Crop yield prediction method and system based on artificial intelligence

Abstract

The invention relates to the field of agricultural yield prediction, and particularly discloses a crop yield prediction method and system based on artificial intelligence, and the method comprises the steps: S1, recognizing crop types through the spectral features of a remote sensing image and field sampling data, and dividing corresponding key growth stages; s2, collecting multi-source feature data for each key growth stage, wherein the multi-source feature data comprises remote sensing image data, an extracted normalized difference vegetation index NDVI, an enhanced vegetation index EVI and a topographic factor; acquiring hourly temperature, humidity, precipitation and illumination intensity of each sub-region through a distributed sensor according to micro-meteorological data; obtaining plant height, leaf area index and fruit development parameters through unmanned aerial vehicle three-dimensional imaging according to crop phenotype data; by adopting the technical scheme of the invention, the terrain, microclimate and crop dynamic growth characteristics can be accurately fused, and the problem of non-uniform sample distribution is solved, so that the accuracy and regional adaptability of yield prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural yield forecasting, and in particular to a method and system for forecasting crop yields based on artificial intelligence. Background Technology

[0002] Crop yield forecasting is a key technology for ensuring food security and sustainable agricultural development. With the development of artificial intelligence technology, forecasting methods based on machine learning and deep learning have been widely used. Existing technologies mostly construct forecasting models by integrating meteorological data (such as temperature and precipitation), remote sensing imagery (such as the NDVI vegetation index), and historical yield data. For example, algorithms such as decision trees and neural networks are used to mine the correlation between features and yield. Some methods also introduce a phased analysis approach, extracting features at different stages of the crop's entire growth cycle to improve forecast accuracy.

[0003] However, in regions with complex terrain and variable climate, existing technologies still have significant limitations: First, the influence of microclimate and topography is ignored. These regions are mostly mountainous and hilly with large differences in altitude, leading to significant discrepancies between microclimate (such as local humidity and light) and macro-meteorological data. Existing models directly use regional average meteorological data, making it difficult to reflect the impact of local environments such as mountains and valleys on crop growth. Second, the fusion of multi-source data is coarse. Existing methods mostly use static weighting to fuse remote sensing data and crop phenotypic data (such as plant height and leaf area), failing to consider the sensitivity differences of key crop growth stages (such as grain filling and tillering stages) to topographic factors (such as slope and soil moisture). For example, the difference in runoff from rainfall during the rainy season between plains and slopes has drastically different effects on crop yield. Third, the sample representativeness is insufficient. Crop planting is scattered in complex terrain areas, and measured yield samples in disturbed areas (such as steep slopes and low-lying, flood-prone areas) are scarce, resulting in large prediction errors in these areas and making it difficult to meet the needs of precision agricultural management.

[0004] Therefore, for the aforementioned regions, there is an urgent need for an artificial intelligence prediction method that can accurately integrate topography, microclimate, and crop dynamic growth characteristics, and solve the problem of uneven sample distribution, in order to improve the accuracy and regional adaptability of yield prediction. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method for predicting crop yield, which can accurately integrate topography, microclimate and crop dynamic growth characteristics, and solve the problem of uneven sample distribution, so as to improve the accuracy and regional adaptability of yield prediction.

[0006] To solve the above-mentioned technical problems, this application provides the following technical solution:

[0007] An artificial intelligence-based method for predicting crop yield includes:

[0008] S1: Identify crop types and delineate corresponding key growth stages by using the spectral characteristics of remote sensing images and field sampling data;

[0009] S2: Collect multi-source feature data for each key growth stage, the multi-source feature data including:

[0010] From remote sensing image data, normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and topographic factors are extracted.

[0011] Micro-meteorological data is obtained by using distributed sensors to acquire hourly temperature, humidity, precipitation, and light intensity for each sub-region.

[0012] Crop phenotypic data, including plant height, leaf area index, and fruit development parameters, were obtained through UAV 3D imaging.

[0013] S3: Construct a dynamic weighted feature fusion model, calculate the influence coefficient of each key growth stage based on the historical yield data of the corresponding crop type to obtain the stage weight; and correct the weight of micro-meteorological data according to slope and altitude to obtain the terrain weight.

[0014] S4: For the sparsely sampled and disturbed regions, virtual samples are generated through a generative adversarial network and combined with the actual samples to form a training set;

[0015] S5: Construct a deep learning model based on an attention mechanism, and train it by inputting the fused stage weights and terrain weights. The model focuses on terrain-sensitive areas through a spatial attention layer and strengthens the features of key growth stages through a temporal attention layer.

[0016] S6: Input the multi-source feature data of the region to be predicted into the trained model, and output the crop yield prediction results by region and stage.

[0017] The basic principle and beneficial effects of this invention are as follows: This invention achieves accurate prediction of crop yield through multi-dimensional collaborative modeling. In S1, the spectral characteristics of remote sensing images (such as differences in reflectance among different crops) are used in combination with field sampling data (such as crop morphological characteristics) to accurately identify crop types (such as rice, wheat, and corn), and specific key growth stages are divided according to their biological characteristics (such as the tillering stage of rice and the overwintering stage of wheat) to ensure that subsequent modeling matches the crop growth patterns.

[0018] In S2, remote sensing vegetation indices (NDVI and EVI reflect crop growth), topographic factors (slope and altitude affect water and light distribution), micrometeorological data (hourly temperature, humidity, precipitation, and other environmental factors), and crop phenotypic data (plant height, leaf area index, and other growth status indicators) are integrated to achieve multi-dimensional feature capture of crops, environment, and topography.

[0019] In S3, the influence coefficients of different key stages of crops are calculated based on historical data (e.g., the influence of wheat grain filling stage on yield is higher than that of seedling stage). At the same time, the weights of micro-meteorological data are adjusted according to slope and altitude (e.g., the weight of precipitation factor on slope is increased to reflect runoff differences), so that feature fusion can be adapted to crop characteristics and terrain heterogeneity.

[0020] In S4-S5, a Generative Adversarial Network (GAN) is used to generate virtual samples for disturbed areas (such as steep slopes and low-lying areas) to address the scarcity of measured data. An attention mechanism is employed to enhance the model's focus on features in terrain-sensitive areas (such as waterlogged areas at the bottom of slopes) and key growth stages (such as the tasseling stage of maize), improving the model's utilization of key information. In S6, based on the aforementioned fused features and model training results, regional and stage-specific predictions are output, achieving accurate predictions in both spatial and temporal dimensions.

[0021] By combining crop-specific stage segmentation with dynamic weight fusion, the problem of traditional models ignoring the sensitivity differences of different crops and different growth stages is solved (such as the high water demand of rice during the tillering stage being modeled in detail), thus reducing prediction errors.

[0022] Multi-source feature fusion (especially the combination of topographic factors and micro-meteorological data) overcomes the limitations of single data dimensions, significantly improving the prediction accuracy in complex terrain areas (such as mountains and slopes), and enhancing the accuracy compared to models using only meteorological data. The topographic weight correction mechanism allows the model to adapt to different landforms (such as the difference in water use between steep slopes and plains), and its prediction stability in mountainous and hilly areas is superior to traditional general models. The generation of virtual samples in disturbed areas solves the problem of uneven sample distribution, enabling the model to maintain high accuracy in data-scarce special areas (such as saline-alkali land and flood-prone areas), thus improving its adaptability.

[0023] Regional and phased output results can provide refined decision support for agricultural management (such as timely irrigation measures for drought warnings during the grain-filling stage in a certain region); the attention mechanism focuses on key features, reduces redundant data interference, improves model training efficiency, and speeds up prediction response. Crop-specific adaptive modeling allows it to be generalized to multiple crops without the need for retraining for a single crop, and its generalization ability is superior to dedicated models.

[0024] In summary, this method can accurately integrate topography, microclimate, and crop dynamic growth characteristics, solve the problem of uneven sample distribution, and improve the accuracy and regional adaptability of yield prediction.

[0025] Furthermore, S1 also includes:

[0026] S11: Perform band analysis on remote sensing images, extract the reflectance ratio of the near-infrared band to the red band, and construct crop spectral feature vectors by combining the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI).

[0027] S12: Obtain crop phenotypic characteristic parameters, including leaf length-to-width ratio, stem color and fruit morphology, through field sampling, and fuse them with spectral feature vectors to form a crop species identification matrix;

[0028] S13: Based on the identification matrix, crop types are matched, and a preset crop phenological period database is called. Combined with the temporal change characteristics of remote sensing images, key growth stages are automatically divided. For crops in mountainous and hilly areas, the start time of the growth stage is further corrected by terrain factors.

[0029] Furthermore, S2 also includes:

[0030] S21: When collecting remote sensing image data, atmospheric correction is first used to eliminate the influence of aerosol scattering. Then, the terrain shadow model is used to calculate the solar altitude angle correction value based on slope and aspect to perform terrain radiometric correction on NDVI and EVI. In addition, a terrain shading coefficient is introduced for areas with a slope > 15°.

[0031] S22: When acquiring micro-meteorological data, in a sparsely distributed area of ​​distributed sensors, the inverse distance weighted interpolation method is used to fuse the data from surrounding sensors to generate hourly meteorological data in a 10m×10m grid, while outlier filtering is performed on the precipitation data.

[0032] S23: When collecting crop phenotypic data, the UAV 3D imaging adopts multi-time period shooting, and generates the dynamic change curve of plant height through point cloud stitching technology. Combined with the time-series change rate of leaf area index, it automatically marks phenotypic abnormal samples, providing a preprocessing basis for subsequent identification of disturbed areas.

[0033] Furthermore, S3 also includes:

[0034] S31: When calculating the stage weights, the influence coefficients of each key growth stage are determined based on the historical yield data of the corresponding crop type using the following formula:

[0035]

[0036] in, Let be the stage weight of the q-th critical growth stage. The correlation coefficient between the characteristic data of this stage and historical output. This represents the historical average yield for this stage, and Q represents the total number of critical growth stages.

[0037] S32: When calculating terrain weights, based on slope and altitude Adjust the weights of micrometeorological data:

[0038]

[0039] in, For the corrected terrain weights, This is the initial meteorological weight, i.e., the baseline value for the plain area;

[0040] S33: Real-time comparison of the current stage and Compared with a preset threshold based on historical qualified intervals, if or Then, guidance information will be automatically generated, including... , The standard deviation is historical; the guidance information includes, for areas with insufficient stage weight, pushing supplementary irrigation / fertilization suggestions, and for areas with abnormal terrain weight, pushing slope drainage / terracing suggestions.

[0041] Furthermore, S3 also includes:

[0042] S34: When calculating terrain weights, in the above... Based on this, a temperature sensitivity coefficient is introduced to address the temperature decrease that occurs for every 100m increase in altitude. The corrected formula is:

[0043]

[0044] in, The weights are the combined temperature-topography corrections. For altitude, This represents the cumulative effect of every 100m increase in altitude. The compensation coefficient for the reduction in temperature on the microclimate weight is quantified;

[0045] S35: Regarding the revised With stage weights Perform collaborative verification and calculate the overall fit using the following formula:

[0046]

[0047] in, The weighted overall fit is 0-1. For the same period in history and The average of the products, The time frame indicates an adaptation anomaly.

[0048] S36: If Then, by combining real-time micro-meteorological data and crop phenotypic data, differentiated guidance can be generated:

[0049] For high-altitude areas, i.e. and For larger, warm-loving crops, we recommend using mulch film to increase temperature and staggering sowing times.

[0050] For areas with a slope greater than 20°, a contour planting runoff interception scheme is proposed to optimize the terrain weight adaptability;

[0051] S37: Collect feedback data after the implementation of the guidance measures, and adjust the weights a second time using the following formula:

[0052]

[0053] in, For the final terrain weight, This is the feedback adjustment coefficient, with a value ranging from 0.05 to 0.1. The amount of temperature change after the implementation of the measures; Synchronize to the deep learning model and update the training parameters.

[0054] Furthermore, S4 also includes:

[0055] S41: Classify the disturbed areas according to terrain type, and construct a dedicated generative adversarial network sub-model for each type of area. The generator input consists of a small number of historical measured samples of that type of area. The measured samples include NDVI, slope, and phenotypic data. The discriminator verifies the authenticity of the virtual samples using the following formula:

[0056]

[0057] in, Scoring the similarity between virtual and real samples. It was deemed qualified. For terrain feature similarity, For crop phenotypic similarity;

[0058] S42: For qualified virtual samples, combine the corrected terrain weights. Filter and retain The sample, The historical weighted average of similar regions, excluding extreme values;

[0059] S43: Mix the selected virtual samples with the actual samples in a 3:7 ratio to form a training set. The virtual samples are labeled to generate labels for the model to distinguish and learn during training in S5, thus avoiding overfitting.

[0060] Furthermore, the spatial attention layer in S5 focuses on terrain-sensitive areas, including:

[0061] S51: The spatial attention layer calculates the attention weight for each pixel using the following formula:

[0062]

[0063] in, Let (i,j) be the spatial attention weight for the (i,j)th pixel. It is the Sigmoid activation function. The vegetation index for that pixel. For slope, For altitude, These are the weighting coefficients. For bias terms;

[0064] S52: Yes In the region, reduce its feature weight; for In terrain-sensitive areas, feature extraction is enhanced through residual connections to retain more detailed information;

[0065] S53: Combine the training set and assign a loss weight of 1.2 times to samples in terrain-sensitive areas during model training to force the model to focus on learning.

[0066] Furthermore, the temporal attention layer in S5 enhances key growth stage features, including:

[0067] S54: The temporal attention layer calculates the stage attention weights for each key growth stage of the crop using the following formula:

[0068]

[0069] in, Let be the temporal weight for the q-th stage. Scaling factor This is the correlation coefficient between this stage and output. This represents the historical output fluctuation value for this period. Total number of stages;

[0070] S55: During model training, for In the critical stage, the duration of its feature input is extended to 1.5 times the original duration to increase the capture dimension of temporal features;

[0071] S56: Combining stage weights ,like If this is triggered, feature resampling is performed to supplement historical similar samples in this stage and balance temporal learning.

[0072] Furthermore, S6 also includes:

[0073] S61: Divide the fused multi-source feature data, including vegetation index, dynamic phenotypic data, and topographic-soil combined features, into unit regions using a 1km×1km grid, and calculate the yield prediction value for each unit using the following formula:

[0074]

[0075] in, Let (i,j) be the predicted output of the (i,j)th grid in stage q. For the regional calibration coefficient, The stage weights are as modified in claim 5. This represents the phenotypic characteristics of the q-th stage of the grid. The terrain weights are as modified in claim 7. This is the micrometeorological data for stage q of the grid. This is the random error term;

[0076] S62: For each grid The time sequence is overlaid according to the key growth stages to generate the cumulative yield curve for each stage, and the points of abrupt change in the slope of the curve are marked, corresponding to abnormal areas.

[0077] S63: The output results are classified into two levels: administrative region and growth stage. At the administrative region level, the affected areas are marked and the differences in yield at the county level are displayed on the electronic map using a heat map. At the growth stage level, the contribution ratio of each stage is output simultaneously to provide a quantitative basis for the allocation of management resources in the later stage.

[0078] S64: Link the prediction results with the virtual sample library. If the predicted value of a certain area is consistently lower than the historical average, the virtual sample supplementation mechanism of S4 will be automatically triggered to improve the regional adaptability of the next round of prediction. Attached Figure Description

[0079] Figure 1 This is a flowchart illustrating an embodiment of an artificial intelligence-based crop yield prediction method. Detailed Implementation

[0080] The following detailed description illustrates the specific implementation method:

[0081] An AI-based method for predicting crop yield (e.g.) Figure 1 (as shown), including:

[0082] S1: Identify crop types and delineate corresponding key growth stages by using the spectral characteristics of remote sensing images and field sampling data;

[0083] S2: Collect multi-source feature data for each key growth stage, the multi-source feature data including:

[0084] From remote sensing image data, normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and topographic factors are extracted.

[0085] Micro-meteorological data is obtained by using distributed sensors to acquire hourly temperature, humidity, precipitation, and light intensity for each sub-region.

[0086] Crop phenotypic data, including plant height, leaf area index, and fruit development parameters, were obtained through UAV 3D imaging.

[0087] S3: Construct a dynamic weighted feature fusion model, calculate the influence coefficient of each key growth stage based on the historical yield data of the corresponding crop type to obtain the stage weight; and correct the weight of micro-meteorological data according to slope and altitude to obtain the terrain weight.

[0088] S4: For the sparsely sampled and disturbed regions, virtual samples are generated through a generative adversarial network and combined with the actual samples to form a training set;

[0089] S5: Construct a deep learning model based on an attention mechanism, and train it by inputting the fused stage weights and terrain weights. The model focuses on terrain-sensitive areas through a spatial attention layer and strengthens the features of key growth stages through a temporal attention layer.

[0090] S6: Input the multi-source feature data of the region to be predicted into the trained model, and output the crop yield prediction results by region and stage.

[0091] In practical application: Taking a mountainous rice-growing area with steep slopes and an altitude of 300-800m as an example, the yield of early rice is predicted. In S1, the near-infrared / red reflectance ratio of 1.2-1.8 is extracted from Sentinel-2 remote sensing images. Combined with the leaf length-to-width ratio of 3.5-4.0 and the stem color of bluish-green from field sampling, the crop type is identified as early rice. The phenological database is called up, and the key stages are assumed to be: seedling stage in March-April, tillering stage in April-May, and grain-filling stage in June. In steep slope areas with a slope >25°, the start time of tillering is delayed by 5 days compared with the plain area (topographic factor correction).

[0092] Remote sensing data in S2: Assuming NDVI is 0.6-0.8, EVI is 0.5-0.7, slope is 15°-30°, and elevation data are extracted;

[0093] The micro-meteorological data includes hourly temperature data of 20-30℃ and humidity data of 60%-90% acquired by distributed sensors. Sparse areas are supplemented by interpolation to achieve 10m×10m grid coverage.

[0094] Phenotypic data were obtained by drone photography weekly, capturing plant heights of 60-100cm and leaf area indices of 3.0-5.0.

[0095] S3 stage weight: Assuming the correlation coefficient between tillering stage and yield is based on historical data from the past 5 years. Historical average yield Assuming the calculation yields (Highest percentage);

[0096] Terrain weights include initial meteorological weights. With a slope of 20° and an altitude of 500m, assuming corrections... .

[0097] S4: For steep slopes with a gradient > 30° and scarce samples, GAN generates virtual samples (NDVI = 0.65, plant height 75cm), which are then used to form a training set with the actual measured samples (30 samples) in a 3:7 ratio.

[0098] S5: Attention model training, the spatial layer focuses on areas above 600m altitude, and the temporal layer strengthens tillering stage features.

[0099] S6: The output results show that the yield in areas affected by steep slopes during the tillering stage is 15% lower than that in plain areas, and the yield in high-temperature areas (>30℃) during the grouting stage fluctuates by ±8%.

[0100] Specifically, S1 also includes:

[0101] S11: Perform band analysis on remote sensing images, extract the reflectance ratio of the near-infrared band to the red band, and construct crop spectral feature vectors by combining the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI).

[0102] S12: Obtain crop phenotypic characteristic parameters, including leaf length-to-width ratio, stem color and fruit morphology, through field sampling, and fuse them with spectral feature vectors to form a crop species identification matrix;

[0103] S13: Based on the identification matrix, crop types are matched, and a preset crop phenological period database is called. Combined with the temporal change characteristics of remote sensing images, key growth stages are automatically divided. For crops in mountainous and hilly areas, the start time of the growth stage is further corrected by terrain factors.

[0104] For specific usage, supplementary explanations will be provided based on wheat-growing areas (assuming an altitude of 800-1200m).

[0105] S11: Extract near-infrared / red light from Landsat-8 images with settings of 1.5-2.0, NDVI=0.4-0.6, and EVI=0.3-0.5 to construct spectral vectors.

[0106] S12: Field sampling yielded a leaf length-to-width ratio of 2.0 and a stem color of gray-green, which were identified as winter wheat after fusion.

[0107] S13: Use the phenology database to divide the overwintering period (December-February) and the jointing period (March), with the start of the jointing period delayed by 7 days in areas above 1000m altitude (corrected by topographic factors).

[0108] S2 further includes:

[0109] S21: When collecting remote sensing image data, atmospheric correction is first used to eliminate the influence of aerosol scattering. Then, the terrain shadow model is used to calculate the solar altitude angle correction value based on slope and aspect to perform terrain radiometric correction on NDVI and EVI. In addition, a terrain shading coefficient is introduced for areas with a slope > 15°.

[0110] S22: When acquiring micro-meteorological data, in a sparsely distributed area of ​​distributed sensors, the inverse distance weighted interpolation method is used to fuse the data from surrounding sensors to generate hourly meteorological data in a 10m×10m grid, while outlier filtering is performed on the precipitation data.

[0111] S23: When collecting crop phenotypic data, the UAV 3D imaging adopts multi-time period shooting, and generates the dynamic change curve of plant height through point cloud stitching technology. Combined with the time-series change rate of leaf area index, it automatically marks phenotypic abnormal samples, providing a preprocessing basis for subsequent identification of disturbed areas.

[0112] Supplementary explanation using hilly cornfields (slope 10°-20°). S21: After atmospheric correction, a shading factor of 0.15 was introduced for the 15° slope area in the Sentinel-2 image. After NDVI correction, the results are consistent with the measured growth. S22: After interpolation in the sparse areas of the sensor, the extreme values ​​of precipitation data were reduced by 20%, and the grid coverage reached 90%. S23: Weekly drone footage shows the plant height dynamic curve marking the abnormal area in mid-June (tasseling stage) (plant height was 10% lower than the average).

[0113] S3 further includes:

[0114] S31: When calculating the stage weights, the influence coefficients of each key growth stage are determined based on the historical yield data of the corresponding crop type using the following formula:

[0115]

[0116] in, Let be the stage weight of the q-th critical growth stage. The correlation coefficient between the characteristic data of this stage and historical output. This represents the historical average yield for this stage, and Q represents the total number of critical growth stages.

[0117] S32: When calculating terrain weights, based on slope and altitude Adjust the weights of micrometeorological data:

[0118]

[0119] in, For the corrected terrain weights, This is the initial meteorological weight, i.e., the baseline value for the plain area;

[0120] S33: Real-time comparison of the current stage and Compared with a preset threshold based on historical qualified intervals, if or Then, guidance information will be automatically generated, including... , The standard deviation is historical. The guidance information includes, for areas with insufficient stage weight, pushing supplementary irrigation / fertilization suggestions (the specific content is matched with the form corresponding to the pre-set insufficient weight situation), and for areas with abnormal terrain weight, pushing slope drainage / terracing improvement suggestions.

[0121] Supplementary explanations will be provided for rapeseed growing areas (for different species).

[0122] S31: The correlation coefficient between flowering period (assumed to be March) and yield is assumed to be... Historical average yield ,calculate .

[0123] S32: Initial meteorological weights After a slope correction of 5° .

[0124] S33: With a threshold of 0.5, suggestions for fertilization during the flowering period are pushed out.

[0125] S3 further includes:

[0126] S34: When calculating terrain weights, in the above... Based on this, a temperature sensitivity coefficient is introduced to address the temperature decrease that occurs for every 100m increase in altitude. The corrected formula is:

[0127]

[0128] in, The weights are the combined temperature-topography corrections. For altitude, This represents the cumulative effect of every 100m increase in altitude. The compensation coefficient for the reduction in temperature on the microclimate weight is quantified;

[0129] S35: Regarding the revised With stage weights Perform collaborative verification and calculate the overall fit using the following formula:

[0130]

[0131] in, The weighted overall fit is 0-1. For the same period in history and The average of the products, The time frame indicates an adaptation anomaly.

[0132] S36: If Then, by combining real-time micro-meteorological data and crop phenotypic data, differentiated guidance can be generated:

[0133] For high-altitude areas, i.e. and For larger, warm-loving crops, we recommend using mulch film to increase temperature and staggering sowing times.

[0134] For areas with a slope greater than 20°, a contour planting runoff interception scheme is proposed to optimize the terrain weight adaptability;

[0135] S37: Collect feedback data after the implementation of the guidance measures, and adjust the weights a second time using the following formula:

[0136]

[0137] in, For the final terrain weight, This is the feedback adjustment coefficient, with a value ranging from 0.05 to 0.1 (set according to actual needs). The amount of temperature change after the implementation of the measures; Synchronize to the deep learning model and update the training parameters.

[0138] In practical application, S34: based on the temperature sensitivity coefficient (For warm-season crops, set according to actual needs), after correction .

[0139] S35: Overall Compatibility An anomaly was detected.

[0140] S36: Recommendation for warming with film coating.

[0141] S4 further includes:

[0142] S41: Classify the disturbed areas according to terrain type, and construct a dedicated generative adversarial network sub-model for each type of area. The generator input consists of a small number of historical measured samples of that type of area. The measured samples include NDVI, slope, and phenotypic data. The discriminator verifies the authenticity of the virtual samples using the following formula:

[0143]

[0144] in, Scoring the similarity between virtual and real samples. It was deemed qualified. For terrain feature similarity, For crop phenotypic similarity;

[0145] S42: For qualified virtual samples, combine the corrected terrain weights. Filter and retain The sample, The historical weighted average of similar regions, excluding extreme values;

[0146] S43: Mix the selected virtual samples with the actual samples in a 3:7 ratio to form a training set. The virtual samples are labeled to generate labels for the model to distinguish and learn during training in S5, thus avoiding overfitting.

[0147] The spatial attention layer in S5 focuses on terrain-sensitive areas, including:

[0148] S51: The spatial attention layer calculates the attention weight for each pixel using the following formula:

[0149]

[0150] in, Let (i,j) be the spatial attention weight for the (i,j)th pixel. It is the Sigmoid activation function. The vegetation index for that pixel. For slope, For altitude, These are the weighting coefficients. For bias terms;

[0151] S52: Yes In the region, reduce its feature weight; for In terrain-sensitive areas, feature extraction is enhanced through residual connections to retain more detailed information;

[0152] S53: Combine the training set and assign a loss weight of 1.2 times to samples in terrain-sensitive areas during model training to force the model to focus on learning.

[0153] The temporal attention layer in S5 enhances key growth stage features, including:

[0154] S54: The temporal attention layer calculates the stage attention weights for each key growth stage of the crop using the following formula:

[0155]

[0156] in, Let be the temporal weight for the q-th stage. Scaling factor This is the correlation coefficient between this stage and output. This represents the historical output fluctuation value for this period. Total number of stages;

[0157] S55: During model training, for In the critical stage, the duration of its feature input is extended to 1.5 times the original duration to increase the capture dimension of temporal features;

[0158] S56: Combining stage weights ,like If this is triggered, feature resampling is performed to supplement historical similar samples in this stage and balance temporal learning.

[0159] S6 further includes:

[0160] S61: Divide the fused multi-source feature data, including vegetation index, dynamic phenotypic data, and topographic-soil combined features, into unit regions using a 1km×1km grid, and calculate the yield prediction value for each unit using the following formula:

[0161]

[0162] in, Let (i,j) be the predicted output of the (i,j)th grid in stage q. For the regional calibration coefficient, The stage weights are as modified in claim 5. This represents the phenotypic characteristics of the q-th stage of the grid. The terrain weights are as modified in claim 7. This is the micrometeorological data for stage q of the grid. This is the random error term;

[0163] S62: For each grid The time sequence is overlaid according to the key growth stages to generate the cumulative yield curve for each stage, and the points of abrupt change in the slope of the curve are marked, corresponding to abnormal areas.

[0164] S63: The output results are classified into two levels: administrative region and growth stage. At the administrative region level, the affected areas are marked and the differences in yield at the county level are displayed on the electronic map using a heat map. At the growth stage level, the contribution ratio of each stage is output simultaneously to provide a quantitative basis for the allocation of management resources in the later stage.

[0165] S64: Link the prediction results with the virtual sample library. If the predicted value of a certain area is consistently lower than the historical average, the virtual sample supplementation mechanism of S4 will be automatically triggered to improve the regional adaptability of the next round of prediction.

[0166] Validated grid prediction: 1km×1km unit yield The actual yield was 492 kg / mu, with an error of 2.5%, achieving a relatively accurate prediction.

[0167] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting crop yield based on artificial intelligence, characterized in that, S1: Identify crop types and delineate corresponding key growth stages by using the spectral characteristics of remote sensing images and field sampling data; S2: Collect multi-source feature data for each key growth stage, the multi-source feature data including: From remote sensing image data, normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and topographic factors are extracted. Micro-meteorological data is obtained by using distributed sensors to acquire hourly temperature, humidity, precipitation, and light intensity for each sub-region. Crop phenotypic data, including plant height, leaf area index, and fruit development parameters, were obtained through UAV 3D imaging. S3: Construct a dynamic weighted feature fusion model, calculate the influence coefficient of each key growth stage based on the historical yield data of the corresponding crop type to obtain the stage weight; and correct the weight of micro-meteorological data according to slope and altitude to obtain the terrain weight. S4: For the sparsely sampled and disturbed regions, virtual samples are generated through a generative adversarial network and combined with the actual samples to form a training set; S5: Construct a deep learning model based on an attention mechanism, and train it by inputting the fused stage weights and terrain weights. The model focuses on terrain-sensitive areas through a spatial attention layer and strengthens the features of key growth stages through a temporal attention layer. S6: Input the multi-source feature data of the region to be predicted into the trained model, and output the crop yield prediction results by region and stage.

2. The method for predicting crop yield based on artificial intelligence according to claim 1, characterized in that, S1 further includes: S11: Perform band analysis on remote sensing images, extract the reflectance ratio of the near-infrared band to the red band, and construct crop spectral feature vectors by combining the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI). S12: Obtain crop phenotypic characteristic parameters, including leaf length-to-width ratio, stem color and fruit morphology, through field sampling, and fuse them with spectral feature vectors to form a crop species identification matrix; S13: Based on the identification matrix, crop types are matched, and a preset crop phenological period database is called. Combined with the temporal change characteristics of remote sensing images, key growth stages are automatically divided. For crops in mountainous and hilly areas, the start time of the growth stage is further corrected by terrain factors.

3. The method for predicting crop yield based on artificial intelligence according to claim 2, characterized in that, S2 further includes: S21: When collecting remote sensing image data, atmospheric correction is first used to eliminate the influence of aerosol scattering. Then, the terrain shadow model is used to calculate the solar altitude angle correction value based on slope and aspect to perform terrain radiometric correction on NDVI and EVI. In addition, a terrain shading coefficient is introduced for areas with a slope > 15°. S22: When acquiring micro-meteorological data, in a sparsely distributed area of ​​distributed sensors, the inverse distance weighted interpolation method is used to fuse the data from surrounding sensors to generate hourly meteorological data in a 10m×10m grid, while outlier filtering is performed on the precipitation data. S23: When collecting crop phenotypic data, the UAV 3D imaging adopts multi-time period shooting, and generates the dynamic change curve of plant height through point cloud stitching technology. Combined with the time-series change rate of leaf area index, it automatically marks phenotypic abnormal samples, providing a preprocessing basis for subsequent identification of disturbed areas.

4. The method for predicting crop yield based on artificial intelligence according to claim 3, characterized in that, S3 further includes: S31: When calculating the stage weights, the influence coefficients of each key growth stage are determined based on the historical yield data of the corresponding crop type using the following formula: in, Let be the stage weight of the q-th critical growth stage. The correlation coefficient between the characteristic data of this stage and historical output. This represents the historical average yield for this stage, and Q represents the total number of critical growth stages. S32: When calculating terrain weights, based on slope and altitude Adjust the weights of micrometeorological data: in, For the corrected terrain weights, This is the initial meteorological weight, i.e., the baseline value for the plain area; S33: Real-time comparison of the current stage and Compared with a preset threshold based on historical qualified intervals, if or Then, guidance information will be automatically generated, including... , The standard deviation is historical; the guidance information includes, for areas with insufficient stage weight, pushing supplementary irrigation / fertilization suggestions, and for areas with abnormal terrain weight, pushing slope drainage / terracing suggestions.

5. The method for predicting crop yield based on artificial intelligence according to claim 4, characterized in that, S3 further includes: S34: When calculating terrain weights, in the above... Based on this, a temperature sensitivity coefficient is introduced to address the temperature decrease that occurs for every 100m increase in altitude. The corrected formula is: in, The weights are the combined temperature-topography corrections. For altitude, This represents the cumulative effect of every 100m increase in altitude. The compensation coefficient for the reduction in temperature on the microclimate weight is quantified; S35: Regarding the revised With stage weights Perform collaborative verification and calculate the overall fit using the following formula: in, The weighted overall fit is 0-1. For the same period in history and The average of the products, The time frame indicates an adaptation anomaly. S36: If Then, by combining real-time micro-meteorological data and crop phenotypic data, differentiated guidance can be generated: For high-altitude areas, i.e. and For larger, warm-loving crops, we recommend using mulch film to increase temperature and staggering sowing times. For areas with a slope greater than 20°, a contour planting runoff interception scheme is proposed to optimize the terrain weight adaptability; S37: Collect feedback data after the implementation of the guidance measures, and adjust the weights a second time using the following formula: in, For the final terrain weight, This is the feedback adjustment coefficient, with a value ranging from 0.05 to 0.

1. The amount of temperature change after the implementation of the measures; Synchronize to the deep learning model and update the training parameters.

6. The method for predicting crop yield based on artificial intelligence according to claim 5, characterized in that, S4 further includes: S41: Classify the disturbed areas according to terrain type, and construct a dedicated generative adversarial network sub-model for each type of area. The generator input consists of a small number of historical measured samples of that type of area. The measured samples include NDVI, slope, and phenotypic data. The discriminator verifies the authenticity of the virtual samples using the following formula: in, Scoring the similarity between virtual and real samples. It was deemed qualified. For terrain feature similarity, For crop phenotypic similarity; S42: For qualified virtual samples, combine the corrected terrain weights. Filter and retain The sample, The historical weighted average of similar regions, excluding extreme values; S43: Mix the selected virtual samples with the actual samples in a 3:7 ratio to form a training set. The virtual samples are labeled to generate labels for the model to distinguish and learn during training in S5, thus avoiding overfitting.

7. The method for predicting crop yield based on artificial intelligence according to claim 6, characterized in that, The spatial attention layer in S5 focuses on terrain-sensitive areas, including: S51: The spatial attention layer calculates the attention weight for each pixel using the following formula: in, Let (i,j) be the spatial attention weight for the (i,j)th pixel. It is the Sigmoid activation function. The vegetation index for that pixel. For slope, For altitude, These are the weighting coefficients. For bias terms; S52: Yes In the region, reduce its feature weight; for In terrain-sensitive areas, feature extraction is enhanced through residual connections to retain more detailed information; S53: Combine the training set and assign a loss weight of 1.2 times to samples in terrain-sensitive areas during model training to force the model to focus on learning.

8. The method for predicting crop yield based on artificial intelligence according to claim 7, characterized in that, The temporal attention layer in S5 enhances key growth stage features, including: S54: The temporal attention layer calculates the stage attention weights for each key growth stage of the crop using the following formula: in, Let be the temporal weight for the q-th stage. Scaling factor This is the correlation coefficient between this stage and output. This represents the historical output fluctuation value for this period. Total number of stages; S55: During model training, for In the critical stage, the duration of its feature input is extended to 1.5 times the original duration to increase the capture dimension of temporal features; S56: Combining stage weights ,like If this is triggered, feature resampling is performed to supplement historical similar samples in this stage and balance temporal learning.

9. The method for predicting crop yield based on artificial intelligence according to claim 8, characterized in that, S6 further includes: S61: Divide the fused multi-source feature data, including vegetation index, dynamic phenotypic data, and topographic-soil combined features, into unit regions using a 1km×1km grid, and calculate the yield prediction value for each unit using the following formula: in, Let (i,j) be the predicted output of the (i,j)th grid in stage q. For the regional calibration coefficient, The stage weights are as modified in claim 5. This represents the phenotypic characteristics of the q-th stage of the grid. The terrain weights are as modified in claim 7. This is the micrometeorological data for stage q of the grid. This is the random error term; S62: For each grid The time sequence is overlaid according to the key growth stages to generate the cumulative yield curve for each stage, and the points of abrupt change in the slope of the curve are marked, corresponding to abnormal areas. S63: The output results are classified into two levels: administrative region and growth stage. At the administrative region level, the affected areas are marked and the differences in yield at the county level are displayed on the electronic map using a heat map. At the growth stage level, the contribution ratio of each stage is output simultaneously to provide a quantitative basis for the allocation of management resources in the later stage. S64: Link the prediction results with the virtual sample library. If the predicted value of a certain area is consistently lower than the historical average, the virtual sample supplementation mechanism of S4 will be automatically triggered to improve the regional adaptability of the next round of prediction.

10. An artificial intelligence-based crop yield prediction system, characterized in that, The method described in any one of claims 1-9 was adopted.

Images