Weakly supervised crop classification method based on full phenology features and timing similarity constraints

Abstract

The application discloses a kind of weakly supervised crop classification method based on full phenology feature and timing similarity constraint, belong to agricultural remote sensing monitoring technical field, solve the problem that current crop classification method is driven by multiple uncertainties and limits agricultural remote sensing monitoring capability.The application obtains the classification and boundary quality comparable to strong supervision or even better with less manual labeling in precision and robustness;In generalization and migration, through template iteration and parameter self-adaption, support stable reuse across crop structure and across climate year;In engineering usability, with pixel-level reliability and consistency index to support batch quality inspection, difference positioning and incremental recalculation, meet the normal mapping needs of business scale.Overall, the application organically penetrates the four links of "feature-matching-learning-evaluation", establishes an interpretable, controllable and transferable remote sensing crop mapping technology chain, and provides a practical technical support for provincial to national scale agricultural monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural remote sensing monitoring, and in particular to a weakly supervised crop classification method based on full phenological features and time sequence similarity constraints. BACKGROUND

[0002] Agricultural remote sensing monitoring uses satellite and unmanned aerial vehicle technologies to achieve accurate monitoring and management of crop growth, area, disasters, etc. by analyzing remote sensing image data. Multi-source remote sensing data has rapidly improved in terms of spatial and temporal resolution and coverage frequency. However, complex atmospheric conditions, crop phenological differences, and regional management mode heterogeneity make traditional strong supervision mapping face the bottleneck of high labeling cost and weak transferability. At the same time, the business scenario requires the results to be verifiable, updatable, and robust across years, which puts higher standards on the robustness, explainability, and engineering usability of the method.

[0003] However, due to the fact that existing crop classification methods are generally driven by multiple uncertainties, such as the dependence on experience or a single index for determining key phenological periods, it is difficult to simultaneously consider inter-class separability and intra-class redundancy along the entire phenological curve. The time series distortion caused by clouds, shadows, and missing data makes the curve matching fixed threshold highly sensitive. In the weak supervision framework, the quality of pseudo-labels is uncontrollable, and the boundary expansion and noise spread are particularly prominent. The regularized phenological matching and deep network training lack mechanism coupling, and the time sequence consistency is difficult to convert into effective learning signals. At the same time, the results lack pixel-level, quantifiable confidence representation, making it difficult to support large-scale quality inspection and incremental update. These problems limit the ability of automated and normalized agricultural remote sensing monitoring. SUMMARY

[0004] To solve the problem that existing crop classification methods are driven by multiple uncertainties and limit the ability of agricultural remote sensing monitoring, the present application provides a weakly supervised crop classification method based on full phenological features and time sequence similarity constraints.

[0005] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0006] A weakly supervised crop classification method based on full phenological features and time sequence similarity constraints is provided, comprising the following steps:

[0007] S1, obtaining crop ground sample data and historical remote sensing image data in the study area, pre-processing the historical remote sensing image data to obtain a multi-temporal spectral time series corresponding to each pixel, and using the crop ground sample data to construct a standard phenological template curve for each type of crop in the study area;

[0008] S2, according to the multi-temporal spectral time series of each pixel and the standard phenology template of each crop, a weighted dynamic time warping distance reflecting the similarity between each pixel and each crop is calculated, and the difference between the maximum two weighted dynamic time warping distances is taken as the classification confidence score of the corresponding pixel;

[0009] S3, parameterized mapping and order-preserving regression correction are performed on each classification confidence score to obtain the misjudgment probability of the corresponding pixel;

[0010] S4, the misjudgment probability of all pixels is analyzed by using the error discovery rate statistical method, and a pixel set that does not exceed the error discovery rate threshold is selected as the initial anchor sample set of weak supervision learning;

[0011] S5, the initial anchor sample set is input into a spatio-temporal deep learning network to train a weak supervision classification model, and the weak supervision classification model is evaluated and optimized by using an independent verification sample set;

[0012] S6, the current remote sensing image in the research area is obtained and input into the pre-trained weak supervision classification model to obtain the crop classification result of the research area.

[0013] S6, the weak supervision classification model is evaluated and optimized by using an independent verification sample set.

[0014] Further, the calculation expression of the weighted dynamic time warping distance is:

[0015]

[0016] wherein, is the weighted dynamic time warping distance between pixel i and the kth crop; t is the time point of pixel i in the multi-temporal spectral time series; s is the time point of pixel i in the standard phenology template curve, and π is the pairing path between t and s in the time weighted dynamic time warping algorithm; is the time weight vector; is the standard phenology template curve of the kth crop; is the multi-temporal spectral time series of pixel i; is the time mismatch penalty function.

[0017] Further, the expression of the classification confidence score is:

[0018]

[0019]

[0020]

[0021] wherein, a classification confidence score for the pixel i; and are the most similar and the second similar crop class for the pixel i, respectively.

[0022] Further, step S3 further comprises the following steps:

[0023] S31, using each classification confidence score for constructing a feature vector z of the probability calibration input, and mapping z to an initial correct probability estimate function with the expression:

[0024]

[0025]

[0026] wherein, is a Logistic function; is a linear transformation function, and are the intercept and the weight coefficient, respectively;

[0027] S32, performing an isotonic regression correction on to obtain a corrected correct probability estimate function with the expression:

[0028]

[0029] wherein, is a monotone non-decreasing mapping function from the definition domain [0, 1] to the value domain [0, 1] based on the isotonic regression;

[0030] S33, calculating a pixel misjudgment probability function .

[0031] Further, step S4 further comprises the following steps:

[0032] S41, performing an ascending arrangement on to obtain a sequence wherein, , and are the 1st, 2nd and smallest pixel misjudgment probability in , respectively; is the largest pixel misjudgment probability;

[0033] S42, calculating a false discovery rate threshold with the expression:

[0034]

[0035]

[0036] wherein, to meet the maximum index; is the number of pixels; is the preset FDR control level;

[0037] S43, selecting all pixels less than the and assigning the most similar prediction category to the corresponding pixel to form an initial anchor sample set .

[0038] Further, step S4 further comprises calculating:

[0039] The minimum achievable false discovery rate corresponding to each pixel is calculated by the Benjamini-Hochberg method, and the expression is:

[0040]

[0041] wherein, is the minimum value of starting from the position of the jth smallest pixel misclassification probability; is the mth smallest pixel misclassification probability in is the minimum achievable false discovery rate corresponding to pixel i.

[0042] Further, the loss function of the weakly supervised classification model includes a weighted cross-entropy loss term , and the expression is:

[0043]

[0044]

[0045]

[0046]

[0047] wherein, is the multi-factor dynamic weight of pixel i; is the initial anchor sample set; a, b and c are all hyperparameters; is the key phenological period coverage of pixel i; is the average observation quality of pixel i; H is the time period of the key phenological period, is the length of H, ​​is an indicator function for indicating whether the pixel i is valid observation at time t, is a judgment function; is the multi-temporal spectral time series of pixel i after matching by step S2, with dimension , , , and are spatial height, spatial width, temporal phase number and spectral band number per phase, respectively; is a spatio-temporal deep learning network; is a spatio-temporal encoder; is a semantic segmentation decoder; is a classification probability map of whether the pixel i belongs to each crop category.

[0048] Further, the boundary consistency regularization term of the weakly supervised classification model is expressed as:

[0049]

[0050] wherein, is a weight coefficient of ; is the intensity of the boundary predicted by the weakly supervised classification model; is a gradient operator; is the spatial gradient of ; is an L1 norm; is the intensity of the real boundary of ; is an edge detection operator for detecting the edges of an image.

[0051] Further, the phenological shape constraint term of the weakly supervised classification model is expressed as:

[0052]

[0053] wherein, is a weight coefficient of ; is the derivable time sequence feature of pixel i extracted by the weakly supervised classification model from the encoder; is the total number of crop categories; is a Soft-DTW function (Soft-Dynamic Time Warping).

[0054] Further, the evaluation indexes of the weakly supervised classification model include overall accuracy , Kappa coefficient , producer accuracy and user accuracy , which are respectively expressed as:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] the number of samples correctly classified; the total number of all verified samples; the proportion of samples in the kth class that are correctly predicted as the kth class; the number of samples with a true class of k; the proportion of samples predicted as the kth class that are truly the kth class; the number of samples predicted as the kth class; the probability of random consistency.

[0061] The application discloses a weakly supervised crop classification method based on full phenological characteristics and timing similarity constraints, which has the beneficial effects that:

[0062] 1、The application significantly reduces the dependence of the model on labeled data, reduces the cost of manual labeling, and thus obtains classification and boundary quality comparable to strong supervision or even better with less manual labeling in terms of accuracy and robustness through phenological knowledge driving and timing similarity constraints; in terms of generalization and migration, the high confidence of weak supervision anchor points is ensured and the reliability of classification results is improved by combining misjudgment probability calibration and error discovery rate control, supporting stable reuse across crop structures and across climate years; in terms of engineering usability, the generalization ability and boundary recognition accuracy of the model in complex environments are enhanced by introducing boundary consistency regularization terms and phenological shape constraint terms, and pixel-level reliability and consistency indicators support batch quality inspection, difference positioning and incremental recalculation, meeting the needs of normal mapping at the business scale. Overall, the application organically connects the four links of "feature-matching-learning-evaluation", establishes an interpretable, controllable and transferable remote sensing crop mapping technology chain, and provides practical technical support for provincial to national scale agricultural monitoring.

[0063] 2、The application quantifies the similarity between pixels and crop phenological templates through the time-weighted dynamic time warping (TWDTW) algorithm, considers the importance differences of different time points within the phenological period, improves the accuracy of similarity measurement, and thus enhances the discrimination ability of crop classification.

[0064] 3. The application constructs a pixel-level classification confidence score by calculating the distance difference between the most similar class and the less similar class, which can effectively identify the uncertainty in the classification result, and provides a basis for subsequent misjudgment probability calibration and high-quality sample screening.

[0065] 4. The application converts the geometric margin into a misjudgment probability with statistical significance through parameterized mapping and order-preserving regression correction, and converts the weighted dynamic time warping distance difference without probability attribute into a pixel misjudgment probability that can be used for multiple tests. As the final correction link, the order-preserving regression ensures that no matter how the output of the intermediate parameterized mapping is, the misjudgment probability of the subsequent false discovery rate (FDR) method is monotonically decreasing with the increase of the weighted dynamic time warping distance difference.

[0066] 5. The application uses the Benjamini-Hochberg (BH) method to control the false discovery rate (FDR), which can automatically screen high-confidence pixels under the multiple test framework, significantly reduce the proportion of false positive samples, and ensure the quality of weakly supervised training samples.

[0067] 6. The application calculates the minimum reachable false discovery rate for each pixel to provide a quantifiable confidence index. is the original misjudgment probability of the pixel i, and quantifies the minimum false discovery rate (FDR) that the pixel will bear if it is selected as a high-confidence seed point. When training a weakly supervised classification model, the loss function will impose stronger constraints on samples with high weights and low values, forcing the model to preferentially learn the features of the most reliable pixels, effectively suppressing the noise interference of low-quality pseudo-labels.

[0068] 7. The application uses multi-factor dynamic weights to integrate key phenological coverage, observation quality and classification confidence, so that the model can preferentially learn high-quality anchor features during training, improving the efficiency and classification accuracy of weakly supervised learning.

[0069] 8. The application uses a boundary consistency regularization term to constrain the consistency of the model prediction boundary with the real remote sensing image edge, effectively suppressing the boundary blur problem caused by noise and improving the recognition accuracy of crop plot boundaries.

[0070] 9. The application introduces the time sequence shape information of the crop phenological template into the model training in a differentiable form through the phenological shape constraint term, enhances the learning ability of the model for crop phenological rules, and improves the generalization performance across regions and years.

[0071] 10. This invention provides a comprehensive and quantifiable model performance evaluation system by integrating multiple evaluation indicators, supporting the comparability and verifiability of classification results, and meeting the requirements for accuracy transparency and reliability in business applications. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating a weakly supervised crop classification method based on full phenological characteristics and temporal similarity constraints. Detailed Implementation

[0073] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0074] refer to Figure 1 This is a flowchart illustrating a weakly supervised crop classification method based on full phenological characteristics and temporal similarity constraints. Its purpose is to address the limitation of agricultural remote sensing monitoring capabilities caused by multiple uncertainties in existing crop classification methods. The method includes the following steps:

[0075] S1. Obtain crop ground sample data and historical remote sensing image data in the study area. Preprocess the historical remote sensing image data to obtain the multi-temporal spectral time series corresponding to each pixel, and use the crop ground sample data to construct the standard phenological template curve for each type of crop in the study area.

[0076] S2. Compare the multi-temporal spectral time series of each pixel with the standard phenological template curve of each crop type based on weighted temporal matching, and calculate the weighted dynamic time warping distance to reflect the similarity between the two. Based on the multi-temporal spectral time series of each pixel and the standard phenological template of each crop type, calculate the weighted dynamic time warping distance to reflect the similarity between each pixel and each crop type, and take the difference between the two largest weighted dynamic time warping distances as the classification confidence score of the corresponding pixel.

[0077] Specifically, the weighted dynamic time-warped distance between pixel i and the k-th crop category. The calculation expression is:

[0078]

[0079] in, represents the dissimilarity between the time series of pixel i and the standard template curve of the kth crop, the smaller the value, the more likely the pixel is planted with the kth crop. t is the time point of pixel i in the multi-temporal spectral time series; s is the time point of pixel i in the standard template curve, and π is the pairing path between t and s in the time-weighted dynamic time warping algorithm, which solves the problem of different synchronization of time series and standard template curve on the time axis; is the time weight vector; is the standard template curve of the kth crop, specifically the average value of the standard template curve of the kth crop (such as wheat, corn, etc.) at time point s; is the multi-temporal spectral time series of pixel i; is the time mismatch penalty function, which represents the time difference between the two points aligned on the two curves. This function will punish those alignments with too large time difference. For example, aligning the wheat template in May with the target pixel in January will be severely punished due to the large time difference, even if the spectral values happen to be similar, thus avoiding false alignment.

[0080] Specifically, the classification confidence score of pixel i is The expression of

[0081]

[0082]

[0083]

[0084] wherein, and are the most similar crop category and the second similar crop category corresponding to pixel i, respectively.

[0085] S3, parameterize mapping and order-preserving regression correction are performed on each classification confidence score to obtain the misjudgment probability of the corresponding pixel.

[0086] S4, the error discovery rate statistical method is used to analyze the misjudgment probability of all pixels, and the pixel set that does not exceed the error discovery rate threshold is selected as the initial anchor point sample set of weak supervision learning.

[0087] S5, input the initial anchor sample set into a spatio-temporal deep learning network composed of a U-TempoNet architecture combined with CNN and LSTM to train a weakly supervised classification model, the loss function of the weakly supervised classification model includes a weighted cross-entropy loss term, a boundary consistency regularization term and a phenological shape constraint term, wherein the phenological shape constraint term is to calculate the differentiable shape similarity distance between the time sequence features extracted by the weakly supervised classification model based on the standard phenological template curve and the weighted time sequence by using a Soft-DTW algorithm, and the weakly supervised classification model is evaluated and optimized using an independent validation sample set.

[0088] S6, obtaining the current remote sensing image in the study area and inputting it into the pre-trained weakly supervised classification model to obtain the crop classification result corresponding to the study area.

[0089] As a further scheme of the embodiment, in step S3, the inventors consider that directly using the confidence is unreliable because it does not have the probability attribute and cannot be used for subsequent statistical testing, so the geometric margin is converted into a false judgment probability that can be used for multiple testing. Thus, based on the verification sample, a calibration independent variable is constructed from the two-class margin obtained by time-weighted dynamic time warping, and a joint feature is optionally formed by the logarithmic ratio of the minimum and second minimum two-class distances, which is scaled and input into the parameterized logistic regression mapping (Platt) to obtain a correct probability estimate between zero and one; the mapping satisfies the constraint that the output probability monotonically increases when the input margin monotonically increases, and suppresses the variance fluctuation under the small sample condition through regularization and partition folding cross-validation. When the number of verification samples does not reach the preset threshold, the system uses the parameter mapping as the only calibration function; when the number of verification samples reaches or exceeds the threshold (the default number of samples is 300), the ordered regression (PAV) is performed on the output score and the corresponding true value pair to obtain a monotonic correction function defined on [0, 1], and the two are combined to form the final corrected correct probability estimation function . which guarantees the monotonic consistency from the geometric margin to the probability scale, wherein the parameterized part provides stable tail extrapolation and small sample availability, and the ordered part is used to eliminate the systematic deviation in the middle section.

[0090] Specifically, step S3 includes the following steps:

[0091] S31, use each classification confidence score to construct a feature vector z for probability calibration input, and map z to an initial correct probability estimation function by logistic regression.

[0092]

[0093]

[0094] wherein, is a Logistic function; is a linear transformation function, and are intercept and weight coefficient, respectively.

[0095] S32, performing an order-preserving regression correction on to obtain a corrected correct probability estimation function , whose expression is.

[0096]

[0097] wherein, is a monotone non-decreasing mapping function from the definition domain [0, 1] to the value domain [0, 1] based on the order-preserving regression.

[0098] S33, calculating a pixel misjudgment probability function .

[0099] As a further scheme of the embodiment, in step S4, the inventor considers that when all pixels in the whole image are classified and predicted, if a traditional single threshold is used, a large number of false positive errors (i.e. low confidence pixels are incorrectly judged as high confidence) will be accumulated. In order to strictly control the overall quality of the weakly supervised anchor points generated finally, the false discovery rate (FDR) control under the multiple test framework is introduced in this scheme. The FDR control can control the proportion of false discoveries to be at a preset level (such as 5%) while ensuring a high recall rate, thereby realizing statistical guarantee of the overall error rate of the anchor point set.

[0100] Specifically, step S4 includes the following steps:

[0101] S41, performing ascending arrangement on to obtain a sequence, wherein, , and are the first smallest, the second smallest and the smallest pixel misjudgment probability in , respectively; is the largest pixel misjudgment probability;

[0102] S42, calculating a false discovery rate threshold , whose expression is:

[0103]

[0104]

[0105] wherein, is the maximum index satisfying ; is the number of pixels; is the preset FDR control level;

[0106] S43, selecting all pixels less than and assigning the most similar predicted class to the corresponding pixel to form an initial anchor sample set .

[0107] S44, calculating the minimum achievable false discovery rate corresponding to each pixel by the Benjamini-Hochberg method, the expression is:

[0108]

[0109] wherein, is the minimum value of starting from the position of the jth smallest pixel misclassification probability; is the mth smallest pixel misclassification probability in is the minimum achievable false discovery rate corresponding to pixel i. By calculating the minimum achievable false discovery rate , a quantifiable confidence index is provided for each pixel. is the original misclassification probability of pixel i, while quantifies the minimum false discovery rate (FDR) that will be borne if this pixel is selected as a high-confidence seed point. When training a weakly supervised classification model, the loss function will impose stronger constraints on samples with high weights and low values, forcing the model to preferentially learn the most reliable pixel features, effectively suppressing the noise interference of low-quality pseudo-labels.

[0110] As a further scheme of the embodiment, the loss function wherein, , and are the weighted cross-entropy loss term, the boundary consistency regularization term and the phenological shape constraint term, respectively.

[0111] Specifically, the weighted cross-entropy loss term is a comprehensive confidence index used to assign different weights to each anchor sample in weakly supervised training, the expression of which is:

[0112]

[0113]

[0114] ​​​

[0115]

[0116] where, is the multi-factor dynamic weight of pixel i, forcing the model to preferentially learn and rely on high-quality anchor samples with "complete and good observations and high classification confidence"; is the initial anchor sample set; a, b and c are all hyperparameters; is the key phenological coverage of pixel i; is the average observation quality of pixel i; H is the time period of the key phenological period, is the length of H, is an indicator function used to represent whether pixel i is effectively observed at time t, is a judgment function; is the multi-temporal spectral time series of pixel i after matching in step S2, with a dimension of , , , and are spatial height, spatial width, number of temporal phases and spectral band per phase, respectively; is a spatio-temporal deep learning network; is a spatio-temporal encoder; is a semantic segmentation decoder; is a classification probability map of pixel i belonging to each crop category.

[0117] Specifically, the boundary consistency regularization term encourages the model to predict the classification probability map to be consistent with the inherent object boundary of the remote sensing image itself, so as to improve the spatial smoothness and boundary accuracy of the model output and suppress noise, and the expression of is:

[0118]

[0119] where, is the weight coefficient of ; is the intensity of the boundary predicted by the weakly supervised classification model; is the gradient operator; is the spatial gradient of ; is the L1 norm; is the intensity of the true boundary of ; is an edge detection operator used to detect the edges of an image.

[0120] Specifically, the phenological shape constraint term encourages the model to learn deep temporal features a phenology template curve of a crop standard Similar in shape, the enhanced model improves the interpretability and generalization ability across years, and its expression is:

[0121]

[0122] wherein, is a weight coefficient of the weakly supervised classification model; is a derivable time sequence feature of the pixel i extracted from the encoder by the weakly supervised classification model; is the total number of crop categories; is a Soft-DTW function (Soft-Dynamic Time Warping).

[0123] As a further scheme of the embodiment, the weakly supervised classification model evaluation indicators include overall accuracy , Kappa coefficient , producer accuracy and user accuracy , and their expressions are respectively:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] is the number of correctly classified samples; is the total number of all verified samples; is the proportion of the k-th crop sample correctly predicted as the k-th class; is the number of samples with the true class being k; is the proportion of the k-th class in the samples predicted as the k-th class; is the number of samples predicted as the k-th class; is the random consistency probability.

[0130] In the embodiment, by constructing a matrix with the true value class as rows and the predicted class as columns, and comprehensively considering various evaluation indicators, a comprehensive and quantifiable model performance evaluation system is provided, which supports the comparability and verifiability of the classification results and meets the requirements of accuracy transparency and reliability in business applications.

[0131] In summary, the beneficial effects of the scheme are:

[0132] ​The scheme significantly reduces the dependence of the model on labeled data and reduces the cost of manual labeling by driving with phenology knowledge and time sequence similarity constraint, thereby obtaining classification and boundary quality comparable to strong supervision or even better with less manual labeling; in terms of generalization and migration, combined with misjudgment probability calibration and error discovery rate control, it ensures high confidence of weakly supervised anchor points and improves the reliability of classification results, supporting stable reuse across crop structures and across climate years; in terms of engineering usability, by introducing boundary consistency regularization terms and phenology shape constraint terms, the generalization ability and boundary recognition accuracy of the model in complex environments are enhanced, and pixel-level reliability and consistency indicators support batch quality inspection, difference positioning and incremental recalculation, meeting the needs of normal mapping at the business scale. Overall, the present application organically connects the four links of "feature-matching-learning-evaluation", establishes a remote sensing crop mapping technology chain that is interpretable, controllable and transferable, and provides practical technical support for agricultural monitoring at the provincial to national scale.

[0133] Although the specific embodiments of the application are described in detail with reference to the accompanying drawings, it should not be understood as limiting the scope of protection of the present application. Various modifications and variations made by those skilled in the art within the scope described in the claims are still within the scope of the present application.

Claims

1. A weakly supervised crop classification method based on full phenological features and temporal similarity constraints, characterized in that, The method comprises the following steps: S1, obtaining crop ground sample data and historical remote sensing image data in a study area, pre-processing the historical remote sensing image data to obtain a multi-temporal spectral time series corresponding to each pixel, and using the crop ground sample data to construct a standard phenology template curve of each type of crop in the study area; S2, calculating a weighted dynamic time warping distance reflecting the similarity between each pixel and each type of crop according to the multi-temporal spectral time series of each pixel and the standard phenology template of each type of crop, and taking the difference between the maximum two weighted dynamic time warping distances as a classification confidence score of the corresponding pixel; S3, performing parameterized mapping and order-preserving regression correction on each classification confidence score to obtain a misjudgment probability of the corresponding pixel; S31, construct a feature vector z with each classification confidence score as a probability calibration input, and map z logistic regression to an initial correct probability estimate function whose expression is: wherein, is a logistic function; is a linear transformation function, and are intercept and weight coefficients, respectively. S32、to Performing an order-preserving regression correction to obtain a corrected probability estimation function : wherein, is a monotone non-decreasing mapping function from the domain [0,1] to the range [0,1] based on an order preserving regression; S33, calculating the pixel misjudgment probability function ; S4, analyzing the misjudgment probability of all pixels using a false discovery rate statistical method, and screening out a pixel set that does not exceed a false discovery rate threshold as an initial anchor sample set for weakly supervised learning; S41、to ascending order, obtaining sequence, wherein, , and are respectively the first, the second and the smallest pixel error probability in the maximum pixel error probability; S42, calculate a false discovery rate threshold : wherein, to meet the maximum index; the number of pixels; a preset FDR control level; S43, select all less than of and give the corresponding pixel its most similar prediction class , forming the initial anchor sample set ; S44, calculating a minimum reachable false discovery rate corresponding to each pixel by a Benjamini-Hochberg method, and the expression is: in, Starting from the position of the misclassification probability of the j-th smallest pixel, find all middle The minimum value; for The probability of misclassification of the m-th smallest pixel; Let i be the minimum reachable error detection rate for pixel i. S5, inputting the initial anchor sample set into a spatio-temporal deep learning network to train a weakly supervised classification model, and evaluating and optimizing the weakly supervised classification model using an independent validation sample set; S6, obtaining a current remote sensing image in the study area, and inputting the current remote sensing image into the pre-trained weakly supervised classification model to obtain a crop classification result corresponding to the study area.

2. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 1, characterized in that, The expression for calculating the weighted dynamic time warping distance is: wherein, is the weighted dynamic time warping distance between the pixel i and the kth crop type; t is the time point of the pixel i in the multi-temporal spectral time series; s is the time point of the pixel i in the standard phenology template curve, π is the pairing path between t and s in the time-weighted dynamic time warping algorithm; is the time weight vector; is the standard phenology template curve of the kth crop type; is the multi-temporal spectral time series of the pixel i; is the time mismatch penalty function.

3. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 2, characterized in that, The expression for calculating the classification confidence score is: wherein, is a classification confidence score for the pixel i; and are the most similar crop class and the second most similar crop class, respectively, corresponding to the pixel i.

4. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 2, characterized in that, The loss function of the weakly supervised classification model comprises a weighted cross-entropy loss term The expression is: wherein, is the multi-factor dynamic weight of the pixel i; is the initial anchor sample set; a, b and c are all hyperparameters; is the key phenology coverage of the pixel i; is the observation quality mean of the pixel i; H is the time period of the key phenology, is the length of H, is the indicator function used to represent whether the pixel i is validly observed at time t, is the judgment function; is the multi-temporal spectral time series of the pixel i after matching in step S2, with the dimension of , , , and are spatial height, spatial width, temporal phase number and spectral band per phase, respectively; is the spatio-temporal deep learning network; is the spatio-temporal encoder; is the semantic segmentation decoder; is the classification probability map of the pixel i belonging to each crop category.

5. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 4, characterized in that, The loss function of the weakly supervised classification model comprises a boundary consistency regularization term The expression is: wherein, is a weight coefficient; is a weight coefficient; is a strength of a boundary predicted by the weakly supervised classification model; is a gradient operator; is a spatial gradient of is a spatial gradient of is an LI norm; is a strength of a true boundary of is a strength of a true boundary of is an edge detection operator for detecting edges of an image.

6. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 5, characterized in that, The loss function of the weakly supervised classification model comprises a phenological shape constraint term The expression is: wherein, is a weight coefficient; is a differentiable temporal feature of the pixel i extracted from the encoder by the weakly supervised classification model; is the total number of crop classes; is a Soft-DTW function.

7. The weakly supervised crop classification method based on full phenological features and timing similarity constraints according to claim 1, characterized in that, Evaluation metrics for weakly supervised classification models include overall accuracy , Kappa coefficient , producer accuracy , and user accuracy , whose expressions are respectively: Number of samples correctly classified; Total number of samples verified; Proportion of samples of class k that were correctly predicted as class k; Number of samples of true class k; Proportion of samples predicted as class k that are truly class k; Number of samples predicted as class k; Random uniform probability.

Images