Small-sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge

Abstract

This invention discloses a small-sample crop mapping method using semi-supervised learning guided by phenological prior knowledge. The method involves acquiring time-series VH data, VV data, and vegetation index data for a study area; obtaining the crop category and geographic coordinates of each sampling point within the study area and constructing labeled samples; and constructing unlabeled samples based on the overall phenological prior knowledge; constructing a semi-supervised learning network; calculating the total loss function and training the backbone network model; and obtaining time-series VH and VV data for cultivated land pixels in the predicted area, inputting them into the trained backbone network model to obtain the final crop category. This invention employs a pseudo-label-based semi-supervised learning strategy, which overcomes the high reliance of supervised learning-based crop recognition models on field samples and improves the performance of simultaneous mapping of multiple crops using a small number of samples collected within a small area.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural remote sensing technology, and specifically relates to a small-sample crop mapping method using semi-supervised learning guided by phenological priori knowledge, which is suitable for small-sample crop mapping. Background Art

[0002] Regional and even global-scale spatiotemporal distribution data on crops is crucial scientific data for adjusting crop planting structures, estimating crop yields, and estimating methane emissions. It is also crucial for scientifically allocating agricultural resources and achieving sustainable agricultural development. Compared to field surveys, which require significant human, material, and time resources, satellite remote sensing, with its large-scale, synchronous, and real-time Earth observation capabilities, has become an effective means of monitoring the spatiotemporal distribution of crops.

[0003] With the development of satellites with varying temporal and spatial resolutions and multiple sensor arrays, a wide range of crop mapping methods based on remote sensing data are now available, primarily including phenology-based and decision tree-based methods, machine learning methods, and deep learning methods. Because crop phenology is highly correlated with its physical and chemical properties and agricultural practices, and does not significantly change over time or space, phenology-based and decision tree-based methods use remote sensing data to establish rule sets that highlight the unique phenological characteristics of crops and employ specific thresholds to identify these crops. While these methods improve the efficiency of large-scale crop mapping, they are highly dependent on image quality during key phenological periods and specific thresholds, making it difficult to simultaneously extract multiple crops. Machine learning-based crop mapping methods employ machine learning models such as random forests (RFs), support vector machines (SVMs), and artificial neural networks (ANNs) to identify different crops by establishing mathematical relationships between remote sensing image features and crop samples. However, these machine learning models require manual feature engineering to select appropriate classification features, and the temporal dependencies between time-series remote sensing images are not considered during model training. Deep learning models have been widely used in crop remote sensing mapping in recent years, achieving promising results due to their ability to automatically extract and learn deep features from remote sensing imagery from a large number of training samples. However, supervised deep learning models require sufficient, high-quality, and representative samples to learn the complex spectral characteristics of different crops. In contrast, semi-supervised learning, by combining a small number of field samples with a large amount of unlabeled remote sensing data, can reduce reliance on field samples while maintaining classification accuracy. Therefore, designing a suitable semi-supervised learning framework that fully utilizes the rich information contained in a small number of field samples collected in a small area and the massive amount of unlabeled remote sensing data acquired by satellites, thereby alleviating the high reliance of existing crop mapping methods on large samples, is key to achieving efficient crop mapping over large areas. Summary of the Invention

[0004] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and provide a small-sample crop mapping method guided by semi-supervised learning based on phenological prior knowledge, so as to alleviate the high dependence of crop remote sensing mapping on field samples in large areas and areas with complex planting structures.

[0005] In order to achieve the above object, the following is the technical solution of the present invention:

[0006] The small-sample crop mapping method using semi-supervised learning guided by phenological prior knowledge includes the following steps:

[0007] Step S1: Acquire Sentinel-1 data for the study area and generate time-series VH data and time-series VV data. Acquire Sentinel-2 surface reflectance data for the study area, perform cloud removal on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data. Obtain the crop category and geographic coordinates of each sampling point in the study area, use the time-series VH data and time-series VV data of the pixels corresponding to the sampling points as labeled samples, and use the crop category of the pixels corresponding to the sampling points as labels.

[0008] Step S2: extract the phenological prior knowledge of each phenological prior crop type in the study area, calculate the union of the phenological prior knowledge of each phenological prior crop type as the overall phenological prior knowledge, and construct unlabeled samples based on the overall phenological prior knowledge;

[0009] Step S3: constructing a semi-supervised learning network, which includes a backbone network model based on a pseudo-label semi-supervised training strategy and an adaptive threshold adjustment strategy;

[0010] Step S4: Calculate the total loss function, and train the backbone network model based on the pseudo-label semi-supervised training strategy and adaptive threshold adjustment strategy in step S3 under the condition of minimizing the total loss function;

[0011] Step S5: Obtain the time series VH data and time series VV data of the cultivated land pixels in the area to be predicted and input them into the trained backbone network model to obtain the final crop category.

[0012] The phenological prior knowledge of each phenological prior crop type in the study area extracted in step S2 is generated based on the following steps:

[0013] According to the phenological priori crop types of the study area, the phenological priori crop subtypes of the study area are confirmed, and the phenological priori classification rules corresponding to the phenological priori crop subtypes are confirmed. The phenological priori classification rules are combined with the time series VH data, time series VV data and Sentinel-2 vegetation index data obtained in step 1 to calculate the phenological prior knowledge corresponding to the phenological priori crop subtypes;

[0014] Based on the phenological prior knowledge corresponding to the phenological prior crop subtypes, the phenological prior knowledge of each phenological prior crop type in the study area was obtained;

[0015] The step S2 in which unlabeled samples are constructed based on the overall phenological prior knowledge is based on the following steps:

[0016] Multiple pixels are randomly selected based on the overall phenological prior knowledge, and the time series VH data and time series VV data corresponding to each pixel are used as unlabeled samples.

[0017] As mentioned above, the phenological prior crop types in the study area include: single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops, winter rapeseed-other crops, rice-shrimp fields, and other crops;

[0018] The phenological prior crop subtypes in the study area are: rice, rice-shrimp fields, winter rapeseed, winter wheat, non-rice crops, and non-winter crops.

[0019] The phenological priori classification rules corresponding to the phenological priori crop subtypes mentioned above are as follows:

[0020] The rice phenology prior classification rule is: the pixel is a cultivated land pixel, the time proportion of the first flooding signal of the pixel is greater than 10%, and the phenology prior type of the cultivated land pixel is not a rice-shrimp field; the judgment basis of the first flooding signal of the pixel is: Normalized Difference Vegetation Index (NDVI) < Land Surface Water Index (LSWI) or Enhanced Vegetation Index (EVI) < Land Surface Water Index (LSWI):

[0021] The phenological prior classification rule of the rice-shrimp field is as follows: the pixel is a cultivated land pixel, the time proportion of the first flooding signal of the pixel in the PW1 period is greater than 60%, the time proportion of the vegetation signal of the pixel in the PW2 period is greater than 10%, and the time proportion of the first flooding signal of the pixel in the PW3 period is greater than 10%; the judgment basis of the first flooding signal of the pixel is: the normalized vegetation index NDVI < the surface water index LSWI or the enhanced vegetation index EVI < the surface water index LSWI, and the judgment basis of the vegetation signal of the pixel is: the normalized vegetation index NDVI > the surface water index LSWI or the enhanced vegetation index EVI > the surface water index LSWI;

[0022] The phenological prior classification rules for non-rice crops are as follows: the pixel is a cultivated land pixel, the pixel has at least one cloud-free Sentinel-2 surface reflectance data from May to mid-August, and the proportion of time the second flooding signal appears in the pixel is 0; the second flooding signal of the pixel is determined based on: the Normalized Difference Vegetation Index (NDVI) < the Land Water Index (LSWI) + 0.05 or the Enhanced Vegetation Index (EVI) < the Land Water Index (LSWI) + 0.05;

[0023] The phenological prior classification rules for winter wheat are as follows: the winter crop index of the pixel is greater than 0.3 and the median of the April time series VH data is less than -17;

[0024] The phenological prior classification rules for winter rapeseed are: the winter crop index of the pixel is greater than 0.3 and the median of the April time series VH data is greater than -15;

[0025] The phenological priori classification rule for non-winter crops is: the winter crop index of the pixel is greater than 0 and less than 0.1;

[0026] The PW1 period, the PW2 period and the PW3 period are designated periods.

[0027] As mentioned above, the phenological prior knowledge of each crop type in the study area was obtained based on the following steps:

[0028] Calculate the rice phenology prior knowledge and the non-winter crop phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the non-winter crop phenology prior knowledge to obtain the single-season rice phenology prior knowledge;

[0029] Calculate prior knowledge of rice-shrimp field phenology;

[0030] Calculate the rice phenology prior knowledge and the winter wheat phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the winter wheat phenology prior knowledge to obtain the winter wheat-rice phenology prior knowledge;

[0031] Calculate the rice phenology prior knowledge and the winter rapeseed phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the winter rapeseed phenology prior knowledge to obtain the winter rapeseed-rice phenology prior knowledge;

[0032] Calculate the non-rice crop phenology prior knowledge and winter wheat phenology prior knowledge, and calculate the intersection of the non-rice crop phenology prior knowledge and the winter wheat phenology prior knowledge to obtain the winter wheat-other crop phenology prior knowledge;

[0033] Calculate the non-rice crop phenology prior knowledge and the winter rapeseed phenology prior knowledge, and calculate the intersection of the non-rice crop phenology prior knowledge and the winter rapeseed phenology prior knowledge to obtain the winter rapeseed-other crop phenology prior knowledge;

[0034] The non-rice crop phenology prior knowledge and the non-winter crop phenology prior knowledge are calculated, and the intersection of the non-rice crop phenology prior knowledge and the non-winter crop phenology prior knowledge is calculated to obtain the other crop phenology prior knowledge.

[0035] The backbone network model in step S3 is a bidirectional LSTM network based on the attention mechanism. The backbone network model is used to extract the labeled deep semantic features and unlabeled deep semantic features corresponding to the labeled samples and unlabeled samples respectively, and further calculate the probability values ​​corresponding to various crop categories of the labeled samples and unlabeled samples based on the labeled deep semantic features and unlabeled deep semantic features;

[0036] The pseudo-label semi-supervised training strategy in step S3 is specifically as follows: selecting the crop category corresponding to the maximum probability value of the unlabeled sample, selecting the unlabeled sample whose crop category corresponding to the maximum probability value is consistent with the phenological prior crop type to which it belongs, further selecting the unlabeled sample whose maximum probability value is greater than a set adaptive threshold, and using the crop category corresponding to the maximum probability value of the further selected unlabeled sample as the pseudo label of the corresponding unlabeled sample;

[0037] The adaptive threshold adjustment strategy in step S3 is specifically as follows: the adaptive threshold corresponding to the current training cycle is dynamically adjusted based on the changing trend of the number of pseudo labels generated in the previous two training cycles and the 10th quantile of the pseudo label prediction probability distribution in the previous training cycle.

[0038] As mentioned above, the corresponding relationship between the adaptive threshold corresponding to the current training cycle, the changing trend of the number of pseudo labels generated in the previous two training cycles, and the 10th quantile of the pseudo label prediction probability distribution in the previous training cycle is as follows:

[0039] When the number of pseudo labels found increases, or the number of pseudo labels found decreases by less than or equal to 20%, a continuous Q10 distribution interval is set. When Q10 falls in different Q10 distribution intervals, the adaptive threshold selects the corresponding value. The larger the Q10 distribution interval, the larger the adaptive threshold value.

[0040] When the number of pseudo labels found decreases by more than 20%, the adaptive threshold corresponding to the Q10 distribution interval is smaller than the adaptive threshold corresponding to the same Q10 distribution interval when the number of pseudo labels increases or decreases by less than or equal to 20%. When Q10 falls in different Q10 distribution intervals, the adaptive threshold selects the corresponding value. The larger the Q10 distribution interval, the larger the adaptive threshold value.

[0041] Among them, Q10 is the 10th percentile value of the pseudo-label prediction probability in the previous training cycle.

[0042] The calculation of the total loss function in step S4 as described above is based on the following formula:

[0043] ;

[0044] ;

[0045] ;

[0046] in, is the total loss function, is the loss function calculated based on labeled samples and predicted crop categories, is the loss function calculated based on unlabeled samples and pseudo labels; Indicates the total number of labeled samples in each batch of training in each round of training cycle, Represents the first labeled samples The probability value distribution results are: Indicates the labels for labeled examples; Indicates the calculation of cross entropy loss value; represents the total number of unlabeled samples for each batch of training in each round, is a multiple; Represents the first unlabeled samples The probability value distribution results of ; Represents the adaptive threshold used to judge pseudo labels; Indicates the The maximum probability value of unlabeled samples, For the The pseudo label corresponding to the maximum probability value of unlabeled samples, Represents unlabeled samples The phenological priori crop type, == is the match operator, 1() is the indicator function, which returns a value of 1 when the set conditions in the brackets are met, and returns a value of 0 when the set conditions in the brackets are not met.

[0047] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned crop mapping method when executing the computer program.

[0048] A computer-readable storage medium stores a computer program, which implements the steps of the above-mentioned crop mapping method when executed by a processor.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This paper adopts a semi-supervised learning strategy based on pseudo-labels, breaking through the high dependence of supervised learning-based crop recognition models on field samples. At the same time, it introduces phenological prior knowledge and designs an adaptive threshold strategy to improve the quantity, quality and representativeness of pseudo-labels generated in the semi-supervised learning strategy based on pseudo-labels, thereby increasing the sample richness used for model training and thus improving the accuracy of crop mapping in small sample scenarios.

[0051] This invention effectively improves the performance of simultaneous mapping of multiple crops using a small number of samples collected in a small area, providing a new perspective for large-scale crop remote sensing monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the present invention;

[0053] Figure 2 This is the distribution map of the study area of ​​the present invention;

[0054] Figure 3 This is the F1-score indicator evaluation result diagram of the crop classification accuracy of the present invention in the study area;

[0055] Figure 4 This is a diagram showing the User's Accuracy evaluation results of the crop classification accuracy of the present invention in the study area;

[0056] Figure 5 This is a graph showing the Producer's Accuracy evaluation results of the crop classification accuracy of the present invention in the study area. DETAILED DESCRIPTION

[0057] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to examples. The implementation examples described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0058] Example 1:

[0059] like Figure 1 As shown in Figure 2, the small-sample crop mapping method using phenological prior knowledge guided semi-supervised learning (PhenoSSL) includes the following steps:

[0060] Step S1: Acquire Sentinel-1 data in the study area and generate time-series VH data and time-series VV data. Acquire Sentinel-2 surface reflectance data in the study area, perform cloud removal on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data. At the same time, obtain the crop categories and geographic coordinates of each sampling point in the study area. The crop categories include single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops, winter rapeseed-other crops, rice-shrimp fields, and other crops. The time-series VH data and time-series VV data of the pixels corresponding to the sampling points are used as labeled samples, and the crop categories of the pixels corresponding to the sampling points are used as labels (see the distribution of the study area and field sampling points for details). Figure 2 ), the detailed steps are:

[0061] Step S1.1: Collect Sentinel-1 data for the study area from January 1, 2021, to October 30, 2021, based on the GEE platform. For Sentinel-1 data, select the VH backscatter coefficient data and VV backscatter coefficient data in the IW mode of the Sentinel-1 data. Perform preprocessing on the VH backscatter coefficient data and VV backscatter coefficient data in the IW mode of the Sentinel-1 data by removing high-incident angle data in the image overlap area, performing 20-day median synthesis, Refined Lee spatial filtering, and adaptive polynomial temporal filtering to obtain time-series VH data and time-series VV data as research data.

[0062] Step S1.2: Collect Sentinel-2 surface reflectance data for the study area from September 30, 2020, to December 31, 2021, using the GEE platform. For the Sentinel-2 surface reflectance data, use the Sentinel-2 cloud probability product provided by the GEE platform to remove pixels with a corresponding cloud probability greater than 65% to obtain the de-clouded Sentinel-2 surface reflectance data. Calculate the Sentinel-2 vegetation index data from the de-clouded Sentinel-2 surface reflectance data. The Sentinel-2 vegetation index data includes the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), the Land Surface Water Index (LSWI), and the Winter Crop Index (WCI). The formula is as follows:

[0063]

[0064]

[0065]

[0066]

[0067] in, 、 、 as well as They represent the blue, red, near-infrared, and short-wave infrared spectral reflectances of the Sentinel-2 surface reflectance data. 、 and They refer to the maximum NDVI value of winter crops in the vigorous growth period, the minimum NDVI value in the early growth stage, and the minimum NDVI value in the late growth stage; the time ranges of Heading in the vigorous growth period, EGS in the early growth stage, and LGS in the late growth stage in the study area are 2021-02-01 to 2021-05-01, 2020-09-30 to 2020-12-01, and 2021-05-01 to 2021-07-01, respectively.

[0068] Step S1.3: Field personnel went to the study area in late April and late August 2021 to conduct field sampling. They used a handheld GPS to record the crop category and geographic coordinates of each sampling point in the study area. Based on the geographic coordinates, the pixels in the time series VH data and time series VV data were associated with the sampling points. The time series VH data and time series VV data corresponding to the pixels of the sampling points were used as labeled samples, and the crop category corresponding to the pixels of the sampling points was used as the label. In this case, 64 labeled samples were selected as research data for application example.

[0069] Step S2: Extract the phenological prior knowledge of each phenological prior crop type in the study area. The phenological prior knowledge is a layer composed of pixels belonging to the corresponding phenological prior crop type in the study area. The union of the phenological prior knowledge of each phenological prior crop type is calculated as the overall phenological prior knowledge. Based on the overall phenological prior knowledge, unlabeled samples are constructed. The true crop category of each pixel in the unlabeled sample is unknown, and only the phenological prior crop type is known. The specific steps include:

[0070] Step S2.1: Confirm the phenological priori crop subtype of the study area based on the phenological priori crop type of the study area, confirm the phenological priori classification rules corresponding to the phenological priori crop subtype, and combine the phenological priori classification rules with the time series VH data, time series VV data and Sentinel-2 vegetation index data obtained in step 1 to obtain the phenological prior knowledge corresponding to the phenological prior crop subtype.

[0071] In some embodiments, the phenological prior crop types for the study area include: single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops, winter rapeseed-other crops, rice-shrimp paddies, and other crops. The phenological prior crop subtypes for the study area are: rice, rice-shrimp paddies, winter rapeseed, winter wheat, non-rice crops, and non-winter crops.

[0072] Based on the rice phenology prior classification rules, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Land Surface Water Index (LSWI), the rice phenology prior knowledge was calculated.

[0073] Based on the rice-shrimp field phenology prior classification rules, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Land Surface Water Index (LSWI), the rice-shrimp field phenology prior knowledge was calculated.

[0074] Based on the prior classification rules of winter rapeseed phenology, the winter crop index (WCI), and time series VH data, the prior knowledge of winter rapeseed phenology was calculated.

[0075] Based on the prior classification rules of winter wheat phenology, the winter crop index (WCI), and time-series VH data, the prior knowledge of winter wheat phenology was calculated.

[0076] Based on the prior classification rules of non-rice crop phenology, the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), and the Land Surface Water Index (LSWI), the prior knowledge of non-rice crop phenology was calculated.

[0077] Based on the prior classification rules of non-winter crops' phenology and the winter crop index (WCI), the prior knowledge of non-winter crops' phenology is calculated.

[0078] The intersection of cultivated land pixels in the Globeland30 data and the CNLUCC land cover data of the study area is calculated to obtain the cultivated land layer as the prior knowledge of cultivated land.

[0079] The intersection of non-cultivated land pixels in the Globeland30 data and the CNLUCC land cover data is calculated to obtain the non-cultivated land layer as the prior knowledge of non-cultivated land.

[0080] Among them, the prior classification rules for rice phenology are: Cropland = 1 and Frequency flooding signals >10% and Pixel RiceCrayfish =False; Cropland is 1, indicating that this pixel is cultivated land;

[0081] Among them, the phenological prior classification rules of rice-shrimp fields are: Cropland=1 and Frequency flooding signals in PW1 >60% and Frequency vegetation signals in PW2>10% and Frequency flooding signals in PW3 >10%;

[0082] Among them, the phenological prior classification rules for non-rice crops are: the cultivated land pixels have more than one cloud-free Sentinel-2 surface reflectance data from May to August, and the Frequency flooding signals+0.05 = 0 ;

[0083] Among them, the phenological prior classification rules for winter wheat are: WCI > 0.3 and VH April < -17;

[0084] Among them, the phenological prior classification rules for winter rapeseed are: WCI > 0.3 and VH April >-15;

[0085] Among them, the phenological prior classification rules for non-winter crops are: <WCI < 0.1;

[0086] Among them, Cropland represents the cultivated land layer (0 and 1 represent non-cultivated land pixels and cultivated land pixels, respectively);

[0087] Frequency flooding signals Indicates the time proportion of the first flooding signal of the pixel. The judgment basis for the first flooding signal of the pixel is: Normalized Difference Vegetation Index NDVI < Land Surface Water Index LSWI or Enhanced Vegetation Index EVI < Land Surface Water Index LSWI;

[0088] Pixel RiceCrayfish Indicates whether the phenological prior type of this pixel is a rice-shrimp field. False indicates a non-rice-shrimp field, while True indicates a rice-shrimp field.

[0089] Frequency flooding signals in PW1 and Frequency flooding signals in PW3 They represent the time proportion of the first flooding signal of the pixel in the PW1 period and the PW3 period, respectively. Similarly, the judgment basis of the first flooding signal of the pixel in the PW1 period and the PW3 period is: Normalized Difference Vegetation Index NDVI < Land Surface Water Index LSWI or Enhanced Vegetation Index EVI < Land Surface Water Index LSWI;

[0090] Frequency vegetation signals in PW2 It indicates the time proportion of the vegetation signal of the pixel in the PW2 period. The judgment basis of the vegetation signal of the pixel in the PW2 period is: Normalized Difference Vegetation Index NDVI> Surface Water Index LSWI or Enhanced Vegetation Index EVI> Surface Water Index LSWI.

[0091] Frequency flooding signals+0.05Indicates the time proportion of the second flooding signal of the pixel. The judgment basis of the second flooding signal of the pixel is: Normalized Difference Vegetation Index NDVI < Surface Water Index LSWI + 0.05 or Enhanced Vegetation Index EVI < Surface Water Index LSWI + 0.05.

[0092] PW1, PW2, and PW3 are designated periods. PW1 is from January 1, 2021, to April 30, 2021; PW2 is from July 15, 2021, to September 30, 2021; PW3 is from November 10, 2021, to December 31, 2021; VH April It represents the median value of the time series VH data in April;

[0093] Step S2.2: Based on the phenological prior knowledge corresponding to the phenological prior crop subtypes, the phenological prior knowledge of each phenological prior crop type in the study area is obtained. The phenological prior knowledge of each phenological prior crop type cannot represent the accurate crop type, and the union of the phenological prior knowledge of each phenological prior crop type is calculated as the overall phenological prior knowledge.

[0094] The intersection of rice phenology prior knowledge and non-winter crop phenology prior knowledge is calculated to obtain single-season rice phenology prior knowledge. The phenology prior crop type of each pixel in the single-season rice phenology prior knowledge is single-season rice.

[0095] Calculate the prior knowledge of rice-shrimp field phenology. The prior crop type of each pixel in the prior knowledge of rice-shrimp field phenology is rice-shrimp field.

[0096] The intersection of rice phenology prior knowledge and winter wheat phenology prior knowledge is calculated to obtain winter wheat-rice phenology prior knowledge. The phenology prior crop type of each pixel in the winter wheat-rice phenology prior knowledge is rice-winter wheat.

[0097] The intersection of rice phenology prior knowledge and winter rapeseed phenology prior knowledge is calculated to obtain winter rapeseed-rice phenology prior knowledge. The phenology prior crop type of each pixel in the winter rapeseed-rice phenology prior knowledge is winter rapeseed-rice.

[0098] The intersection of non-rice crop phenology prior knowledge and winter wheat phenology prior knowledge is calculated to obtain winter wheat-other crop phenology prior knowledge. The phenology prior crop type of each pixel in the winter wheat-other crop phenology prior knowledge is winter wheat-other crops.

[0099] The intersection of non-rice crop phenology prior knowledge and winter rapeseed phenology prior knowledge is calculated to obtain winter rapeseed-other crop phenology prior knowledge. The phenology prior crop type of each pixel in the winter rapeseed-other crop phenology prior knowledge is winter rapeseed-other crops.

[0100] The intersection of non-rice crop phenology prior knowledge and non-winter crop phenology prior knowledge is calculated to obtain other crop phenology prior knowledge. The phenology prior crop type of each pixel in the other crop phenology prior knowledge is other crops.

[0101] The union of the above phenological prior crop types is calculated as the overall phenological prior knowledge.

[0102] Step S2.3: Randomly select 10,240 pixels based on the overall phenological prior knowledge. The time series VH data and time series VV data corresponding to these 10,240 pixels are used as unlabeled samples. The true crop category of the unlabeled samples is unknown, and only the phenological prior crop type is known.

[0103] Step S3: Construct a semi-supervised learning network, which includes a backbone network model based on a pseudo-label semi-supervised training strategy and an adaptive threshold adjustment strategy.

[0104] The backbone network model uses a bidirectional LSTM network (DCM model) based on the attention mechanism. The backbone network model is used to extract the labeled deep semantic features and unlabeled deep semantic features corresponding to the labeled samples and unlabeled samples respectively. The probability values ​​corresponding to various crop categories of the labeled samples and unlabeled samples are further calculated based on the labeled deep semantic features and unlabeled deep semantic features.

[0105] Pseudo-label semi-supervised training strategy: select the crop category corresponding to the maximum probability value of the unlabeled sample, select the unlabeled sample whose crop category corresponding to the maximum probability value is consistent with the phenological prior crop type, further select the unlabeled sample whose maximum probability value is greater than the set adaptive threshold, and use the crop category corresponding to the maximum probability value of the further selected unlabeled samples as the pseudo label of the corresponding unlabeled sample;

[0106] Adaptive threshold adjustment strategy: The core idea of ​​the strategy adopted in this step is to dynamically adjust the probability threshold by calculating two key indicators of the pseudo labels of unlabeled samples: (1) the changing trend of the number of pseudo labels in consecutive training cycles; (2) the 10th percentile value (Q10) of the pseudo label prediction probability in the previous training cycle. The adaptive threshold corresponding to the current training cycle is dynamically adjusted based on the changing trend of the number of pseudo labels generated in the previous two training cycles and the 10th percentile value of the pseudo label prediction probability distribution in the previous training cycle. Specifically,

[0107] First, compare the changes in the number of pseudo labels in the previous two training cycles under the current training cycle. If the number of pseudo labels found increases or the number of pseudo labels found decreases by less than or equal to 20%, further calculate the 10th percentile (Q10) of the pseudo label prediction probability of the previous training cycle; when the 10th percentile is greater than or equal to 0.55 and less than 0.6, the adaptive threshold of the current training cycle is set to 0.55; when the 10th percentile is greater than or equal to 0.6 and less than 0.65, the adaptive threshold of the current training cycle is set to 0.6; when the 10th percentile is greater than or equal to 0.65 and less than 0.7, the adaptive threshold of the current training cycle is set to 0.65; when the 10th percentile is greater than or equal to 0.7, the adaptive threshold of the current training cycle is set to 0.7.

[0108] If the number of pseudo-labels found decreases by more than 20%, the 10th percentile (Q10) of the pseudo-label prediction probability for the previous training cycle is calculated. When the 10th percentile is greater than or equal to 0.55 and less than 0.6, the adaptive threshold for the current training cycle is set to 0.5; when the 10th percentile is greater than or equal to 0.6 and less than 0.65, the adaptive threshold for the current training cycle is set to 0.55; when the 10th percentile is greater than or equal to 0.65 and less than 0.7, the adaptive threshold for the current training cycle is set to 0.6; and when the 10th percentile is greater than or equal to 0.7, the adaptive threshold for the current training cycle is set to 0.65. This flexible adjustment mechanism based on actual training performance enables the model to autonomously select the optimal threshold, achieving a dynamic balance between the quality and representativeness of pseudo-label generation.

[0109] Step S4: Calculate the total loss function and train the backbone network model under the semi-supervised learning framework guided by the phenological prior knowledge in step S3 under the condition of minimizing the total loss function.

[0110] Calculate the total loss function , the total loss function Based on the following formula:

[0111]

[0112] in, is the loss function calculated based on labeled samples and predicted crop categories, is the loss function calculated based on unlabeled samples and pseudo labels.

[0113] Loss function calculated based on labeled samples and crop categories Based on the following formula:

[0114]

[0115] in, Indicates the total number of labeled samples in each batch of training in each round of training cycle, Represents the first labeled samples The probability value distribution results are: Indicates the labels for labeled examples; Indicates the calculation of cross entropy loss value.

[0116] The loss function calculated for unlabeled data and pseudo labels as described above Based on the following formula:

[0117] ;

[0118] in, Indicates the total number of labeled samples in each batch of training in each round of training cycle, represents the total number of unlabeled samples for each batch of training in each round, is a multiple; represents the first unlabeled samples The probability value distribution results of ; Represents the adaptive threshold used to judge pseudo labels; Indicates the The maximum probability value of unlabeled samples, For the The pseudo label corresponding to the maximum probability value of unlabeled samples, Represents unlabeled samples == is the matching operator. 1() is the indicator function, which returns a value of 1 when the conditions in the brackets are met and a value of 0 when the conditions in the brackets are not met. Indicates the calculation of cross entropy loss value.

[0119] Step S5: Obtain the time series VH data and time series VV data of the cultivated land pixels in the area to be predicted and input them into the trained backbone network model to obtain the final crop category.

[0120] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0121] The final crop classification of all cultivated land pixels in the predicted area is compared with the crop identification results of the traditional machine learning model (Random Forest RF model) and the backbone network model based on supervised learning (DCM model). The results are evaluated using indicators such as F1-score, User's Accuracy and Producer's Accuracy. Figures 3-5 shown.

[0122] As can be seen from the figure, for the F1-score indicator ( Figure 3 ), the accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for single-season rice, winter wheat-rice, winter rapeseed-rice, winter rapeseed-other crops, rice-shrimp fields and other crops is higher than that of the RF model (pink column) and the DCM model (blue column). The accuracy evaluation of the PhenoSSL DCM model (green column) for winter wheat-other crops is the same as that of the RF model and the DCM model. For the User's Accuracy index ( Figure 4 ), the accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops and other crops is higher than that of the RF model and the DCM model. The accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for winter rapeseed-other crops and shrimp-rice fields is between that of the RF model and the DCM model. Figure 5 ), the accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for single-season rice, winter wheat-rice, winter rapeseed-rice, winter rapeseed-other crops and rice-shrimp fields is higher than that of the RF model and the DCM model. The accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for winter wheat-other crops is weaker than that of the RF model and the same as that of the DCM model. The accuracy evaluation of the method of the present invention (PhenoSSL DCM model) for other crops is weaker than that of the RF model and the DCM model.

[0123] In summary, the method of the present invention (PhenoSSL DCM model) significantly improves the simultaneous recognition effect of multiple types of crops in small sample scenarios. Compared with the random forest RF and the base DCM model, it has higher accuracy and better recognition effect, which greatly alleviates the high dependence of deep learning crop remote sensing identification on field samples.

[0124] Example 2:

[0125] In this embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0126] Example 3:

[0127] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0128] Example 4:

[0129] In this embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0130] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A small-sample crop mapping method using semi-supervised learning guided by phenological prior knowledge, characterized by: The following steps are involved: Step S1: Acquire Sentinel-1 data for the study area and generate time-series VH data and time-series VV data. Acquire Sentinel-2 surface reflectance data for the study area, perform cloud removal on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data. Obtain the crop category and geographic coordinates of each sampling point in the study area, use the time-series VH data and time-series VV data of the pixels corresponding to the sampling points as labeled samples, and use the crop category of the pixels corresponding to the sampling points as labels. Step S2: extract the phenological prior knowledge of each phenological prior crop type in the study area, calculate the union of the phenological prior knowledge of each phenological prior crop type as the overall phenological prior knowledge, and construct unlabeled samples based on the overall phenological prior knowledge; Step S3: constructing a semi-supervised learning network, which includes a backbone network model based on a pseudo-label semi-supervised training strategy and an adaptive threshold adjustment strategy; Step S4: Calculate the total loss function, and train the backbone network model based on the pseudo-label semi-supervised training strategy and adaptive threshold adjustment strategy in step S3 under the condition of minimizing the total loss function; Step S5: Obtain the time series VH data and time series VV data of the cultivated land pixels in the area to be predicted and input them into the trained backbone network model to obtain the final crop category. The phenological prior knowledge of each phenological prior crop type in the study area extracted in step S2 is generated based on the following steps: According to the phenological priori crop types of the study area, the phenological priori crop subtypes of the study area are confirmed, and the phenological priori classification rules corresponding to the phenological priori crop subtypes are confirmed. The phenological priori classification rules are combined with the time series VH data, time series VV data and Sentinel-2 vegetation index data obtained in step 1 to calculate the phenological prior knowledge corresponding to the phenological priori crop subtypes; Based on the phenological prior knowledge corresponding to the phenological prior crop subtypes, the phenological prior knowledge of each phenological prior crop type in the study area was obtained; The step S2 in which unlabeled samples are constructed based on the overall phenological prior knowledge is based on the following steps: Multiple pixels are randomly selected based on the overall phenological prior knowledge, and the time series VH data and time series VV data corresponding to each pixel are used as unlabeled samples.

2. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 1 is characterized in that: The phenological prior crop types in the study area include: single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops, winter rapeseed-other crops, rice-shrimp fields and other crops; The phenological prior crop subtypes in the study area are: rice, rice-shrimp fields, winter rapeseed, winter wheat, non-rice crops, and non-winter crops.

3. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 2 is characterized in that: The phenological priori classification rules corresponding to the phenological priori crop subtypes are as follows: The rice phenology prior classification rule is: the pixel is a cultivated land pixel, the proportion of the time when the first flooding signal of the pixel occurs is greater than 10%, and the phenology prior type of the cultivated land pixel is not a rice-shrimp field; the judgment basis of the first flooding signal of the pixel is: the normalized difference vegetation index NDVI < the land surface water index LSWI or the enhanced vegetation index EVI < the land surface water index LSWI; The phenological prior classification rule of the rice-shrimp field is as follows: the pixel is a cultivated land pixel, the time proportion of the first flooding signal of the pixel in the PW1 period is greater than 60%, the time proportion of the vegetation signal of the pixel in the PW2 period is greater than 10%, and the time proportion of the first flooding signal of the pixel in the PW3 period is greater than 10%; the judgment basis of the first flooding signal of the pixel is: the normalized vegetation index NDVI < the surface water index LSWI or the enhanced vegetation index EVI < the surface water index LSWI, and the judgment basis of the vegetation signal of the pixel is: the normalized vegetation index NDVI > the surface water index LSWI or the enhanced vegetation index EVI > the surface water index LSWI; The phenological prior classification rules for non-rice crops are as follows: the pixel is a cultivated land pixel, the pixel has at least one cloud-free Sentinel-2 surface reflectance data from May to mid-August, and the proportion of time the second flooding signal appears in the pixel is 0; the second flooding signal of the pixel is determined based on: the Normalized Difference Vegetation Index (NDVI) < the Land Water Index (LSWI) + 0.05 or the Enhanced Vegetation Index (EVI) < the Land Water Index (LSWI) + 0.05; The phenological prior classification rules for winter wheat are as follows: the winter crop index of the pixel is greater than 0.3 and the median of the April time series VH data is less than -17; The phenological prior classification rules for winter rapeseed are: the winter crop index of the pixel is greater than 0.3 and the median of the April time series VH data is greater than -15; The phenological priori classification rule for non-winter crops is: the winter crop index of the pixel is greater than 0 and less than 0.1; The PW1 period, the PW2 period and the PW3 period are designated periods.

4. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 3 is characterized in that: The phenological prior knowledge of each phenological prior crop type in the study area was obtained based on the following steps: Calculate the rice phenology prior knowledge and the non-winter crop phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the non-winter crop phenology prior knowledge to obtain the single-season rice phenology prior knowledge; Calculate prior knowledge of rice-shrimp field phenology; Calculate the rice phenology prior knowledge and the winter wheat phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the winter wheat phenology prior knowledge to obtain the winter wheat-rice phenology prior knowledge; Calculate the rice phenology prior knowledge and the winter rapeseed phenology prior knowledge, and calculate the intersection of the rice phenology prior knowledge and the winter rapeseed phenology prior knowledge to obtain the winter rapeseed-rice phenology prior knowledge; Calculate the non-rice crop phenology prior knowledge and winter wheat phenology prior knowledge, and calculate the intersection of the non-rice crop phenology prior knowledge and the winter wheat phenology prior knowledge to obtain the winter wheat-other crop phenology prior knowledge; Calculate the non-rice crop phenology prior knowledge and the winter rapeseed phenology prior knowledge, and calculate the intersection of the non-rice crop phenology prior knowledge and the winter rapeseed phenology prior knowledge to obtain the winter rapeseed-other crop phenology prior knowledge; The non-rice crop phenology prior knowledge and the non-winter crop phenology prior knowledge are calculated, and the intersection of the non-rice crop phenology prior knowledge and the non-winter crop phenology prior knowledge is calculated to obtain the other crop phenology prior knowledge.

5. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 1, characterized in that: The backbone network model in step S3 is a bidirectional LSTM network based on the attention mechanism, which is used to extract the labeled deep semantic features and unlabeled deep semantic features corresponding to the labeled samples and the unlabeled samples respectively, and further calculate the probability values ​​corresponding to various crop categories of the labeled samples and the unlabeled samples based on the labeled deep semantic features and the unlabeled deep semantic features; The pseudo-label semi-supervised training strategy in step S3 is specifically as follows: selecting the crop category corresponding to the maximum probability value of the unlabeled sample, selecting the unlabeled sample whose crop category corresponding to the maximum probability value is consistent with the phenological prior crop type to which it belongs, further selecting the unlabeled sample whose maximum probability value is greater than a set adaptive threshold, and using the crop category corresponding to the maximum probability value of the further selected unlabeled sample as the pseudo label of the corresponding unlabeled sample; The adaptive threshold adjustment strategy in step S3 is specifically as follows: the adaptive threshold corresponding to the current training cycle is dynamically adjusted based on the changing trend of the number of pseudo labels generated in the previous two training cycles and the 10th quantile of the pseudo label prediction probability distribution in the previous training cycle.

6. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 5, characterized in that: The corresponding relationship between the adaptive threshold corresponding to the current training cycle and the changing trend of the number of pseudo labels generated in the previous two training cycles and the 10th quantile of the pseudo label prediction probability distribution in the previous training cycle is as follows: When the number of pseudo labels found increases, or the number of pseudo labels found decreases by less than or equal to 20%, a continuous Q10 distribution interval is set. When Q10 falls in different Q10 distribution intervals, the adaptive threshold selects the corresponding value. The larger the Q10 distribution interval, the larger the adaptive threshold value. When the number of pseudo labels found decreases by more than 20%, the adaptive threshold corresponding to the Q10 distribution interval is smaller than the adaptive threshold corresponding to the same Q10 distribution interval when the number of pseudo labels increases or decreases by less than or equal to 20%. When Q10 falls in different Q10 distribution intervals, the adaptive threshold selects the corresponding value. The larger the Q10 distribution interval, the larger the adaptive threshold value. Among them, Q10 is the 10th percentile value of the pseudo-label prediction probability in the previous training cycle.

7. The small sample crop mapping method based on semi-supervised learning guided by phenological prior knowledge according to claim 6, characterized in that: The calculation of the total loss function in step S4 is based on the following formula: ； ； ； in, is the total loss function, is the loss function calculated based on labeled samples and predicted crop categories, is the loss function calculated based on unlabeled samples and pseudo labels; Indicates the total number of labeled samples in each batch of training in each round of training cycle, Represents the first labeled samples The probability value distribution results are: Indicates the labels for labeled examples; Indicates the calculation of cross entropy loss value; represents the total number of unlabeled samples for each batch of training in each round, is a multiple; Represents the first unlabeled samples The probability value distribution results of ; Represents the adaptive threshold used to judge pseudo labels; Indicates the The maximum probability value of unlabeled samples, For the The pseudo label corresponding to the maximum probability value of unlabeled samples, Represents unlabeled samples The phenological priori crop type, == is the match operator, 1() is the indicator function, which returns a value of 1 when the set conditions in the brackets are met, and returns a value of 0 when the set conditions in the brackets are not met.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the crop mapping method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the crop mapping method according to any one of claims 1 to 7 are implemented.

Images