A method for real-time monitoring of corn growth stages in a farm by synergizing ground cameras and remote sensing

Abstract

The application discloses a kind of corn growth period real-time monitoring methods of farm cooperation ground camera and remote sensing.The method utilizes ground camera to continuously obtain the corn photo of observation plot, and the growth period of corn photo is classified and identified by deep learning model ResNet-50, and the Sentinel-2 satellite time series data of the corresponding area of observation point is continuously obtained, after radiation calibration, atmospheric correction, vegetation index calculation, the vegetation index time series is constructed, and the smooth curve is generated by interpolation reconstruction and Savitzky-Golay filtering, and the growth period date identified on the ground is accurately calibrated, to form the satellite scale standard reference curve with growth period label.Simultaneously obtain the Sentinel-2 satellite data of other plots in farm area, and the standard reference curve is used to fit it, to preliminarily predict the growth period of other plots.The method provides a feasible scheme for realizing the automation, high-precision monitoring of crop phenology at farm scale, and has important significance for guiding agricultural production management.

Description

Technical Field

[0001] This invention relates to the field of agricultural information technology, specifically to a method for near real-time monitoring of maize growth stages in farms using a combination of ground cameras and satellite remote sensing. Background Technology

[0002] The crop growth stage is a key phenological phase characterizing the growth and development of crops. Timely and accurate monitoring and forecasting of the growth stage are of great significance for guiding field management (such as irrigation and fertilization), assessing crop growth, predicting yield, and responding to the impacts of climate change. Although traditional field manual survey methods yield reliable results, they rely heavily on the experience of farmers and have drawbacks such as high cost, low efficiency, strong subjectivity, and difficulty in large-scale implementation, failing to meet the needs of modern agriculture for real-time, dynamic, and wide-area monitoring.

[0003] In recent years, satellite remote sensing technology has become a major technical means for agricultural remote sensing monitoring at regional and even global scales due to its advantages such as wide coverage, stable revisit cycles, and rich historical data accumulation. By interpreting time-series remote sensing images acquired by medium- and high-resolution satellites such as Landsat, Sentinel-2, and MODIS, and calculating spectral indicators such as the normalized vegetation index, the growth dynamic curve of crops can be retrieved, thereby inferring their critical growth stages. However, existing satellite remote sensing-based growth stage monitoring usually relies on synchronous ground-based manual observation records for calibration and verification. This not only introduces additional labor costs and subjective errors but also limits its ability to monitor growth stages over large areas. Near-ground observation technologies, represented by IoT sensors, have developed rapidly. These technologies can quickly acquire crop canopy spectra and images, are unaffected by cloud cover, and are sensitive to crop growth and subtle changes, achieving high-precision monitoring at single-point or field scales. However, their observation range is limited, and the equipment deployment and maintenance costs are high, making it difficult to directly use for agricultural monitoring at the farm scale. Meanwhile, near-ground observation technologies, such as IoT sensors and ground-based phenological cameras, have developed rapidly in recent years due to their ability to collect crop canopy spectra, images, and environmental parameters at high frequency and high precision. These technologies offer fast response times, are less affected by weather, and can capture subtle changes in crop growth, enabling high-precision monitoring at the field or even individual plant scale, effectively compensating for the insufficient spatiotemporal resolution of satellite data. However, their observation range is limited by the density of equipment deployment, and large-scale deployment and maintenance costs are high, making it difficult to independently support the systematic monitoring needs at the farm or regional scale, and they also face challenges in practical applications.

[0004] Near-Earth observation and satellite remote sensing have formed a complementary yet fragmented pattern in crop growth monitoring: near-Earth observation offers high precision and fast response, but is difficult to extend to large areas; satellite remote sensing provides wide coverage and is easily acquired, but relies on ground-based calibration and is susceptible to environmental noise. Currently, the two have not achieved efficient synergy and deep complementarity, failing to construct a crop growth monitoring system that combines point and area measurements with spatiotemporal continuity. Automated, high-precision monitoring and prediction of crop growth at the farm scale still faces technical bottlenecks.

[0005] There is an urgent need to develop a technical approach that can organically integrate the advantages of near-ground observation and satellite remote sensing, enabling dynamic calibration and correction of the satellite remote sensing inversion process using high-precision ground data, improving the automation and accuracy of farm-scale growth period identification, and building an integrated perception system from field plots to the entire farm area, thus providing technical support for precise agricultural monitoring in smart agriculture. Summary of the Invention

[0006] The purpose of this invention is to provide a near-real-time monitoring method for maize growth stages in farms that combines ground cameras and satellite remote sensing, in order to solve the problems of limited coverage, high cost, and difficulty in direct application to large-scale monitoring at the farm scale caused by ground camera observations.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for real-time monitoring of maize growth stages in a farm using a combination of ground cameras and remote sensing includes the following steps:

[0009] Step 1: Select the observation plot and deploy ground cameras. Process the corn photos acquired by the ground cameras to remove blurry or invalid photos. Inspect the corn growth period and divide the dataset. Train the ResNet-50 model.

[0010] Step 2: Input the photos taken by the ground camera in that year into the ResNet-50 model trained in Step 1 to infer the different growth stages of maize;

[0011] Step 3: Obtain Sentinel-2 satellite time series data of the observed plots and calculate the corresponding vegetation indices. Use interpolation to complete the vegetation index time series data, and then use Savitzky-Golay filtering to generate continuous and smooth vegetation index curves. Mark the maize growth period inferred in Step 2 on the curves to form a reference curve with growth period markings.

[0012] Step 4: Obtain Sentinel-2 data covering all target plots during the current year's maize growing season, calculate the corresponding vegetation index, and fit the obtained reference curve using the vegetation index time series reconstruction method to obtain the preliminary predicted growing season of maize in the target area.

[0013] Step 5: In the early part of the current year's growing season, use ground cameras deployed on the target plot to acquire time-series photos and input them into the recognition model trained in Step 1 to determine the exact time of at least one maize growth period that has occurred; compare and analyze the maize growth period date predicted in Step 4 with the corresponding identified growth period, and calculate the prediction deviation ΔT, where ΔT = model inferred growth period - preliminary prediction date;

[0014] Step 6 uses the early prediction deviation ΔT calculated in Step 5 to perform an overall correction on the preliminary prediction date of the late fertility period that has not yet occurred in Step 4: Corrected prediction date = Preliminary prediction date + ΔT; Output the prediction result of the entire fertility period after dynamic correction.

[0015] The method, wherein step 1 includes the following steps:

[0016] Step A: Select the observation plot and deploy ground cameras to acquire time-series photos taken continuously by the ground cameras at the observation plot. Filter the photos, remove blurry, overexposed, and underexposed invalid images, and retain valid photos that can clearly reflect the growth status of corn.

[0017] Step B: Based on the standard definition of maize growth period, agronomic experts or through confirmed phenological records, label the growth period of valid photos and construct a labeled time-series image dataset;

[0018] Step C: Divide the above dataset into training set, validation set and test set according to an appropriate ratio; use the training set and validation set to train and fine-tune the ResNet-50 deep learning model; the training goal is to enable the model to accurately identify the key growth period of maize based on the input time series image, evaluate the model performance using the test set, and save the final model that meets the accuracy requirements.

[0019] The method, wherein step 2 includes the following steps:

[0020] Step D: Input the complete time series of ground camera photos of the same observation plot into the recognition model trained in Step 1; the model performs reproductive period recognition on the input image data, infers and outputs the reproductive period category of each image, and then analyzes the temporal changes of the model output results to determine the start date of each key reproductive period.

[0021] The method, wherein step 3 includes the following steps:

[0022] Step E: Acquire satellite time-series data matching the observation plot: Obtain multi-temporal images from the Sentinel-2 satellite data platform that completely correspond to the geographical location of the observation plot described in Step D and cover the entire maize growing season of the previous year; Perform necessary preprocessing on the acquired images, including radiometric calibration, atmospheric correction and orthorectification, to eliminate sensor and atmospheric effects and obtain high-quality surface reflectance data;

[0023] Step F: Construct the NDVI time series for this observation plot: according to the formula Calculate its normalized vegetation index (NDVI), where NIR is the near-infrared reflectance and RED is the red reflectance;

[0024] Step G: Interpolation and smoothing of the vegetation index time series: For missing or outlier NDVI data caused by cloud and snow weather factors, time series interpolation is used to reconstruct the missing parts to ensure the continuity of the time series; then, the Savitzky-Golay filtering algorithm is applied to smooth the reconstructed NDVI series to filter out high-frequency fluctuations caused by sensor noise, atmospheric residual effects, etc., and generate a continuous and smooth curve that clearly and stably reflects the vegetation growth trend.

[0025] Step H calibrates the corn growth period of the reference curve: The key growth periods obtained and accurately determined in step D are calibrated onto the smooth NDVI curve generated in step G to form a standard reference curve with corn growth information.

[0026] The method, wherein step 4 includes the following steps:

[0027] Step 1: Acquire remote sensing data of the target area: For the target area in the current year that needs to be predicted for the growing season, acquire Sentinel-2 satellite time series images of the entire growing season; perform standardized preprocessing such as radiometric calibration, atmospheric correction, and orthorectification on the images to ensure data quality and consistency;

[0028] Step J: Calculate the vegetation index sequence for the farm area: Calculate the normalized vegetation index for each Sentinel-2 image after preprocessing in Step I.

[0029] Step K: Matching and fitting with the standard reference curve: Match and fit the NDVI time series data obtained in step J with the standard reference curve of the calibrated maize phenological period generated in step H.

[0030] In the formula, This represents the length of time corn takes to grow. The corn curve function is obtained by fitting Sentinel-2 image data at the scale. The initial reference curve function represents the set of corn reference curve samples. This represents the time shift of corn during its growth period, and the value is within ±30 days of a time interval of 1 day. and The fitting coefficient represents the difference between the two curves;

[0031] Step L involves using satellite remote sensing vegetation index reconstruction to preliminarily predict the maize growth period in the farm area.

[0032] The method, wherein step 5 includes the following steps:

[0033] Step M: Obtain early phenological observations for the current year: In the early growing season of the current year, use ground cameras deployed in the target area to continuously collect time-series images of maize; input these images into the maize growth stage recognition model trained and saved in step C, the model automatically analyzes and identifies at least one early growth stage that has actually occurred (e.g., seedling stage, three-leaf stage), and records the actual calendar date of its occurrence.

[0034] Step N calculates the systematic bias of the preliminary forecast: From the preliminary forecast results generated in step L, the preliminary forecast dates for the same fertility period of these same pixels are extracted, and their forecast bias is calculated: ΔT = Model inference date - Preliminary forecast date, which is used as the forecast bias ΔT of the forecast model for the current year in this region; ΔT quantifies the overall degree of advance or lag of the preliminary forecast based on the historical model in this year.

[0035] The method, wherein step 6 includes the following steps:

[0036] Step O involves dynamically correcting the prediction of the later fertility period: The regional average prediction deviation ΔT calculated in step N is uniformly applied to the preliminary prediction dates of all later fertility periods that have not yet actually occurred, obtained in step L; the correction formula is: the final prediction date after correction for a certain fertility period = the preliminary prediction date + ΔT; this operation is equivalent to performing a one-time overall translation correction on the entire later phenological calendar based on the actual situation in the early stage, so as to eliminate systematic errors between years.

[0037] Step P, Iterative Optimization and Final Output: Steps M, N, and O can be executed iteratively within a growing season; whenever new ground observations of the growth period are obtained, ΔT is recalculated, and the predictions for all remaining non-growing periods are immediately updated using the latest ΔT, achieving incremental optimization of the prediction results; finally, the predicted dates of each later growth period after dynamic correction are integrated with the actual dates of the observed earlier growth periods to form and output a complete and high-precision spatial prediction map of the entire maize growth period for the target area.

[0038] In the method described above, in step 1, the corn growth period identification model is a ResNet-50 deep learning image recognition model, which can classify and identify key growth period images of corn seedling stage, three-leaf stage, seven-leaf stage, tasseling stage, milk stage, and maturity stage from time-series images captured by ground cameras.

[0039] The method described assumes that, at the high-resolution Sentinel-2 image data scale, the maize time series curve of a certain pixel has the following relationship with the extracted standard reference curve:

[0040]

[0041] In the formula, This represents the length of time corn takes to grow. This represents the vegetation index curve function fitted at the scale of Sentinel-2 image data. The initial reference curve function represents the set of corn reference curve samples. For a single cell, This is the corresponding reference curve; This represents the time shift of corn during its growth period, and the value is within ±30 days of a time interval of 1 day. and The fitting coefficient represents the difference between the two curves.

[0042] The method described herein uses the Normalized Difference Vegetation Index (NDVI), which is calculated using the following formula: Where NIR is the near-infrared reflectance and RED is the red reflectance; during the entire NDVI time series curve reconstruction process, the number of days the curve is shifted due to phenological differences will be output synchronously; the NDVI time series of each maize pixel is matched and fitted with the reference curve to obtain the number of days of shift for each pixel, and then the growth period of each maize pixel is obtained based on the number of days of shift and the growth period information contained in the reference curve.

[0043] Beneficial effects:

[0044] 1. Since the ground camera continuously observes data in steps 1 and 2 and combines it with a deep learning model to automatically identify the maize growth period, the present invention can obtain high temporal resolution and high precision growth period information at the observation point, avoiding the problems of traditional manual surveys that rely on experience, are inefficient and highly subjective, and significantly improves the automation level of growth period acquisition.

[0045] 2. Since NDVI standard reference curves with reproductive period labels were constructed in steps 3 and 4, and curve matching models were used... Extending this technology to the farm scale enables the effective transfer of high-precision "plot-scale" information to the "farm-scale" scale, significantly improving the accuracy and scalability of large-scale fertility period prediction.

[0046] 3. Because the prediction deviation dynamic correction mechanism based on previous observation data is introduced in steps 5 and 6, and supports multiple iterative updates during the corn growing season, this invention can correct systematic errors between different years in real time, realize the gradual optimization of the growth period prediction results, and improve the applicability of the method in practical applications. Attached Figure Description

[0047] Figure 1 This is the flowchart described in this method.

[0048] Figure 2 It is a schematic diagram of a standard reference curve with the corn growth period marked.

[0049] Figure 3 This is a spatial distribution map of the predicted results for the key growth stages of maize at Hongxing Farm: tasseling stage and maturity stage. Detailed Implementation

[0050] The present invention will be described in detail below with reference to specific embodiments.

[0051] like Figure 1 As shown, the present invention provides a near-real-time monitoring method for the growth stage of maize in a farm using a combination of ground cameras and satellite remote sensing, comprising the following steps:

[0052] S1: A corn-growing area in Hongxing Farm, Bei'an City, Heilongjiang Province, was selected as a typical observation plot. Ground-based cameras were installed at unobstructed locations within this plot, and daily time-series images were taken to obtain images throughout the entire corn growing season. The acquired photos underwent quality screening, removing invalid images due to weather, lighting, or other factors, such as blurriness, overexposure, or underexposure. Then, agricultural experts manually labeled the valid photos to determine the corresponding corn growth stage for each image. The labeled time-series images were divided into training, validation, and test sets in a 7:2:1 ratio for training a deep learning model based on an improved ResNet-50.

[0053] S2: Input the time-series images of the observed plot from the ground camera into the pre-trained growth stage recognition model. The model identifies different growth stages in the input images, sorts the recognition results according to the time series, and thus determines the growth stage of the corn, which is then used for subsequent calibration of satellite remote sensing data.

[0054] S3: First, multi-temporal imagery data was acquired from the Sentinel-2 satellite data platform, perfectly matching the geographical location of the Hongxing Farm observation plot and covering the entire maize growing season of the previous year. The images underwent standardized preprocessing including radiometric calibration, atmospheric correction, and orthorectification to generate high-quality surface reflectance products. Next, based on the preprocessed images, the Normalized Difference Vegetation Index (NDVI) was calculated scene-by-scene to construct the NDVI time series for the plot. For data gaps or anomalies caused by cloud and snow interference in this NDVI time series, interpolation was used for reconstruction. The Savitzky-Golay filtering algorithm was then used to smooth the reconstructed sequence, resulting in a smooth NDVI curve that stably reflects vegetation growth trends. Finally, the start dates of each key growth period automatically determined in step 2 were used as growth period feature points and mapped onto this smooth NDVI curve, thus generating a labeled NDVI time series curve. This curve establishes a correspondence between the temporal morphology of the remote sensing index and the ground-based growth period calendar.

[0055] S4: For the Hongxing Farm area in the current year, acquire Sentinel-2 time-series images during its growing season, and perform the same preprocessing and NDVI calculation as in step 3 to obtain the NDVI time series of the target area for the current year. Compare this series with the maize reference curve generated in step 3 according to... Perform matching and fitting, where This represents the length of time corn takes to grow. The corn curve function is obtained by fitting Sentinel-2 image data at the scale. The initial reference curve function represents the set of corn reference curve samples. This represents the time shift of corn during its growth period, and the value is within ±30 days of a time interval of 1 day. and The fitting coefficient represents the difference between the two curves.

[0056] S5: In the early part of the growing season of the current predicted year, continuously acquire time-series images of maize using ground cameras deployed at Hongxing Farm. Input the acquired early images into the same recognition model trained in Step 1 to automatically identify the date corresponding to the maize jointing stage. Extract the preliminary prediction date for the same growth stage for the corresponding pixels from the preliminary prediction results generated in Step 4. Calculate the deviation between the two: ΔT = 2 days.

[0057] S6: Apply the prediction deviation ΔT calculated in step S5 to the preliminary prediction dates of the yet-to-occur late-stage phenological periods obtained in step S4. The correction formula is: Final prediction date = Preliminary prediction of phenological period + ΔT. This operation is equivalent to performing a one-time overall translation correction on the entire preliminary prediction of the late-stage phenological calendar based on the actual situation in the early stages of this year. To improve accuracy, this step can be designed as an iterative process: whenever a new phenological period is observed through a ground camera, ΔT is recalculated, and the predictions for all remaining phenological periods are immediately updated. Finally, the dynamically corrected prediction dates of the late-stage phenological periods are output, and a high spatiotemporal accuracy spatial distribution map of maize phenological period predictions for the current year is generated and output for Hongxing Farm, as shown below. Figure 3 As shown, the root mean square error (RMSE) for the ejaculation stage is 2.24 days, and the root mean square error (RMSE) for the maturity stage is 2.47 days.

[0058] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for real-time monitoring of maize growth stages in farms using a combination of ground cameras and remote sensing, characterized in that, Includes the following steps: Step 1: Select the observation plot and deploy ground cameras. Process the corn photos acquired by the ground cameras and remove blurry or invalid photos. The growth period of maize was calibrated and the dataset was divided, and the ResNet-50 model was trained. Step 2: Input the photos taken by the ground camera in that year into the ResNet-50 model trained in Step 1 to infer the different growth stages of maize; Step 3: Obtain Sentinel-2 satellite time series data of the observed plots and calculate the corresponding vegetation indices. Use interpolation to complete the vegetation index time series data, and then use Savitzky-Golay filtering to generate continuous and smooth vegetation index curves. Mark the maize growth period inferred in Step 2 on the curves to form a reference curve with growth period markings. Step 4: Obtain Sentinel-2 data covering all target plots during the current year's maize growing season, calculate the corresponding vegetation index, and fit the obtained reference curve using the vegetation index time series reconstruction method to obtain the preliminary predicted growing season of maize in the target area. Step 5: In the early part of the current year's growing season, use ground cameras deployed on the target plot to acquire time-series photos and input them into the recognition model trained in Step 1 to determine the exact time of at least one maize growth period that has occurred; compare and analyze the maize growth period date predicted in Step 4 with the corresponding identified growth period, and calculate the prediction deviation ΔT, where ΔT = model inferred growth period - preliminary prediction date; Step 6 uses the early prediction deviation ΔT calculated in Step 5 to perform an overall correction on the preliminary prediction date of the late fertility period that has not yet occurred in Step 4: Corrected prediction date = Preliminary prediction date + ΔT; Output the prediction result of the entire fertility period after dynamic correction.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step A: Select the observation plot and deploy ground cameras to acquire time-series photos taken continuously by the ground cameras at the observation plot. Filter the photos, remove blurry, overexposed, and underexposed invalid images, and retain valid photos that can clearly reflect the growth status of corn. Step B: Based on the standard definition of maize growth period, agronomic experts or through confirmed phenological records, label the growth period of valid photos and construct a labeled time-series image dataset; Step C: Divide the above dataset into training set, validation set and test set according to an appropriate ratio; use the training set and validation set to train and fine-tune the ResNet-50 deep learning model; the training goal is to enable the model to accurately identify the key growth period of maize based on the input time series image, evaluate the model performance using the test set, and save the final model that meets the accuracy requirements.

3. The method according to claim 1, characterized in that, Step 2 The process includes the following steps: Step D: Input the complete time series of ground camera photos of the same observation plot into the recognition model trained in Step 1; The model identifies the reproductive period of the input image data, infers and outputs the reproductive period category of each image, and then analyzes the temporal changes of the model output results to determine the start date of each key reproductive period.

4. The method according to claim 3, characterized in that, Step 3 includes the following steps: Step E: Obtain satellite time-series data matching the observation plot: Obtain multi-temporal images from the Sentinel-2 satellite data platform that completely correspond to the geographical location of the observation plot described in Step D and cover the entire maize growing season of the previous year; Perform necessary preprocessing on the acquired images, including radiometric calibration, atmospheric correction and orthorectification, to eliminate sensor and atmospheric effects and obtain high-quality surface reflectance data; Step F: Construct the NDVI time series for this observation plot: according to the formula Calculate its normalized vegetation index (NDVI), where NIR is the near-infrared reflectance and RED is the red reflectance; Step G: Interpolation and smoothing of the vegetation index time series: For missing or outlier NDVI data caused by cloud and snow weather factors, time series interpolation is used to reconstruct the missing parts to ensure the continuity of the time series; then, the Savitzky-Golay filtering algorithm is applied to smooth the reconstructed NDVI series to filter out high-frequency fluctuations caused by sensor noise, atmospheric residual effects, etc., and generate a continuous and smooth curve that clearly and stably reflects the vegetation growth trend. Step H: Calibrate the maize growth period on the reference curve: The key growth periods obtained and accurately determined in step D are calibrated onto the smooth NDVI curve generated in step G to form a standard reference curve with maize growth information.

5. The method according to claim 4, characterized in that, Step 4 includes the following steps: Step 1: Acquire remote sensing data of the target area: For the target area in the current year that needs to be predicted for the growing season, acquire Sentinel-2 satellite time series images of the entire growing season; perform standardized preprocessing such as radiometric calibration, atmospheric correction, and orthorectification on the images to ensure data quality and consistency; Step J: Calculate the vegetation index sequence for the farm area: Calculate the normalized vegetation index for each Sentinel-2 image after preprocessing in Step I. Step K: Matching and fitting with the standard reference curve: Match and fit the NDVI time series data obtained in step J with the standard reference curve of the calibrated maize phenological period generated in step H. In the formula, This represents the length of time corn takes to grow. The corn curve function is obtained by fitting Sentinel-2 image data at the scale. The initial reference curve function represents the set of corn reference curve samples. This represents the time shift of corn during its growth period, and the value is within ±30 days of a time interval of 1 day. and The fitting coefficient represents the difference between the two curves; Step L involves using satellite remote sensing vegetation index reconstruction to preliminarily predict the maize growth period in the farm area.

6. The method according to claim 5, characterized in that, Step 5 includes the following steps: Step M: Obtain early phenological observations for the current year: In the early growing season of the current year, use ground cameras deployed in the target area to continuously collect time-series images of maize; input these images into the maize growth period recognition model trained and saved in step C, the model automatically analyzes and identifies at least one early growth period that has actually occurred, and records the actual calendar date of its occurrence. Step N: Calculate the systematic bias of the preliminary forecast: From the preliminary forecast results generated in Step L, extract the preliminary forecast dates for the same fertility period of these same pixels, and calculate their forecast bias: ΔT = Model inference date - Preliminary forecast date, which is used as the forecast bias ΔT of the forecast model for the current year in this region; ΔT quantifies the overall degree of advance or lag of the preliminary forecast based on the historical model in this year.

7. The method according to claim 6, characterized in that, Step 6 includes the following steps: Step O, dynamically correct the prediction of the later fertility period: apply the regional average prediction deviation ΔT calculated in step N to the preliminary prediction dates of all later fertility periods that have not yet actually occurred, obtained in step L; the correction formula is: the final prediction date of a certain fertility period after correction = the preliminary prediction date + ΔT; this operation is equivalent to performing a one-time overall translation correction on the entire later phenological calendar of the preliminary prediction based on the actual situation in the early stage, so as to eliminate systematic errors between years; Step P, Iterative Optimization and Final Output: Steps M, N, and O can be executed iteratively within a growing season; whenever new ground observations of the growth period are obtained, ΔT is recalculated, and the predictions for all remaining non-growing periods are immediately updated using the latest ΔT, achieving incremental optimization of the prediction results; finally, the predicted dates of each later growth period after dynamic correction are integrated with the actual dates of the observed earlier growth periods to form and output a complete and high-precision spatial prediction map of the entire maize growth period for the target area.

8. The method according to claim 1, characterized in that: In step 1, the corn growth period identification model is a ResNet-50 deep learning image recognition model, which can classify and identify key growth period images of corn seedling stage, three-leaf stage, seven-leaf stage, tasseling stage, milk stage, and maturity stage from time-series images captured by ground cameras.

9. The method according to claim 1, characterized in that: Suppose that, at the high-resolution Sentinel-2 image data scale, the maize time series curve of a certain pixel has the following relationship with the extracted standard reference curve: In the formula, This represents the length of time corn takes to grow. This represents the vegetation index curve function fitted at the scale of Sentinel-2 image data. The initial reference curve function represents the set of corn reference curve samples. For a single cell, This is the corresponding reference curve; This represents the time shift of corn during its growth period, and the value is within ±30 days of a time interval of 1 day. and The fitting coefficient represents the difference between the two curves.

10. The method according to claim 1, characterized in that: The vegetation index mentioned is the Normalized Difference Vegetation Index (NDVI), which is calculated using the following formula: Wherein, NIR is the near-infrared reflectance and RED is the red reflectance; during the entire NDVI time series curve reconstruction process, the number of days the curve is shifted due to phenological differences will be output synchronously; the NDVI time series of each maize pixel is matched and fitted with the reference curve to obtain the number of days of shift for each pixel, and then the growth period of each maize pixel is obtained based on the number of days of shift and the growth period information contained in the reference curve.

Images