Search

CN-122019798-A - Grain yield prediction method, device and medium based on large model and time sequence retrieval

CN122019798ACN 122019798 ACN122019798 ACN 122019798ACN-122019798-A

Abstract

The application provides a grain yield prediction method, a grain yield prediction device and a grain yield prediction medium based on a large model and time sequence retrieval. The method comprises the steps of obtaining multi-mode data corresponding to a target area, target crops and a prediction period, processing the multi-mode data to obtain alignment data, fusing all encoding results based on a cross-mode attention mechanism to obtain multi-mode fusion characteristics, constructing a time sequence segment representation of a current growth stage, obtaining yield label information related to similar historical segments, inputting a prediction context into a supervised training large model, outputting interpretation information, triggering self-thinking verification on initial yield predicted values and interpretation information, correcting based on a reinforcement learning strategy, outputting yield predicted values and uncertainty intervals of the target area and the target crops, and outputting influence factors related to the yield predicted values and similar historical segment reference information. The application can improve the multi-mode fusion modeling capability, improve the prediction stability of extreme years and enhance the interpretation of the prediction result.

Inventors

  • XIE SHULEI

Assignees

  • 北京衔远有限公司

Dates

Publication Date
20260512
Application Date
20260414

Claims (10)

  1. 1. A grain yield prediction method based on a large model and time sequence retrieval is characterized by comprising the following steps: Acquiring multi-mode data corresponding to a target area, a target crop and a prediction period, wherein the multi-mode data comprises a meteorological time sequence, a remote sensing image and an agricultural condition text; Performing time alignment, space mapping and growth stage labeling on the multi-mode data to obtain alignment data associated with a unified space unit and a unified time axis; Respectively executing feature coding on the meteorological time sequence, the remote sensing image and the agricultural text in the alignment data, and fusing all coding results based on a cross-modal attention mechanism to obtain multi-modal fusion features; constructing a current growth stage time sequence segment representation based on the multi-mode fusion characteristics and in combination with the growth stage labeling; The current growth stage time sequence segment is used as a retrieval condition, similar historical segments are retrieved from a historical time sequence segment knowledge base, and yield label information associated with the similar historical segments is obtained; combining the multi-mode fusion characteristics, the current growth stage time sequence segment representation, the similar history segments and the yield label information into a prediction context, inputting the prediction context into a supervised and trained large model, and outputting an initial yield predicted value and interpretation information corresponding to the initial yield predicted value; triggering self-thinking verification on the initial yield predicted value and the interpretation information, correcting based on the reinforcement learning strategy, outputting the yield predicted value and the uncertainty interval of the target area and the target crop, and outputting the influence factors and similar historical fragment reference information associated with the yield predicted value.
  2. 2. The method of claim 1, wherein performing time alignment, spatial mapping, and growth phase labeling on the multi-modal data results in alignment data associated with a unified spatial unit and a unified timeline, comprising: Determining a unified space unit associated with a target area, and mapping the weather time sequence, the remote sensing image and the agricultural text to the unified space unit respectively to form a multi-mode data entry associated with the unified space unit; Determining a unified time axis corresponding to the predicted period, and performing time scale unification and time index alignment on the multi-modal data entry to form a multi-modal time series associated with the unified time axis; Marking a growth stage mark on each time position in the multi-mode time sequence based on a physical stage division rule of the target crop, and writing the growth stage mark into the corresponding multi-mode time sequence to obtain the alignment data.
  3. 3. The method according to claim 1, wherein the performing feature encoding on the weather time sequence, the remote sensing image and the agricultural text in the aligned data, and fusing the encoding results based on a cross-modal attention mechanism, to obtain a multi-modal fusion feature, includes: Performing sequence modeling on a meteorological time sequence input time sequence encoder associated with a unified space unit and a unified time axis to output a meteorological embedded sequence carrying time sequence information; Performing blocking processing on the remote sensing image which is associated with the unified space unit and falls into a prediction period, and inputting the blocked image blocks into an image encoder to output an image block embedding sequence; Generating text semantic embedding for an agricultural text input text encoder associated with the unified space unit and the prediction period, and extracting event labels related to disaster or agricultural management from the agricultural text; And carrying out alignment fusion on the meteorological embedding sequence, the image block embedding sequence, the text semantic embedding and the event tag based on a cross-modal attention mechanism, and outputting the multi-modal fusion characteristics.
  4. 4. The method of claim 1, wherein constructing the current growth stage timing segment representation based on the multimodal fusion feature in combination with the growth stage annotation comprises: determining a target growth stage corresponding to the current moment based on the growth stage label, and determining a time sequence window parameter associated with the target growth stage from the multi-mode fusion characteristic, wherein the time sequence window parameter comprises a window length or a time span; Intercepting multi-mode time sequence segment characteristics aligned with a unified time axis from the multi-mode fusion characteristics according to the time sequence window parameters, and carrying out association packaging on the multi-mode time sequence segment characteristics, a unified space unit and a target growth stage; And executing aggregation or sequence representation generation processing on the packaged multi-mode time sequence segment characteristics to obtain the time sequence segment representation of the current growth stage.
  5. 5. The method of claim 1, wherein the retrieving similar historical segments in the historical time series segment knowledge base and obtaining yield tag information associated with the similar historical segments using the current growth stage time series segment representation as a retrieval condition comprises: Constructing a historical time sequence segment knowledge base, wherein a plurality of historical time sequence segment entries are stored in the historical time sequence segment knowledge base, each historical time sequence segment entry comprises a historical growth stage time sequence segment representation and yield label information associated with the historical growth stage time sequence segment representation, and the yield label information comprises actual yield or relatively perennial yield deviation; generating a retrieval representation based on the current growth stage timing sequence segment representation, and calculating the similarity between the retrieval representation and each historical growth stage timing sequence segment representation in the historical timing sequence segment knowledge base; And selecting a preset number of similar historical fragments from the historical time sequence fragment knowledge base according to the similarity, and outputting yield label information corresponding to the similar historical fragments and fragment identification information associated with the similar historical fragments.
  6. 6. The method of claim 1, wherein inputting the prediction context into a supervised trained large model, outputting an initial yield prediction value and interpretation information corresponding to the initial yield prediction value, comprises: Generating current state description information based on the multi-mode fusion characteristics and the current growth stage time sequence segment representation, and generating historical case description information based on yield label information and segment identification information of the similar historical segments; Combining the current state description information and the historical case description information according to a preset context structure to obtain a prediction context, wherein the prediction context comprises a target area identifier, a target crop identifier, a prediction period identifier, a current growth stage mark and historical case description information; Inputting the prediction context into the supervised trained large model, outputting an initial yield prediction value by the large model, and synchronously outputting interpretation information associated with the initial yield prediction value, wherein the interpretation information comprises main influence factor description and similar history fragment reference information.
  7. 7. The method of claim 1, wherein triggering a self-jeopardy check on the initial yield prediction value and interpretation information and correcting based on reinforcement learning strategies, outputting yield prediction values and uncertainty intervals for the target area, target crop, and outputting impact factors and similar historical segment reference information associated with the yield prediction values, comprises: Generating a thinking-back input based on the initial yield predicted value, the interpretation information and the prediction context, and taking the thinking-back input as the input of the large model to obtain a consistency check result, wherein the consistency check result is used for indicating the matching state between the influence factors in the interpretation information and the similar historical fragment reference information and the prediction context; Generating correction guidance information based on preset rewarding rules in the condition that the consistency check result indicates that mismatch or omission exists, wherein the preset rewarding rules comprise rules related to true yield deviation, yield change direction consistency and interpretation consistency; invoking reinforcement learning strategies based on the correction guide information to execute correction on the initial yield predicted value and the interpretation information, so as to obtain corrected yield predicted value and corrected interpretation information; and generating the uncertainty interval based on the corrected yield predicted value, and outputting an influence factor and similar historical fragment reference information associated with the yield predicted value based on the corrected interpretation information.
  8. 8. Grain yield prediction device based on big model and time sequence retrieval, characterized by comprising: The acquisition module is used for acquiring multi-mode data corresponding to the target area, the target crop and the prediction period, wherein the multi-mode data comprises a meteorological time sequence, a remote sensing image and an agricultural condition text; The alignment module is used for executing time alignment, space mapping and growth stage labeling on the multi-mode data to obtain alignment data associated with the unified space unit and the unified time axis; The fusion module is used for respectively executing feature coding on the meteorological time sequence, the remote sensing image and the agricultural text in the alignment data, and fusing all coding results based on a cross-modal attention mechanism to obtain multi-modal fusion features; The construction module is used for constructing the current growth stage time sequence segment representation based on the multi-mode fusion characteristics and in combination with the growth stage annotation; the retrieval module is used for retrieving similar historical fragments in the historical time sequence fragment knowledge base and obtaining yield label information associated with the similar historical fragments by taking the current growth stage time sequence fragments as retrieval conditions; The prediction module is used for combining the multi-mode fusion characteristics, the time sequence segment representation of the current growth stage, the similar historical segments and the yield label information into a prediction context, inputting the prediction context into a large model subjected to supervision training, and outputting an initial yield predicted value and interpretation information corresponding to the initial yield predicted value; The output module is used for triggering the self-thinking verification on the initial yield predicted value and the interpretation information, correcting based on the reinforcement learning strategy, outputting the yield predicted value and the uncertainty interval of the target area and the target crop, and outputting the influence factors and similar historical fragment reference information associated with the yield predicted value.
  9. 9. An electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.
  10. 10. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any one of claims 1 to 7.

Description

Grain yield prediction method, device and medium based on large model and time sequence retrieval Technical Field The application relates to the technical field of artificial intelligence, in particular to a grain yield prediction method, device and medium based on a large model and time sequence retrieval. Background Grain yield prediction is an important foundation for agricultural production organization, grain storage regulation and control and disaster handling, and is generally required to combine meteorological conditions, growth condition monitoring and agricultural condition information in the crop growth process to judge the yield or unit yield of a target area and a target crop in a prediction period. In the prior art, the yield prediction method is based on statistical regression, a mechanism model or a machine learning model based on remote sensing and meteorological features, wherein one type of method establishes an empirical relationship between historical yield and meteorological indexes, the other type of method characterizes growth vigor by remote sensing features such as vegetation indexes and carries out regression estimation, and part of methods try to introduce text agriculture condition or disaster records as auxiliary variables. However, although the method can realize certain prediction capability in the conventional year, under the conditions of inconsistent spatial-temporal resolution of multi-source data, obvious difference of physical phases, frequent extreme weather and the like, unified modeling and consistent alignment of cross-modal information are difficult to realize at the same time, meanwhile, the regression inference based on fixed features is difficult to effectively utilize the evolution mode and corresponding yield labels of similar historical years, so that the inference basis of the extreme years is insufficient, in addition, the inference process of the existing method is mostly black box output, a referenceable historical evidence and traceable influence factor link are lacked, and a self-thinking and continuous error correction mechanism is lacked when prediction deviation appears, so that the requirements of a service side on stability and interpretation presentation are difficult to meet. Disclosure of Invention In view of the above, the embodiment of the application provides a grain yield prediction method, a grain yield prediction device and a grain yield prediction medium based on a large model and time sequence retrieval, which are used for solving the problems that multi-mode data are difficult to align and fuse uniformly, history similar cases are not utilized enough and prediction interpretation lacks traceable basis in the prior art. According to a first aspect of the embodiment of the application, a grain yield prediction method based on a large model and time sequence retrieval is provided, the grain yield prediction method comprises the steps of obtaining multi-mode data corresponding to a target area, a target crop and a prediction period, wherein the multi-mode data comprises a weather time sequence, a remote sensing image and a plot text, performing time alignment, space mapping and growth period labeling on the multi-mode data to obtain alignment data associated with a unified space unit and a unified time axis, performing feature coding on the weather time sequence, the remote sensing image and the plot text in the alignment data respectively, fusing each coding result based on a cross-mode attention mechanism to obtain multi-mode fusion features, constructing a current growth period time sequence segment representation based on the multi-mode fusion features and combining with the growth period labeling, searching a similar history segment in a history time sequence segment knowledge base and obtaining yield label information associated with the similar history segment, combining the multi-mode fusion features, the current growth period time sequence segment representation, the similar history segment and the yield label information into a prediction context, inputting the prediction context into a supervised large model, outputting an initial predicted yield value and an initial predicted yield value, an initial yield interpretation value and a predicted yield value, and a corresponding yield interpretation value and a target yield verification threshold, and a prediction factor, and a self-interpretation factor, and a prediction factor and a self-interpretation factor and a prediction factor are output. The second aspect of the embodiment of the application provides a grain yield prediction device based on a large model and time sequence retrieval, which comprises an acquisition module, a prediction module and a prediction module, wherein the acquisition module is used for acquiring multi-mode data corresponding to a target area, a target crop and a prediction period, and the multi-mode data comprises a meteorological time sequence, a remot