasonable range of options for subsequent AI analysis.

[0066] Step 3: Extract intuitive appearance features from sugarcane images. Using physical visual feature extraction methods, basic identification is performed on images in the transition period to quantify the physical state of the crop. This primarily utilizes OpenCV computer vision technology to perform pixel-level and frequency domain analysis on the test images at the developmental boundary, extracting the greenness ratio (HSV color space filtering), yellowing ratio, total vegetation coverage, plant verticality features (frequency domain energy ratio based on Fourier transform), and edge density (Canny edge detection).

[0067] Step 4: Launch multi-path AI models for image depth analysis and prediction. Three different AI algorithms are used for independent inference, each providing its own inference results. For the test image in the transition period, its CLIP visual tensor is extracted, and a three-path inference algorithm competition mechanism is launched. 1) The K-nearest neighbor (KNN) algorithm is launched to retrieve the reference vector library, and the features of the test image are compared with the constructed developmental period vector library for click similarity retrieval; 2) Multimodal image-text matching is launched, and the zero-shot image-text feature matching mechanism of the CLIP large model is used to perform zero-shot probability matching between the tensor of the test image and the text prompt words "Prompt" of the preset developmental period detail features; 3) Visual heuristic rule judgment is performed, based on the physical visual features extracted in Step 3, combined with the preset empirical threshold rule engine for inference judgment.

[0068] Step 5: Integrated Voting Decision and Final Result Map Generation. This step primarily achieves interpretable spatialized output through weighted arbitration and group decision-making of multiple algorithm results. The results generated by the three-way competitive algorithm are validated for legality, eliminating inference results not included in the candidate stage combination. The validated algorithm inference results are weighted and accumulated with the time baseline vote, and the developmental stage with the highest score is taken as the final inference conclusion for a single map. Subsequently, invalid outliers are identified, and a majority vote is conducted on all image conclusions at the same site to filter out outliers, obtaining the dominant developmental stage of the site. The result is then exported as a Shapefile spatial vector file containing geographic coordinate information.

[0069] This invention, based on time prior constraints, visual feature analysis, CLIP multimodal semantic matching, and a hybrid weighted arbitration mechanism, achieves intelligent identification and spatialized output of sugarcane development stages. To clarify the objectives, technical solutions, and advantages of this invention, the following detailed explanation is provided in conjunction with a specific sugarcane monitoring scenario.

[0070] (1) Mechanism for preliminary delineation and boundary determination of phenological windows based on time scale

[0071] This invention utilizes time prior constraints and boundary period determination logic to initially locate the developmental cycle. The date information in the input image file name is parsed to extract the year, month, and day fields, and the date is converted into the number of days in a year for continuous time calculations across months. Based on monitoring experience from agricultural stations, a baseline time segmentation rule for sugarcane growth stages was established, including the sowing period, seedling emergence period, tillering period, stem elongation period, and technological maturity period, as shown in Table 1.

[0072] Table 1. Time Segmentation Rules for Sugarcane Growth Stages

[0073]

[0074] When calculating the distance between the current date and the boundary dates of each baseline developmental stage, a threshold of 20 days is set for boundary determination. If the current date is more than 20 days away from the boundary of an adjacent stage, the current sample is determined to be in a stable developmental stage, and the time inference result is directly used as the final output. If the distance is less than or equal to 20 days, it is determined to be a developmental stage boundary state, and the multi-source algorithm competition and weighted arbitration module is activated. Only adjacent stages are allowed as candidate results, thereby effectively reducing cross-stage misjudgments.

[0075] For example, when the image date is May 20, the system identifies it as being in the "tillering stage," but since it is less than 20 days away from the May 31 boundary, the "tillering stage" and the "stem elongation stage" are allowed to participate in the subsequent competitive determination.

[0076] (2) Multidimensional physical feature quantization and CLIP multimodal semantic feature extraction

[0077] After entering the transitional period of multi-source algorithm competition and weighted arbitration, the first step is to extract visual features from the image. As shown in Table 2, this specifically includes:

[0078] Table 2 Key Visual Features

[0079]

[0080] The proportion of green vegetation was obtained by statistical analysis of the green threshold region in the HSV color space; the proportion of withered leaves was obtained by extraction of the yellowish-brown region; and the texture direction features were calculated by analyzing the ratio of vertical texture energy to horizontal texture energy in the image using two-dimensional Fourier transform, which is used to reflect the longitudinal growth characteristics of sugarcane stalks.

[0081] Secondly, the CLIP visual language model is used to extract image feature vectors, and English agricultural semantic prompts are employed for multimodal matching. This maps image pixel information and agricultural meteorological text descriptions to the same high-dimensional vector space (512-dimensional), achieving cross-modal matching between images and text, and between images. Leveraging the cutting-edge concept of Prompt Tuning, semantic descriptions of sugarcane phenological characteristics during developmental stages (Prompts) are constructed, activating agricultural knowledge in the pre-trained model through natural language descriptions.

[0082] 1) Sowing period: "Soil plowed, sugarcane seeds visible on ground"

[0083] 2) Emergence stage: "Cone-shaped buds emerging from the ground, small green shoots about 2cm"

[0084] 3) Tillering stage: "Sugarcane buds germinating at base, side shoots clustering, bushy" (buds sprout at the base, side branches grow in clusters, forming a shrub-like shape);

[0085] 4) Stem elongation period: "Sugarcane stem internodes elongating, tall plants, mainstem visible"

[0086] 5) Processing maturity stage: "Yellow withered leaves, dry and smooth stem skin, mature sugarcane field".

[0087] Meanwhile, to address the feature differences caused by varying shooting distances, two types of perspective cue words were constructed: close-up and distant views.

[0088] 1) Close-up shot: “close-up detail of sugarcane leaves”;

[0089] 2) Vision: “wide angle landscape of sugarcane field”.

[0090] Finally, the CLIP model is used to calculate the cosine similarity between the image to be tested and the viewpoint cue words to achieve automatic viewpoint recognition. Based on the identified viewpoint category, feature vectors are extracted from the corresponding reference image library for KNN similarity retrieval, and combined with the CLIP image-text matching results to form multi-source candidate developmental stages.

[0091] Meanwhile, this method uses a validity check mechanism to retain only candidate developmental stages that fall within the allowed time window. For example, when the time window only allows the "tillering stage" and "stem elongation stage," even if the KNN retrieval finds the "process maturity stage," the result will be automatically discarded.

[0092] (3) Weighted arbitration and collaborative decision-making process based on multi-source algorithm competition

[0093] This method employs a multi-model weighted voting mechanism to fuse the results of different algorithms, thereby improving the recognition stability and robustness under complex boundary conditions. The main decision sources include the following four categories:

[0094] 1) KNN reference image library retrieval results: A reference image library is constructed through a lazy loading mechanism. The feature vector similarity between the image to be tested and the labeled samples in the image library is extracted, and the nearest neighbor sample category is selected.

[0095] 2) CLIP multimodal image-text matching calculates the image-text similarity between image features and pre-set agricultural knowledge descriptive terms.

[0096] 3) Visual heuristic rules, based on physical features for logical verification.

[0097] 4) The time benchmark is guaranteed, and the initial time window inference score is retained.

[0098] An accumulator is introduced into the fusion decision-making process to perform weighted voting on the effective results of the above algorithm. The weight parameters are set as shown in Table 3.

[0099] Table 3 Weight parameters of each algorithm module

[0100]

[0101] In practice, the time inference results are first added to the voting pool as a baseline, and then the AI ​​recognition results that pass the legality check are weighted accordingly. Finally, the candidate developmental period with the highest cumulative score is output as the final recognition result.

[0102] For example, in a sample from a certain boundary period:

[0103] 1) The time inference result is "tillering stage";

[0104] 2) The KNN result is "stem elongation period";

[0105] 3) The multimodal matching result is "stem elongation period";

[0106] 4) The visual rule result is "tillering stage".

[0107] The final vote count is as follows:

[0108] 1) Tillering stage: 0.8 + 1.0 = 1.8;

[0109] 2) Stem elongation period: 1.5 + 1.2 = 2.7.

[0110] Therefore, the system ultimately determined it to be the "stem elongation period".

[0111] (4) Logical source tracing encapsulation of recognition results and generation of spatial vector files

[0112] After image recognition is completed, the recognition results are bound to the input latitude and longitude information to generate spatial vector data containing station name, image name, time inference result, final recognition result and logical reasoning chain.

[0113] In the specific implementation, GeoPandas is used to construct GeoDataFrame objects, and ESRI Shapefile format files are generated based on the WGS84 coordinate system (EPSG:4326) to realize the GIS spatial expression of sugarcane development period identification results and subsequent agricultural analysis applications.

[0114] Finally, when multiple images exist at the same site, the system further counts the frequency of the final recognition results for each image and outputs the dominant reproductive stage as the overall judgment result for the region, thereby improving the stability of regional scale recognition.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying the sugarcane development stage by fusing temporal prior constraints and multimodal vision, characterized in that, The steps include the following: Step 1: Analyze the time information of the image to be tested and convert it into an ordinal number within the year. Combine the time interval of the sugarcane baseline development period to calculate the minimum distance between the current date and the phenological boundary. Based on the 20-day threshold and the given formula, determine whether the crop is in a stable period or a boundary period, and generate the corresponding set of legal candidate stages. Step 2: For the boundary period image determined in Step 1, use OpenCV computer vision technology combined with color space, Fourier transform, and edge detection technology to extract physical visual features such as vegetation color ratio, texture direction, coverage, and edge density to form low-level visual feature data. Step 3: In the boundary period scenario, based on the candidate range defined in Step 1 and the visual features extracted in Step 2, three types of algorithms are run in parallel: K nearest neighbor vector retrieval, CLIP multimodal image-text matching, and visual heuristic rule determination. The time reference results are retained simultaneously, and multiple preliminary recognition results are output. Step 4: Using the asymmetric legal candidate stage set generated in Step 1 as the verification standard, filter the output results of the multi-way algorithm in Step 3, remove invalid results that exceed the candidate range, and retain compliant results to enter the voting stage; Step 5: Assign fixed weights to the valid results retained in Step 4, perform weighted cumulative voting, and select the developmental stage with the highest cumulative score as the final recognition conclusion for a single image; Step 6: Summarize the single-image recognition results of all images from the same monitoring site from Step 5, filter out outliers using the majority voting method, and output the overall dominant developmental period of the site. Step 7: Based on the single-map results from Step 5 and the site aggregation results from Step 6, bind the single-map inference logic chain, site results, geographic coordinates, and other information to generate a Shapefile spatial vector file with a complete traceability chain, thus completing the spatialized output of the results.

2. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: Extract the timestamp from the filename of the image to be tested, parse the year, month, and day, and convert them into ordinal numbers within the year. ; The baseline upper and lower boundaries are preset for the five major developmental stages of sugarcane: sowing period, seedling emergence period, tillering period, stem elongation period, and technological maturity period. , and the time intervals for each stage; Step 1.2, calculate the year ordinal number of the current date. Minimum distance from the upper and lower boundaries of the baseline developmental period Determine the state and generate a set of valid candidates according to the following formula. : ； If the minimum distance is greater than 20 days: it is determined to be a stable developmental period, and the valid candidate set only includes the current target developmental period; if the minimum distance is less than or equal to 20 days: it is determined to be a developmental boundary period, and the valid candidate set includes the current target developmental period and adjacent developmental periods. Step 1.3: During the stable phase, directly output the current developmental stage result and terminate the process; During the boundary period, the valid candidate set is retained and proceeds to the next visual analysis process.

3. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 2, characterized in that, Step 2 is as follows: Step 2.1, Statistical analysis of pixel proportion based on HSV color space: Define the green range [25, 40, 40]-[95, 255,255], calculate the proportion of green vegetation; extract the region of interest in the image, and statistically analyze the proportion of yellowish-brown pixels to obtain the proportion of withered yellow leaves; Step 2.2: Perform a two-dimensional Fourier transform on the image to perform frequency domain analysis, calculate the ratio of spectral energy in the vertical direction to that in the horizontal direction, and obtain the plant verticality / texture direction features; Step 2.3: Calculate the vegetation coverage rate by combining the proportion of green vegetation and the proportion of withered leaves, extract edge density features by Canny edge detection, and integrate all features to form a physical visual feature set.

4. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 3, characterized in that, Step 3 specifically involves: Step 3.1, K-nearest neighbor algorithm retrieval: The CLIP model is used to extract feature vectors of the image to be tested and the reference image library, and the vector similarity is calculated to complete the nearest neighbor retrieval and output the developmental stage recognition result; the CLIP model is combined to identify the near / far view of the image, and the corresponding view reference image library is dynamically called by the lazy loading mechanism to avoid image library redundancy; Step 3.2, CLIP Multimodal Image-Text Matching: Load the preset English text prompts corresponding to different sugarcane development stages and shooting angles, map the image tensor and text features to the same 512-dimensional semantic space, and output the image-text recognition results through zero-shot cosine similarity matching; Step 3.3: Based on the physical features such as green ratio, yellow ratio, and frequency domain texture extracted in Step 2, and combined with the preset experience threshold rule engine, the recognition result is obtained through logical reasoning. Step 3.4: Use the time inference results from Step 1 as the baseline time result and summarize the preliminary identification results from the four paths.

5. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 4, characterized in that, Step 4 is as follows: Step 4.1: Read the list of legal candidate stages for the current boundary period determined in Step 1, and use it as the only valid result range; Step 4.2: Verify the recognition results of the three algorithms in Step 3: K-nearest neighbor algorithm retrieval, CLIP, and visual rule. If the result belongs to the valid candidate set, it is retained; if it exceeds the candidate range, it is judged as a logical out-of-bounds error and is directly eliminated, and does not participate in subsequent voting. Step 4.3: The time baseline backup result is considered a valid candidate set by default and is directly retained.

6. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1: Preset fixed weights for each algorithm module: K-nearest neighbor algorithm retrieval weight 1.5, CLIP multimodal matching weight 1.2, visual heuristic rule weight 1.0, and time baseline minimum weight 0.8; Step 5.2: For each type of recognition result that passes the verification, calculate the total score for each developmental stage by accumulating the corresponding weights; Step 5.3: Compare the cumulative scores of all candidate developmental stages and determine the stage with the highest score as the final recognition result of a single image.

7. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 6, characterized in that, Step 6 specifically involves: Step 6.1: Collect the final recognition results of all images from the same site and count the frequency of occurrence of results at each developmental stage; Step 6.2: Using the majority voting rule, outlier results with extremely low frequency are removed, and the development period with the highest frequency is selected as the dominant development period of the site.

8. The sugarcane development stage recognition method based on time prior constraints and multimodal visual fusion according to claim 7, characterized in that, Step 7 specifically includes: Step 7.1: Organize logical tracing information: record the time window determination results, boundary period markers, original results from four paths: K nearest neighbor algorithm retrieval / CLIP / visual rules / time benchmark, weighted voting count details, and single graph final conclusions to form a complete reasoning chain; Step 7.2: Combine the site ID corresponding to the image with the geographic coordinates in the WGS84 coordinate system, and use GeoPandas to construct a spatial data object; Step 7.3: Unify the single image recognition results, logical tracing chain, site dominant development period, and geographic information, and export them as an ESRI Shapefile spatial vector file.

Images