A corn planting area extraction method and system based on deep learning

Abstract

This invention relates to a method and system for extracting maize planting area based on deep learning, comprising: determining multiple vegetation indices based on acquired optical images and determining polarization feature indices based on radar images; constructing a set of statistical feature images corresponding to the entire growing season based on the multiple vegetation indices and polarization feature indices; inputting the median EVI and median RVI temporal images in the set into a temporal convolutional network (TCN) to extract target phenological features; aligning, stitching, and standardizing the target phenological features with the mean NDVI image, maximum NDRE image, and standard deviation RVI image in the set according to pixels to obtain a target multidimensional classification feature matrix; inputting the target multidimensional classification feature matrix into a prototype network classifier for preliminary identification of maize planting areas; and performing connected component analysis, inverse hole filling, and quantified statistical area based on the preliminary identification results to obtain the total maize planting area of ​​the target region.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of remote sensing image information extraction and deep learning technology, and in particular to a method and system for extracting corn planting area based on deep learning. Background Technology

[0002] Currently, traditional methods for extracting maize planting area mostly rely on manual visual interpretation or rule-based classification based on phenological characteristics. These methods are not only time-consuming and labor-intensive, but their phenological feature extraction is also highly dependent on expert experience and manually set rules, making them difficult to adapt to complex and ever-changing agricultural environments and climatic conditions, thus limiting extraction accuracy and efficiency. Meanwhile, shallow classifiers (such as support vector machines and random forests) lack sufficient feature representation capabilities when processing high-dimensional remote sensing data, making it difficult to fully extract deeper information from the data, further restricting the accuracy of maize planting area extraction. Furthermore, while cloud computing platforms provide powerful computing resources for large-scale remote sensing data processing with the rapid development of cloud computing technology, existing cloud computing platforms often cannot natively support the training and optimization of deep learning models, leading to technical bottlenecks in the application of deep learning in the field of agricultural remote sensing.

[0003] Therefore, in view of the shortcomings of the existing technologies, such as reliance on manual rules for phenological feature extraction, insufficient feature expression capabilities of shallow classifiers, and the inability of cloud computing platforms to natively support deep learning training, there is an urgent need to develop a new method for extracting corn planting area that is efficient, accurate, and adaptable to the cloud computing environment. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for extracting corn planting area based on deep learning, addressing the above-mentioned deficiencies.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for extracting corn planting area based on deep learning, the method comprising:

[0006] S1. Acquire radar and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature indices include Radar Vegetation Index (RVI).

[0007] S2. Based on the multiple vegetation indices and the polarization characteristic index, construct a statistical characteristic image set corresponding to the entire growing season. The statistical characteristic image set includes EVI median time series image, RVI median time series image, NDVI mean image, NDRE maximum image, and RVI standard deviation image.

[0008] S3. Input the EVI median temporal image and the RVI median temporal image into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism.

[0009] S4. Align, stitch, and standardize the target phenological features with the NDVI mean image, NDRE maximum value image, and RVI standard deviation image according to pixels to obtain the target multidimensional classification feature matrix.

[0010] S5. Input the target multidimensional classification feature matrix into the prototype network classifier to perform preliminary identification of maize planting areas. Non-maize negative samples are generated by multi-strategy phenological perturbation of maize positive samples.

[0011] S6. Based on the preliminary identification results, perform connected component analysis, reverse filling of gaps, and quantitative statistical area to obtain the total corn planting area of ​​the target region.

[0012] In one embodiment, in step S3, the target phenological characteristics include the growing season start date SOS, the growth peak date POS, the growing season end date EOS, and the growth amplitude Amplitude.

[0013] In one embodiment, in step S3, the temporal convolutional network (TCN) includes an input layer, a multi-scale dilated causal convolutional layer, a temporal attention layer, a global average pooling layer, and a fully connected output layer connected in sequence, wherein:

[0014] The input tensor dimension of the input layer is [number of batches, 12, 2], where 12 represents the number of time steps and 2 represents the dual channels of the median time series images of EVI and RVI.

[0015] The multi-scale dilated causal convolutional layer includes four convolutional blocks with dilation rates of 1, 2, 4, and 8, respectively. Each convolutional block adopts a causal filling method, and the corresponding receptive fields cover approximately 45 days, 75 days, 135 days, and the entire growth season, respectively.

[0016] The time attention layer applies time dimension attention weights to the multi-scale feature tensors based on the time step correlation of the multi-scale feature tensors, so as to dynamically focus on key temporal information at different growth stages.

[0017] After the fully connected output layer is activated by 64-dimensional ReLU, the normalized phenological parameters are output by the 4-dimensional sigmoid activated output layer.

[0018] In one embodiment, in step S3, the temporal convolutional network TCN employs a weighted mean square error loss function in the following form:

[0019] ;

[0020] Where N represents the number of samples, The predicted SOS value represents the start date of the growing season. This indicates the actual value of SOS representing the start date of the growing season. This represents the predicted value for the peak growth date (POS). This represents the actual value of the peak growth date POS. This represents the predicted value of EOS for the end date of the growing season. The actual value of EOS representing the end date of the growing season. This represents the predicted value of growth amplitude (Amplitude). This represents the actual value of the growth amplitude (Amplitude).

[0021] In one embodiment, in step S5, the multi-strategy phenological perturbation includes at least one of the following:

[0022] (1) Phenological window offset perturbation strategy: Random sampling is performed from positive maize samples, and a random offset perturbation amount exceeding the normal maize phenological window is applied from the SOS (start date of growing season) and EOS (end date of growing season) feature dimensions of the sampled samples. ,in, The sampling range is [30 days, 60 days], to generate perturbation negative samples that distort phenological rhythms;

[0023] (2) Feature dimension random permutation strategy: For the positive samples of maize, the growth season start date SOS, growth peak date POS and growth season end date EOS dimension features, random dimension permutation is performed between samples. By breaking the temporal logical constraints between phenological features, perturbation negative samples with anomalous temporal relationships are generated.

[0024] (3) Mixed coefficient constrained interpolation strategy: Randomly sample two groups of samples from the positive samples of maize. and and with a preset mixing coefficient After linear interpolation, random perturbations within the range of [-40 days, 40 days] are superimposed on the SOS feature dimension from the start date of the growing season of the sample to generate perturbed negative samples with partially fused phenological features and abnormal SOS perturbations.

[0025] In one embodiment, in step S5, the prototype network classifier uses a phenological perception dual-component contrastive loss function, which consists of two parts: a maize positive sample clustering loss function and a non-maize negative sample rejection loss function. The calculation formula for the maize positive sample clustering loss function is as follows:

[0026] ;

[0027] in, Indicates the number of positive samples. This represents the cosine similarity between positive sample i and other positive samples k. This represents the weighting coefficient for positive sample i. This represents the feature vector of a positive sample i. The L2-normalized vector representing the mean of the batch positive sample embedding vectors;

[0028] The formula for calculating the non-corn negative sample rejection loss function is as follows:

[0029] ;

[0030] ;

[0031] in, Indicates the number of negative samples. This represents the cosine similarity between negative sample j and other negative samples l. This represents the weighting coefficient for negative sample j. Indicates boundary values, Indicates the basic boundary value. This represents the adjustment step size coefficient for boundary values. This indicates the current number of training steps. This represents the distance between the aggregated features of negative sample j and positive sample j. This represents the distance between positive samples and their aggregated features.

[0032] In one embodiment, step S6, which involves performing connected component analysis, reverse filling of gaps, and quantified statistical area based on the preliminary identification results to obtain the total corn planting area of ​​the target region, includes:

[0033] S61. Based on the preliminary identification results, perform connected component analysis, morphological processing, and difference analysis to obtain the void region to be verified.

[0034] S62. For the void area to be verified, the maximum normalized red edge vegetation index (maxNDRE) during the peak growing season is used for spectral secondary verification and reverse filling to obtain the complete area result after filling.

[0035] S63. Based on the results of the filled complete area, perform area quantification statistics to obtain the total corn planting area of ​​the target area.

[0036] Secondly, this application discloses a deep learning-based corn planting area extraction system, which includes a multi-source data index calculation module, a statistical feature image construction module, a phenological feature extraction module, a classification feature matrix generation module, a preliminary corn planting area classification identification module, and a total corn planting area quantification and statistics module, wherein:

[0037] The multi-source data index calculation module is used to acquire radar images and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature index includes Radar Vegetation Index (RVI).

[0038] The statistical feature image construction module is used to construct a set of statistical feature images corresponding to the entire growing season based on the multiple vegetation indices and the polarization feature index. The set of statistical feature images includes EVI median time series images, RVI median time series images, NDVI mean images, NDRE maximum images, and RVI standard deviation images.

[0039] The phenological feature extraction module is used to input the median EVI time-series image and the median RVI time-series image into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism.

[0040] The classification feature matrix generation module is used to align, stitch, and standardize the target phenological features with the NDVI mean image, the NDRE maximum value image, and the RVI standard deviation image according to pixels in sequence to obtain the target multidimensional classification feature matrix.

[0041] The preliminary identification module for corn planting area classification is used to input the target multidimensional classification feature matrix into the prototype network classifier for preliminary identification of corn planting areas. Non-corn negative samples are generated by multi-strategy phenological perturbation of corn positive samples.

[0042] The total corn planting area quantification and statistics module is used to perform connected component analysis, reverse filling of gaps, and quantification and statistics of area based on the preliminary identification results, so as to obtain the total corn planting area of ​​the target area.

[0043] Thirdly, this application discloses a readable storage medium storing a computer program that, when executed by a processor, implements a deep learning-based method for extracting corn planting area as described in any of the preceding claims.

[0044] The beneficial effects of this invention are as follows: by fusing radar and optical image data, comprehensively integrating multiple vegetation indices and polarization feature indices, constructing a rich set of statistical feature images, using a temporal convolutional network and combining it with a temporal attention mechanism to accurately extract phenological features, and then constructing a multi-dimensional classification feature matrix to achieve effective identification of maize planting areas. Finally, through connected component analysis and hole filling operations, the area is accurately calculated, significantly improving the accuracy of maize planting area identification and the reliability of area calculation, providing strong support for maize planting monitoring and management. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a deep learning-based method for extracting corn planting area disclosed in this invention.

[0046] Figure 2 This is a schematic diagram of the structure of a corn planting area extraction system based on deep learning disclosed in this invention;

[0047] Figure 3 This is a schematic diagram of the structure of a readable storage medium disclosed in this invention. Detailed Implementation

[0048] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0049] like Figure 1 As shown, this application discloses a method for extracting corn planting area based on deep learning, the method comprising:

[0050] Step S1: Acquire radar and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature indices include Radar Vegetation Index (RVI).

[0051] Specifically, this application acquires radar and optical images of the target area based on the cloud computing platform Google Earth Engine (GEE). The radar images include Sentinel-1 synthetic aperture radar (SAR) images, and the optical images include Sentinel-2 optical images. Specifically, the Sentinel-1 SAR images utilize interferometric wide-swath (IW) ground distance detection products. These products undergo preprocessing via the Sentinel-1 toolkit, including thermal noise removal, radiometric calibration, and terrain correction. They include VV and VH dual-polarization bands, have a spatial resolution of 10 m, a revisit period of 5 days, and all-weather imaging capabilities, unaffected by cloud or rain conditions. This allows the application to compensate for data loss caused by cloud cover in optical images during critical growth periods. Sentinel-2 optical imagery uses atmospherically corrected L2A-level surface reflectance products. It acquires visible light bands B2 (blue light, center wavelength 492nm), B4 (red light, center wavelength 665nm), and the near-infrared broadband band B8 (center wavelength 842nm) with a spatial resolution of 10m, as well as the red-edge band B5 (center wavelength 705nm) and the near-infrared narrow band B8A (center wavelength 865nm) with a spatial resolution of 20m. These bands form the core data foundation for subsequent calculations of vegetation indices such as NDVI mean imagery, EVI median time-series imagery, and NDRE maximum imagery.

[0052] Specifically, the Normalized Difference Vegetation Index (NDVI) is sensitive to changes in vegetation cover density and is a fundamental indicator for characterizing the growth status of maize. The formula for calculating this index is: NDVI = (B8 − B4) / (B8 + B4).

[0053] Specifically, the Enhanced Vegetation Index (EVI) corrects for the effects of soil background reflectance and atmospheric aerosol scattering by introducing the blue light band. It is less prone to saturation during the high vegetation cover stage (i.e., from the corn tasseling stage to the grain-filling stage) and is more sensitive to changes in canopy biomass. The formula for calculating this index is: .

[0054] Specifically, the Normalized Red Edge Vegetation Index (NDRE) utilizes the high sensitivity of the red edge band to changes in chlorophyll content. Even under conditions of high leaf area index, it can accurately reflect the dynamic changes in chlorophyll content in the maize canopy. Furthermore, it shows significant differences between maize and other crops of the same season, such as soybeans and sorghum, during the grain-filling stage. The calculation formula for this index is: NDRE=(B8A-B5) / (B8A+B5), where B8A is the near-infrared narrow band (center wavelength 865nm) and B5 is the first red edge band (center wavelength 705nm).

[0055] Specifically, the Radar Vegetation Index (RVI) characterizes the volume scattering properties of the vegetation canopy by the ratio of the VH to VV polarization backscattering coefficients. It increases significantly with biomass from the jointing stage to the tasseling stage of maize, effectively distinguishing between the crop growth period and bare land conditions. The formula for calculating this index is: .

[0056] Specifically, this application will calculate the standard deviation RVI_std of the time series values ​​of each pixel during the entire growing season. It can characterize the temporal fluctuation range of the scattering characteristics of ground objects and can serve as an auxiliary statistical feature to distinguish corn from ground objects with stable scattering characteristics during the growing season (such as buildings and evergreen vegetation).

[0057] In one embodiment, this application further utilizes the scene classification layer (SCL) inherent in Sentinel-2 optical imagery to construct an effective pixel mask by filtering pixels categorized as vegetation and bare soil, and removing clouds, cloud shadows, water bodies, and other invalid pixels. Furthermore, based on this effective pixel mask, this application performs masking processing on an image-by-image basis across the entire growing season image set, dividing the surface reflectance values ​​by 10000 to convert them into standardized reflectance, thereby obtaining a reliable optical time-series dataset.

[0058] In one embodiment, this application also uses the nearest neighbor interpolation method to resample the 20m resolution B5 and B8A bands to 10m to ensure that the spatial resolution of all optical bands is consistent, providing a unified spatial benchmark for subsequent multi-band vegetation index calculation and subsequent feature fusion.

[0059] Step S2: Based on the multiple vegetation indices and the polarization characteristic index, construct a statistical characteristic image set corresponding to the entire growing season. The statistical characteristic image set includes EVI median time series image, RVI median time series image, NDVI mean image, NDRE maximum image, and RVI standard deviation image.

[0060] Specifically, this application sets the time interval step to 15 days, dividing the entire maize growing season into 12 equally spaced time windows, numbered sequentially from t1 to t12. Within each time window, this application performs median synthesis on all valid images within that window to obtain representative feature values ​​for that time period. Furthermore, for time windows with excessive cloud cover leading to insufficient effective observations, this application employs a two-step smoothing strategy for reconstruction. Specifically, firstly, this application performs linear interpolation using feature values ​​from adjacent valid time windows to fill consecutive missing data gaps; then, it applies a Savitzky-Golay filter (window length 5 time steps, polynomial order 3) to the complete time series data for smoothing to remove high-frequency noise from the time series curves and retain low-frequency signals of phenological changes.

[0061] Specifically, after completing the time-series composite, this application will export the following 27-band feature images into multi-band GeoTIFF format in the Google Earth Engine (GEE) cloud computing platform, with a spatial resolution of 10m, while retaining the original geographic coordinate information:

[0062] (1) Bands 1 to 12: EVI median time series data, corresponding to the composite results of EVI median values ​​for 12 15-day time windows, denoted as EVI_t1 to EVI_t12;

[0063] (2) Bands 13 to 24: RVI median time series data, corresponding to the composite results of RVI median values ​​for 12 15-day time windows, denoted as RVI_t1 to RVI_t12;

[0064] (3) Band 25: Growing season average NDVI (mean_NDVI), which is obtained by averaging the NDVI values ​​of all valid images throughout the growing season and is used to reflect the average vegetation cover density of maize throughout the growing season.

[0065] (4) Band 26: Maximum NDRE during the growing season (max_NDRE), which is the maximum value selected from all valid images during the entire growing season. This value can effectively characterize the peak chlorophyll content during maize growth.

[0066] (5) Band 27: RVI standard deviation during the growing season (RVI_std). It is obtained by first calculating the RVI values ​​of all valid images during the entire growing season and then using the standard deviation calculation formula. It reflects the dispersion of RVI values ​​during the maize growing season and can indirectly reflect the changes and fluctuations in the scattering characteristics of the vegetation canopy.

[0067] After exporting, this application will download the resulting GeoTIFF file to the local deep learning computing environment, and all subsequent deep learning training and inference steps will be executed locally. The minimum configuration requirements for the local environment are 32GB of system memory and 8GB of GPU memory, Python 3.8 or later, and PyTorch 1.10 or later.

[0068] Step S3: The median temporal image of EVI and the median temporal image of RVI are input into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism.

[0069] Specifically, to address the shortcomings of traditional dynamic thresholding methods in extracting phenological weights, which rely on expert experience and are difficult to adapt to different regions and interannual variations, this application introduces a temporal convolutional network (TCN) combined with a temporal attention mechanism to automatically extract target phenological feature parameters in a data-driven manner, including the start date of the growing season (SOS), the peak date of growth (POS), the end date of the growing season (EOS), and the growth amplitude (Amplitude).

[0070] In one embodiment, the temporal convolutional network TCN used in this application uses maize samples with accurate phenological node annotations as supervision signals and is trained using a weighted mean square error loss function, which enables the network to more accurately capture the changing patterns of maize phenological characteristics and effectively improve the accuracy and stability of phenological node identification.

[0071] Step S4: Align, stitch together and standardize the target phenological features with the NDVI mean image, NDRE maximum value image and RVI standard deviation image according to pixels to obtain the target multidimensional classification feature matrix.

[0072] Specifically, alignment is to ensure that each image data is strictly corresponding in spatial location, so that the subsequent stitched feature matrix can accurately reflect multiple feature information of the same geographical location; stitching aims to integrate phenological features, vegetation indices and other information carried by different types of images to construct a more comprehensive feature description; standardization is to eliminate the influence of differences in units and numerical ranges between different features, and improve the convergence speed and classification accuracy of the classification model.

[0073] Step S5: Input the target multidimensional classification feature matrix into the prototype network classifier for preliminary identification of maize planting areas. Non-maize negative samples are generated by multi-strategy phenological perturbation of maize positive samples.

[0074] Specifically, this application will extract the 7-dimensional standardized feature vectors corresponding to the positive sample points of corn from the target multidimensional classification feature matrix generated in step S4 through geographic coordinate indexing to form a training set with no less than 50 samples.

[0075] Specifically, the feature encoder of the prototype network classifier consists of three fully connected layers. The first layer is a 64-dimensional fully connected layer (ReLU activation) → batch normalization layer → dropout layer (dropout rate 0.2). The second layer is a 32-dimensional fully connected layer (ReLU activation) → batch normalization layer. The third layer is a 16-dimensional fully connected layer (linear activation). Finally, the 16-dimensional embedding vector output by the feature encoder is projected onto the unit hypersphere after L2 normalization to ensure the geometric validity of the cosine similarity calculation.

[0076] In one embodiment, the feature encoder is trained using the Adam optimizer with a learning rate of 0.001, a gradient clipping threshold of 1.0, a batch size of 64 (each batch contains an equal number of positive samples and 32 negative samples generated in real time), and 300 training epochs.

[0077] In one embodiment, after training, this application inputs all positive corn samples into the trained encoder, and after averaging the 16-dimensional embedding vectors of all samples, performs L2 normalization to obtain the final corn prototype vector P_corn. Finally, this application performs block-wise forward inference on the effective pixels of the entire study area, with each batch not exceeding 100,000 pixels, and calculates the embedding vector for each pixel. Cosine distance between the corn prototype vector P_corn and the corn prototype vector: .

[0078] Specifically, this application classifies pixels whose cosine distance is less than a preset judgment threshold θ (in this embodiment, θ = 0.35, determined by maximizing F1 on the verification sample) as corn (i.e., output 1), and otherwise as non-corn (i.e., output 0). After inference, the binary result is reshaped to the same spatial size as the original image to obtain a preliminary spatial distribution map of the corn planting area.

[0079] Step S6: Based on the preliminary identification results, perform connected component analysis, reverse filling of gaps, and quantitative statistical area to obtain the total corn planting area of ​​the target region.

[0080] Specifically, this application will perform connected component analysis to denoise based on the preliminary classification results, and use the Normalized Red Edge Vegetation Index (NDRE) to perform spectral secondary verification and reverse filling of the internal void regions, so as to improve the completeness and accuracy of maize planting area identification and provide a reliable foundation for subsequent accurate statistics of maize planting area.

[0081] As can be seen from the above, the deep learning-based method for extracting maize planting area disclosed in this application integrates radar and optical image data, comprehensively integrates multiple vegetation indices and polarization feature indices, constructs a rich set of statistical feature images, uses a temporal convolutional network and combines it with a temporal attention mechanism to accurately extract phenological features, and then constructs a multi-dimensional classification feature matrix to effectively identify maize planting areas. Finally, the area is accurately calculated through connected component analysis and hole filling, which significantly improves the accuracy of maize planting area identification and the reliability of area calculation, providing strong support for maize planting monitoring and management.

[0082] In one embodiment, in step S3, the target phenological characteristics include the growing season start date SOS, the growth peak date POS, the growing season end date EOS, and the growth amplitude Amplitude.

[0083] In one embodiment, in step S3, the temporal convolutional network (TCN) includes an input layer, a multi-scale dilated causal convolutional layer, a temporal attention layer, a global average pooling layer, and a fully connected output layer connected in sequence, wherein:

[0084] The input tensor dimension of the input layer is [number of batches, 12, 2], where 12 represents the number of time steps and 2 represents the dual channels of the median time series images corresponding to EVI and RVI.

[0085] Specifically, this application uses EVI and RVI median time-series images as inputs simultaneously to fully utilize the complementarity of optical and radar data. The EVI median time-series image can reflect the temporal dynamics of maize canopy photosynthetic activity, while the RVI median time-series image can reflect the temporal dynamics of canopy body scattering characteristics. The combination of the two can more accurately locate phenological turning points.

[0086] The multi-scale dilated causal convolutional layer includes four convolutional blocks with dilation rates of 1, 2, 4, and 8, respectively. Each convolutional block adopts a causal filling method, and the corresponding receptive fields cover approximately 45 days, 75 days, 135 days, and the entire growth season, respectively.

[0087] Specifically, the structure of each convolutional block is as follows: a one-dimensional causal convolutional layer (kernel size 3, output channels 64, causal padding) → batch normalization layer → ReLU activation function → Dropout layer (dropout rate 0.1). Causal padding ensures that the output at any time step depends only on the input at that time step and previous inputs, conforming to the causal constraints of phenological time series. The effective receptive fields corresponding to the four expansion rates cover approximately 45 days, 75 days, 135 days, and the entire growing season, respectively, enabling hierarchical parallel perception of multi-timescale phenological patterns such as short-term seedling emergence and jointing, mid-term tasseling and grain filling, and long-term complete reproductive rhythms. The outputs of the four convolutional blocks are concatenated along the channel dimension to obtain a multi-scale feature tensor with dimensions [batch number, 12, 256].

[0088] The time attention layer applies time dimension attention weights to the multi-scale feature tensors based on the time step correlation of the multi-scale feature tensors, so as to dynamically focus on key temporal information at different growth stages.

[0089] Specifically, the temporal attention layer applies temporal dimension attention weights to the multi-scale feature tensor. It calculates the attention score for each time step using a two-layer fully connected network (256-dimensional → 64-dimensional tanh activation, 64-dimensional → 1-dimensional linear activation). After softmax normalization, an attention weight vector of [batch number, 12, 1] is obtained. This vector is then multiplied sequentially with the multi-scale feature tensor and weighted and summed in the temporal dimension, outputting a context vector of [batch number, 256]. This mechanism enables the model to automatically identify and strengthen the contributions of key growth stages such as seedling emergence, tasseling, grain filling, and maturity, while automatically reducing the weight of the residual cloud pollution time step.

[0090] After the fully connected output layer is activated by 64-dimensional ReLU, the normalized phenological parameters are output by the 4-dimensional sigmoid activated output layer.

[0091] Specifically, after the fully connected output layer is activated by 64-dimensional ReLU, the output layer is activated by 4-dimensional sigmoid to output normalized phenological parameters, corresponding to SOS / 365, POS / 365, EOS / 365 and Amplitude. During inference, the first three are multiplied by 365 to restore the actual accumulated days of the year.

[0092] In one embodiment, in step S3, the temporal convolutional network TCN employs a weighted mean square error loss function in the following form:

[0093] ;

[0094] Where N represents the number of samples, The predicted SOS value represents the start date of the growing season. This indicates the actual value of SOS representing the start date of the growing season. This represents the predicted value for the peak growth date (POS). This represents the actual value of the peak growth date POS. This represents the predicted value of EOS for the end date of the growing season. The actual value of EOS representing the end date of the growing season. This represents the predicted value of growth amplitude (Amplitude). This represents the actual value of the growth amplitude (Amplitude).

[0095] Specifically, the innovation of this application lies in proposing a specific weighted strategy for crop phenology extraction tasks to construct the weighted mean squared error loss function of the Temporal Convolutional Network (TCN). In the agricultural field, accurate identification of maize planting areas depends on a precise grasp of maize phenological characteristics. The start of the growing season (SOS) and end of the growing season (EOS) are crucial for defining the maize growth cycle, directly related to key agricultural activities such as sowing and harvesting, and therefore have higher importance for identifying maize planting areas. Therefore, in the loss function, this application sets the loss weights for SOS and EOS to 2.0 to emphasize the model's prediction accuracy for these key nodes. Furthermore, since the peak growth date (POS) reflects the most vigorous stage of maize growth, although its importance is slightly lower than SOS and EOS, it is still an important component of maize phenological characteristics, so the POS weight is set to 1.0. Additionally, since growth amplitude reflects the intensity change of maize growth, its direct impact on maize planting area identification is relatively small compared to the previous features, so the amplitude weight is set to 0.5.

[0096] In summary, this application uses this training method to enable the model to pay more attention to the phenological features that are more critical for identifying maize planting areas, thereby improving the model's accuracy in identifying maize planting areas.

[0097] In one embodiment, in step S5, the multi-strategy phenological perturbation includes at least one of the following:

[0098] (1) Phenological window offset perturbation strategy: Random sampling is performed from positive maize samples, and a random offset perturbation amount exceeding the normal maize phenological window is applied from the SOS (start date of growing season) and EOS (end date of growing season) feature dimensions of the sampled samples. ,in, The sampling range is [30 days, 60 days], to generate perturbation negative samples that distort phenological rhythms.

[0099] (2) Feature dimension random permutation strategy: For the positive samples of maize, the growth season start date SOS, growth peak date POS and growth season end date EOS dimension features are randomly permuted between samples. By breaking the temporal logic constraints between phenological features, perturbation negative samples with abnormal temporal relationships are generated.

[0100] (3) Mixed coefficient constrained interpolation strategy: Randomly sample two groups of samples from the positive samples of maize. and and with a preset mixing coefficient After linear interpolation, random perturbations within the range of [-40 days, 40 days] are superimposed on the SOS feature dimension from the start date of the growing season of the sample to generate perturbed negative samples with partially fused phenological features and abnormal SOS perturbations.

[0101] In one embodiment, in step S5, the prototype network classifier uses a phenological perception dual-component contrastive loss function, which consists of two parts: a maize positive sample clustering loss function and a non-maize negative sample rejection loss function. The calculation formula for the maize positive sample clustering loss function is as follows:

[0102] ;

[0103] in, Indicates the number of positive samples. This represents the cosine similarity between positive sample i and other positive samples k. This represents the weighting coefficient for positive sample i. This represents the feature vector of a positive sample i. This represents the L2 normalized vector representing the mean of the batch of positive sample embedding vectors.

[0104] The formula for calculating the non-corn negative sample rejection loss function is as follows:

[0105] ;

[0106] ;

[0107] in, Indicates the number of negative samples. This represents the cosine similarity between negative sample j and other negative samples l. This represents the weighting coefficient for negative sample j. Indicates boundary values, Indicates the basic boundary value. This represents the adjustment step size coefficient for boundary values. This indicates the current number of training steps. This represents the distance between the aggregated features of negative sample j and positive sample j. This represents the distance between positive samples and their aggregated features.

[0108] In one embodiment, step S6, which involves performing connected component analysis, reverse filling of gaps, and quantified statistical area based on the preliminary identification results to obtain the total corn planting area of ​​the target region, includes:

[0109] Step S61: Based on the preliminary identification results, perform connected component analysis, morphological processing, and difference analysis to obtain the void region to be verified.

[0110] Specifically, this application uses an 8-connected structure to label the connected components of the classification results, counts the number of pixels in each connected component, and removes isolated noise patches with a number of pixels less than a first preset threshold to obtain a denoised corn classification image.

[0111] Specifically, this application performs morphological closing operations (specifically, dilation followed by erosion, iterated twice each, using 3×3 all-1 structuring elements) on the denoised corn classification image to obtain a reference mask with its internal holes filled. Then, this application performs a difference operation between the reference mask and the denoised corn classification image, extracting regions in the reference mask that belong to corn pixels but not to non-corn pixels in the original denoised classification image, as candidate hole masks. Further, this application labels the candidate hole masks with 8-connected components, defining connected regions with an area less than a second preset threshold and all pixels within a preset cultivated land mask range as holes to be verified, while the remaining connected regions retain the original non-corn classification result.

[0112] In one embodiment, the first preset threshold is 5 pixels, the second preset threshold is 10 pixels, and the preset farmland mask is constructed based on pre-acquired land cover data.

[0113] Step S62: For the void area to be verified, the maximum normalized red edge vegetation index (maxNDRE) during the peak growing season is used for spectral secondary verification and reverse filling to obtain the complete area result after filling.

[0114] Specifically, this application extracts the maximum normalized red edge vegetation index (maxNDRE) of all pixels in each cavity region to be verified during the peak growing season. If the maxNDRE of a cavity region to be verified is greater than the preset spectral threshold and the cavity region is located within the preset cultivated land mask range, the cavity is determined to be a misclassified region caused by missing image phase or local quality fluctuation. This application sets the pixel value of all its pixels to 1. Otherwise, the original non-corn classification result remains unchanged.

[0115] In one embodiment, the preset spectral threshold is set to 0.42.

[0116] Step S63: Based on the results of the filled complete area, perform area quantification statistics to obtain the total corn planting area of ​​the target area.

[0117] Specifically, this application will count the total number of corn pixels in the obtained corn planting distribution map, and combine it with the spatial resolution of the image (10 m) to calculate the total corn planting area of ​​the target area according to the following formula: planting area = total number of corn pixels × area of ​​a single pixel (100 m²), thereby obtaining the total corn planting area of ​​the target area and completing the extraction process.

[0118] Please refer to Figure 2 This application discloses a deep learning-based system for extracting maize planting area. The system includes a multi-source data index calculation module, a statistical feature image construction module, a phenological feature extraction module, a classification feature matrix generation module, a preliminary classification identification module for maize planting areas, and a quantitative statistical module for total maize planting area.

[0119] The multi-source data index calculation module is used to acquire radar images and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature indices include Radar Vegetation Index (RVI).

[0120] The statistical feature image construction module is used to construct a set of statistical feature images corresponding to the entire growing season based on the multiple vegetation indices and the polarization feature index. The set of statistical feature images includes EVI median time series images, RVI median time series images, NDVI mean images, NDRE maximum images, and RVI standard deviation images.

[0121] The phenological feature extraction module is used to input the median EVI temporal image and the median RVI temporal image into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism.

[0122] The classification feature matrix generation module is used to align, stitch together, and standardize the target phenological features with the NDVI mean image, NDRE maximum value image, and RVI standard deviation image according to pixels in sequence to obtain the target multidimensional classification feature matrix.

[0123] The preliminary identification module for corn planting area classification is used to input the target multidimensional classification feature matrix into the prototype network classifier for preliminary identification of corn planting areas. Non-corn negative samples are generated by multi-strategy phenological perturbation of corn positive samples.

[0124] The total corn planting area quantification and statistics module is used to perform connected component analysis, reverse filling of gaps, and quantification and statistics of area based on the preliminary identification results, so as to obtain the total corn planting area of ​​the target area.

[0125] In one embodiment, the above modules are also used to implement a deep learning-based method for extracting corn planting area as described in any of the preceding claims, and this application does not limit this method.

[0126] As can be seen from the above, the corn planting area extraction system disclosed in this application, based on deep learning, integrates radar and optical image data, comprehensively integrates multiple vegetation indices and polarization feature indices, constructs a rich set of statistical feature images, accurately extracts phenological features using a temporal convolutional network combined with a temporal attention mechanism, and then constructs a multi-dimensional classification feature matrix to effectively identify corn planting areas. Finally, it accurately calculates the area through connected component analysis and hole filling operations, significantly improving the accuracy of corn planting area identification and the reliability of area calculation, providing strong support for corn planting monitoring and management.

[0127] Thirdly, this application discloses a readable storage medium storing a computer program that, when executed by a processor, implements a deep learning-based method for extracting corn planting area as described in any of the preceding claims.

[0128] As can be seen from the above, the readable storage medium disclosed in this application integrates radar and optical image data, comprehensively integrates multiple vegetation indices and polarization feature indices, constructs a rich set of statistical feature images, accurately extracts phenological features using a temporal convolutional network combined with a temporal attention mechanism, and then constructs a multi-dimensional classification feature matrix to effectively identify maize planting areas. Finally, it accurately calculates the area through connected component analysis and hole filling operations, significantly improving the accuracy of maize planting area identification and the reliability of area calculation, providing strong support for maize planting monitoring and management.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting corn planting area based on deep learning, characterized in that, The method includes: S1. Acquire radar and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature indices include Radar Vegetation Index (RVI). S2. Based on the multiple vegetation indices and the polarization characteristic index, construct a statistical characteristic image set corresponding to the entire growing season. The statistical characteristic image set includes EVI median time series image, RVI median time series image, NDVI mean image, NDRE maximum image, and RVI standard deviation image. S3. Input the EVI median temporal image and the RVI median temporal image into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism. S4. Align, stitch, and standardize the target phenological features with the NDVI mean image, NDRE maximum value image, and RVI standard deviation image according to pixels to obtain the target multidimensional classification feature matrix. S5. Input the target multidimensional classification feature matrix into the prototype network classifier to perform preliminary identification of maize planting areas. Non-maize negative samples are generated by multi-strategy phenological perturbation of maize positive samples. S6. Based on the preliminary identification results, perform connected component analysis, reverse filling of gaps, and quantitative statistical area to obtain the total corn planting area of ​​the target region.

2. The corn planting area extraction method according to claim 1, characterized by, In step S3, the target phenological characteristics include the start date of the growing season (SOS), the peak growth date (POS), the end date of the growing season (EOS), and the growth amplitude (Amplitude).

3. The method for extracting corn planting area according to claim 1, characterized in that, In step S3, the temporal convolutional network (TCN) comprises, in sequence, an input layer, a multi-scale dilated causal convolutional layer, a temporal attention layer, a global average pooling layer, and a fully connected output layer, wherein: The input tensor dimension of the input layer is [number of batches, 12, 2], where 12 represents the number of time steps and 2 represents the dual channels of the median time series images of EVI and RVI. The multi-scale dilated causal convolutional layer includes four convolutional blocks with dilation rates of 1, 2, 4, and 8, respectively. Each convolutional block adopts a causal filling method, and the corresponding receptive fields cover approximately 45 days, 75 days, 135 days, and the entire growth season, respectively. The time attention layer applies time dimension attention weights to the multi-scale feature tensors based on the time step correlation of the multi-scale feature tensors, so as to dynamically focus on key temporal information at different growth stages. After the fully connected output layer is activated by 64-dimensional ReLU, the normalized phenological parameters are output by the 4-dimensional sigmoid-activated output layer.

4. The method for extracting corn planting area according to claim 1, characterized in that, In step S3, the temporal convolutional network TCN adopts a weighted mean square error loss function in the following form: ； Where N represents the number of samples, The predicted SOS value represents the start date of the growing season. This indicates the actual value of SOS representing the start date of the growing season. This represents the predicted value for the peak growth date (POS). This represents the actual value of the peak growth date POS. This represents the predicted value of EOS for the end date of the growing season. The actual value of EOS representing the end date of the growing season. This represents the predicted value of growth amplitude (Amplitude). This represents the actual value of the growth amplitude (Amplitude).

5. The method for extracting corn planting area according to claim 1, characterized in that, In step S5, the multi-strategy phenological perturbation includes at least one of the following: (1) Phenological window offset perturbation strategy: Random sampling is performed from positive maize samples, and a random offset perturbation amount exceeding the normal maize phenological window is applied from the SOS (start date of growing season) and EOS (end date of growing season) feature dimensions of the sampled samples. ,in, The sampling range is [30 days, 60 days], to generate perturbation negative samples that distort phenological rhythms; (2) Feature dimension random permutation strategy: For the positive samples of maize, the growth season start date SOS, growth peak date POS and growth season end date EOS dimension features, random dimension permutation is performed between samples. By breaking the temporal logical constraints between phenological features, perturbation negative samples with anomalous temporal relationships are generated. (3) Mixed coefficient constrained interpolation strategy: Randomly sample two groups of samples from the positive samples of maize. and and with a preset mixing coefficient After linear interpolation, random perturbation amounts within the range of [-40 days, 40 days] are superimposed on the SOS feature dimension from the start date of the growing season of the sample to generate perturbed negative samples with partially fused phenological features and abnormal SOS perturbation.

6. The method for extracting corn planting area according to claim 1, characterized in that, In step S5, the prototype network classifier uses a phenological perception dual-component contrastive loss function, which consists of two parts: a maize positive sample clustering loss function and a non-maize negative sample rejection loss function. The calculation formula for the maize positive sample clustering loss function is as follows: ； in, Indicates the number of positive samples. This represents the cosine similarity between positive sample i and other positive samples k. This represents the weighting coefficient for positive sample i. This represents the feature vector of a positive sample i. The L2-normalized vector representing the mean of the batch positive sample embedding vectors; The formula for calculating the non-corn negative sample rejection loss function is as follows: ； ； in, Indicates the number of negative samples. This represents the cosine similarity between negative sample j and other negative samples l. This represents the weighting coefficient for negative sample j. Indicates boundary values, Indicates the basic boundary value. This represents the adjustment step size coefficient for boundary values. This indicates the current number of training steps. This represents the distance between the aggregated features of negative sample j and positive sample j. This represents the distance between positive samples and their aggregated features.

7. The method for extracting corn planting area according to claim 1, characterized in that, In step S6, the process of performing connected component analysis, inverse hole filling, and quantitative area quantification based on the preliminary identification results to obtain the total corn planting area of ​​the target region includes: S61. Based on the preliminary identification results, perform connected component analysis, morphological processing, and difference analysis to obtain the void region to be verified. S62. For the void area to be verified, the maximum normalized red edge vegetation index (maxNDRE) during the peak growing season is used for spectral secondary verification and reverse filling to obtain the complete area result after filling. S63. Based on the results of the filled complete area, perform area quantification statistics to obtain the total corn planting area of ​​the target area.

8. A deep learning-based system for extracting corn planting area, characterized in that, The system includes a multi-source data index calculation module, a statistical feature image construction module, a phenological feature extraction module, a classification feature matrix generation module, a preliminary identification module for maize planting area classification, and a quantitative statistical module for total maize planting area, wherein: The multi-source data index calculation module is used to acquire radar images and optical images of the target area, and determine multiple vegetation indices based on the optical images and polarization feature indices based on the radar images. The multiple vegetation indices include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Red Edge Vegetation Index (NDRE). The polarization feature index includes Radar Vegetation Index (RVI). The statistical feature image construction module is used to construct a set of statistical feature images corresponding to the entire growing season based on the multiple vegetation indices and the polarization feature index. The set of statistical feature images includes EVI median time series images, RVI median time series images, NDVI mean images, NDRE maximum images, and RVI standard deviation images. The phenological feature extraction module is used to input the median EVI time-series image and the median RVI time-series image into the temporal convolutional network TCN. During the process, the target phenological features are extracted by combining the temporal attention mechanism through adaptive time step weight allocation and multi-scale feature fusion mechanism. The classification feature matrix generation module is used to align, stitch, and standardize the target phenological features with the NDVI mean image, the NDRE maximum value image, and the RVI standard deviation image according to pixels in sequence to obtain the target multidimensional classification feature matrix. The preliminary identification module for corn planting area classification is used to input the target multidimensional classification feature matrix into the prototype network classifier for preliminary identification of corn planting areas. Non-corn negative samples are generated by multi-strategy phenological perturbation of corn positive samples. The total corn planting area quantification and statistics module is used to perform connected component analysis, reverse filling of gaps, and quantification and statistics of area based on the preliminary identification results, so as to obtain the total corn planting area of ​​the target area.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a deep learning-based method for extracting corn planting area as described in any one of claims 1-7.

Images