Small sample crop drawing method for semi-supervised learning guided by phenological priori knowledge

Abstract

The invention discloses a phenological priori knowledge guided semi-supervised learning small sample crop drawing method. The method comprises the following steps: acquiring time sequence VH data, time sequence VV data and vegetation index data of a research area; acquiring crop categories and geographic coordinates of each sampling point in the research area, constructing labeled samples, and constructing unlabeled samples based on overall phenological priori knowledge; constructing a semi-supervised learning network; calculating a total loss function and training the backbone network model; and obtaining time sequence VH data and time sequence VV data of cultivated land pixels of a to-be-predicted area, and inputting the time sequence VH data and the time sequence VV data into the trained backbone network model to obtain a final crop category. According to the method, a semi-supervised learning strategy based on pseudo labels is adopted, the high dependence of a crop identification model based on supervised learning on field samples is broken through, and the performance of collecting a small number of samples in a small area for synchronous drawing of multiple crops is improved.

Description

Technical Field

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

[0002] The spatio-temporal distribution data of crops at the regional and even global scales are important scientific data for supporting the adjustment of crop planting structures, crop yield estimation, methane emission estimation, etc., and are of great significance for scientifically allocating agricultural resources to achieve sustainable agricultural development. Compared with field surveys that require a large amount of human, material and time costs, satellite remote sensing has become an effective means for monitoring the spatio-temporal distribution information of crops due to its advantages of large area, synchronous and real-time earth observation.

[0003] With the development of satellites with different spatio-temporal resolutions and multiple sensors, there are many existing crop mapping methods based on remote sensing data, mainly including: methods based on phenology and decision trees, machine learning methods, and deep learning methods. Since the phenology of crops is highly correlated with their own physical and chemical characteristics, agricultural measures, etc., and does not change significantly with time and space, the method based on phenology and decision trees uses remote sensing data to establish a rule set to highlight the unique phenological characteristics of crops, and uses specific thresholds for rule judgment to identify crops. Although this method improves the efficiency of large-scale crop mapping, it highly depends on the image quality of the key phenological periods of crops and specific thresholds, and it is difficult to use this method to synchronously extract multiple crops. The crop mapping method based on machine learning refers to using machine learning models such as random forest (RF), support vector machine (SVM), and artificial neural network (ANN) to identify different crops by constructing a mathematical relationship between remote sensing image features and crop samples. However, the above-mentioned machine learning models need to carry out artificial feature engineering to select appropriate classification features, and the time dependence relationship between temporal remote sensing images is not considered in model training. Since deep learning models have the ability to automatically extract and learn the deep features of remote sensing images from a large number of training samples, they have been widely used in crop remote sensing mapping in recent years and achieved good results. However, deep learning models based on supervised learning require a sufficient number, reliable quality, and strong representativeness of samples to learn the complex spectral features of different crops. In contrast, semi-supervised learning can reduce the dependence on field samples while ensuring classification accuracy by combining a small number of field samples with a large amount of unlabeled remote sensing data. Therefore, how to design a suitable semi-supervised learning framework, make full use of the rich information contained in a small number of field samples collected in a small area and a large amount of unlabeled remote sensing data obtained by satellites, and alleviate the high dependence of existing crop mapping methods on a large number of samples is the key to achieving efficient crop mapping in large areas. Summary of the Invention

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

[0005] To achieve the above purpose, the following is the technical solution of the present invention:

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

[0007] Step S1: Obtain Sentinel-1 data of the study area and generate time-series VH data and time-series VV data, obtain Sentinel-2 surface reflectance data of the study area, perform cloud removal processing on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data; obtain the crop categories and geographical 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 categories 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: Construct a semi-supervised learning network, and the semi-supervised learning network 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 the adaptive threshold adjustment strategy in step S3 under the condition of minimizing the total loss function;

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

[0012] As described above, the extraction of the phenological prior knowledge of each phenological prior crop type in the study area in step S2 is generated based on the following steps:

[0013] Confirm the phenological prior crop subtypes in the study area according to the phenological prior crop types in the study area, confirm the phenological prior classification rules corresponding to the phenological prior crop subtypes, and calculate the phenological prior knowledge corresponding to the phenological prior crop subtypes by combining the time-series VH data, time-series VV data and Sentinel-2 vegetation index data obtained in step 1;

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

[0015] In step S2, constructing the unlabeled samples based on the overall phenological prior knowledge is based on the following steps:

[0016] Randomly select multiple pixels from the overall phenological prior knowledge, and the corresponding time-series VH data and time-series VV data of each pixel are used as unlabeled samples.

[0017] The phenological prior crop types in the study area as described above 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 prior classification rules corresponding to the phenological prior crop subtypes as described above are as follows:

[0020] The phenological prior classification rule for rice is: the pixel is an arable land pixel, and the time proportion of the first flooding signal of the pixel is greater than 10%, and the phenological prior type of the arable land pixel is not a rice - shrimp field; 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) < LSWI;

[0021] The phenological prior classification rule for rice - shrimp fields is: the pixel is an arable 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 for the first flooding signal of the pixel is: NDVI < LSWI or EVI < LSWI, and the judgment basis for the vegetation signal of the pixel is: NDVI > LSWI or EVI > LSWI;

[0022] The phenological prior classification rule for non - rice crops is: the pixel is an arable land pixel, there is more than 1 cloud - free Sentinel - 2 surface reflectance data in the pixel from May to August, and the time proportion of the second flooding signal of the pixel = 0; the judgment basis for the second flooding signal of the pixel is: NDVI < LSWI + 0.05 or EVI < LSWI + 0.05;

[0023] The prior phenological classification rule for winter wheat is that the winter crop index of the pixel is greater than 0.3 and the median value of the VH data in April is less than -17;

[0024] The prior phenological classification rule for winter rapeseed is that the winter crop index of the pixel is greater than 0.3 and the median value of the VH data in April is greater than -15;

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

[0026] Both the PW1 period, the PW2 period, and the PW3 period are specified periods.

[0027] The prior phenological knowledge of each prior phenological crop type in the study area as described above is obtained based on the following steps:

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

[0029] Calculate the prior phenological knowledge of rice-crayfish fields;

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

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

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

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

[0034] Calculate the prior phenological knowledge of non-rice crops and non-winter crops, and calculate the intersection of the prior phenological knowledge of non-rice crops and non-winter crops to obtain the prior phenological knowledge of other crops.

[0035] As described above, 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: select the crop category corresponding to the maximum probability value of the unlabeled sample, select the unlabeled samples whose crop category corresponding to the maximum probability value is consistent with the prior crop type of the phenological period, further select the unlabeled samples 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 samples;

[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 change trend of the number of pseudo-labels generated in the previous two training cycles and the 10th percentile of the pseudo-label prediction probability distribution in the previous training cycle.

[0038] The corresponding relationship between the adaptive threshold corresponding to the current training cycle as described above and the change trend of the number of pseudo-labels generated in the previous two training cycles and the 10th percentile of the pseudo-label prediction probability distribution in the previous training cycle is as follows:

[0039] When the number of identified pseudo-labels increases, or the decrease amplitude of the number of identified pseudo-labels is less than or equal to 20%, set continuous Q10 distribution intervals. When Q10 falls into 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 decrease amplitude of the number of identified pseudo-labels is greater than 20%, the adaptive threshold corresponding to the Q10 distribution interval is smaller than the adaptive threshold corresponding to the same Q10 distribution interval in the case where the increase or decrease amplitude of the number of pseudo-labels is less than or equal to 20%. When Q10 falls into 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] As described above, the calculation of the total loss function in step S4 is based on the following formula:

[0043] ;

[0044] ;

[0045] ;

[0046] Among them, is the total loss function, is the loss function calculated based on the labeled samples and the predicted crop categories, is the loss function calculated based on the unlabeled samples and the pseudo-labels; represents the total number of labeled samples in each batch of training in each training cycle, represents the th labeled sample predicted by the backbone network model probability value distribution result, represents the th label of the labeled sample; represents calculating the cross-entropy loss value; represents the total number of unlabeled samples in each batch of training in each round, is a multiple; represents the th unlabeled sample predicted by the backbone network model probability value distribution result; represents the adaptive threshold for judging the pseudo-labels; represents the th maximum probability value of the unlabeled sample, is the th pseudo-label corresponding to the maximum probability value of the unlabeled sample, represents the phenological prior crop type of the unlabeled sample . == is the conforming operator, and 1() is the indicator function, which returns 1 when the set conditions in the parentheses are met and returns 0 when the set conditions in the parentheses are not met.

[0047] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned crop mapping method are implemented.

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

[0049] The present invention has the following beneficial effects compared with the prior art:

[0050] By adopting a semi-supervised learning strategy based on pseudo-labels, the present invention breaks through the high dependence of the crop recognition model based on supervised learning on field samples. At the same time, phenological prior knowledge is introduced, and an adaptive threshold strategy is designed to improve the quantity, quality, and representativeness of the pseudo-labels generated in the semi-supervised learning strategy based on pseudo-labels, enhance the richness of the samples used for model training, and thus improve the accuracy of crop mapping in the small-sample scenario.

[0051] The present invention effectively improves the performance of synchronous 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 is a distribution map of the research area of the present invention;

[0054] Figure 3 is a diagram of the evaluation results of the F1-score index of the crop classification accuracy of the present invention in the research area;

[0055] Figure 4 is a diagram of the evaluation results of the User’s Accuracy index of the crop classification accuracy of the present invention in the research area;

[0056] Figure 5 is a diagram of the evaluation results of the Producer’s Accuracy index of the crop classification accuracy of the present invention in the research area. Detailed Embodiments

[0057] To facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail below with reference to examples. The embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0058] Example 1:

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

[0060] Step S1: Obtain Sentinel-1 data of the study area and generate time-series VH data and time-series VV data, obtain Sentinel-2 surface reflectance data of the study area, perform cloud removal on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data; meanwhile, obtain the crop categories and geographical 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. 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 categories of the pixels corresponding to the sampling points as labels (see the distribution of the study area and field sampling points in Figure 2 ) The detailed operation steps are as follows:

[0061] Step S1.1: Collect Sentinel-1 data of the study area with a time range 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 Sentinel-1 data; perform preprocessing on both the VH backscatter coefficient data and VV backscatter coefficient data in the IW mode of Sentinel-1 data, including removing high-incidence angle data in the image overlay area, 20-day median synthesis, Refined Lee spatial filtering, and adaptive order polynomial time 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 of the study area with a time range from September 30, 2020 to December 31, 2021 based on the GEE platform; for Sentinel-2 surface reflectance data, use the Sentinel-2 cloud probability product provided by the GEE platform to remove the pixels with a corresponding cloud probability higher than 65% to obtain the cloud-removed Sentinel-2 surface reflectance data, and calculate the Sentinel-2 vegetation index data of the cloud-removed Sentinel-2 surface reflectance data. The Sentinel-2 vegetation index data includes normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), land surface water index (LSWI), and winter crop index (WCI). The formulas are as follows:

[0063]

[0064]

[0065]

[0066]

[0067] Among them, , , and respectively represent the spectral reflectance of blue, red, near-infrared, and short-wave infrared of Sentinel-2 surface reflectance data. , and respectively refer to the maximum NDVI of winter crops in the vigorous growth stage, the minimum NDVI in the early growth stage, and the minimum NDVI in the late growth stage; the time ranges of the vigorous growth stage Heading, the early growth stage EGS, and the late growth stage LGS in the study area are from February 1, 2021 to May 1, 2021, from September 30, 2020 to December 1, 2020, and from May 1, 2021 to July 1, 2021, respectively.

[0068] Step S1.3: Field personnel went to the study area in late April 2021 and late August 2021 respectively to conduct on-site sampling, used handheld GPS to record the crop categories and geographical coordinates of each sampling point in the study area, associated the pixels on the time-series VH data and time-series VV data with the sampling points based on the geographical coordinates, used the time-series VH data and time-series VV data of the pixels corresponding to the sampling points as labeled samples, and used the crop categories of the pixels corresponding to the sampling points as labels. In this case, 64 labeled samples were selected as the research data for application examples;

[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. 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. The true crop categories of the pixels of the unlabeled samples are unknown, and only the phenological prior crop types are available. The specific steps include:

[0070] Step S2.1: Confirm the phenological prior crop subtypes in the study area according to the phenological prior crop types in the study area, confirm the phenological prior classification rules corresponding to the phenological prior crop subtypes, and the phenological prior classification rules combine 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 subtypes.

[0071] In some embodiments, 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 - crayfish fields, and other crops. Then the phenological prior crop subtypes in the study area are: rice, rice - crayfish fields, winter rapeseed, winter wheat, non - rice crops, and non - winter crops.

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

[0073] Based on the phenological prior classification rules for rice - crayfish fields, the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), and the Land Surface Water Index (LSWI), the phenological prior knowledge of rice - crayfish fields is calculated.

[0074] Based on the phenological prior classification rules for winter rapeseed, the Winter Crop Index (WCI), and the time - series VH data, the phenological prior knowledge of winter rapeseed is calculated.

[0075] Based on the phenological prior classification rules for winter wheat, the Winter Crop Index (WCI), and the time - series VH data, the phenological prior knowledge of winter wheat is calculated.

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

[0077] Based on the phenological prior classification rules for non - winter crops and the Winter Crop Index (WCI), the phenological prior knowledge of non - winter crops is calculated.

[0078] Calculate the intersection of the cropland pixels in the Globeland30 data and the CNLUCC land - cover data in the study area to obtain the cropland layer, which is used as the prior knowledge of cropland.

[0079] Calculate the intersection of the non - cropland pixels in the Globeland30 data and the CNLUCC land - cover data to obtain the non - cropland layer, which is used as the prior knowledge of non - cropland.

[0080] Among them, the phenological prior classification rules for rice are: Cropland = 1 and Frequency flooding signals > 10% and Pixel RiceCrayfish = False; Cropland being 1 means this pixel is cropland.

[0081] Among them, the phenological prior classification rules for rice - crayfish 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 prior phenological classification rule for non - rice crops is: there is more than 1 cloud - free Sentinel - 2 surface reflectance data in the arable land pixels from May to August, and Frequency flooding signals+0.05 = 0;

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

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

[0085] Among them, the prior phenological classification rule for non - winter crops is: 0 < WCI < 0.1;

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

[0087] Frequency flooding signals represents 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 represents whether the prior phenological type of this pixel is a rice - crayfish field. False means non - rice - crayfish field, and True means rice - crayfish field;

[0089] Frequency flooding signals in PW1 and Frequency flooding signals in PW3 respectively represent the time proportion of the first flooding signal of the pixel in the PW1 period and the PW3 period. Similarly, the judgment basis for 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 represents the time proportion of the vegetation signal of the pixel in the PW2 period. The judgment basis for the vegetation signal of the pixel in the PW2 period is: Normalized Difference Vegetation Index NDVI > Land Surface Water Index LSWI or Enhanced Vegetation Index EVI > Land Surface Water Index LSWI,

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

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

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

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

[0095] Calculate the phenological prior knowledge of rice-crayfish fields. The phenological prior crop type of each pixel in the phenological prior knowledge of rice-crayfish fields is rice-crayfish fields.

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

[0097] Calculate the intersection of the phenological prior knowledge of rice and the phenological prior knowledge of winter rape to obtain the phenological prior knowledge of winter rape-rice. The phenological prior crop type of each pixel in the phenological prior knowledge of winter rape-rice is winter rape-rice.

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

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

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

[0101] Calculate the union of the above phenological prior crop types as the overall phenological prior knowledge.

[0102] Step S2.3: Randomly select 10,240 pixels from the overall phenological prior knowledge. The temporal VH data and temporal VV data corresponding to the 10,240 pixels are used as unlabeled samples. The true crop categories of the unlabeled samples are unknown, and only the phenological prior crop types are available.

[0103] Step S3: Construct a semi - supervised learning network. The semi - supervised learning network 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 selects a bidirectional LSTM network based on the attention mechanism (DCM model). Through the backbone network model, the labeled deep semantic features and unlabeled deep semantic features corresponding to the labeled samples and unlabeled samples are extracted respectively. Further, the probability values corresponding to various crop categories of the labeled samples and unlabeled samples are 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 samples. Select the unlabeled samples whose crop category corresponding to the maximum probability value is consistent with the belonging phenological prior crop type. Further, select the unlabeled samples 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 samples.

[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 the unlabeled samples: (1) The change 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 change trend of the number of generated pseudo - labels in the previous two training cycles and the 10th percentile of the pseudo - label prediction probability distribution in the previous training cycle. Specifically:

[0107] First, compare the change in the number of pseudo-labels in the first two training cycles of the current training cycle. If the number of identified pseudo-labels increases, or the decrease in the number of identified pseudo-labels is less than or equal to 20%, further calculate the 10th percentile value (Q10) of the predicted probabilities of the pseudo-labels in the previous training cycle; when the 10th percentile value is greater than or equal to 0.55 and less than 0.6, the adaptive threshold for the current training cycle is set to 0.55; when the 10th percentile value is greater than or equal to 0.6 and less than 0.65, the adaptive threshold for the current training cycle is set to 0.6; when the 10th percentile value is greater than or equal to 0.65 and less than 0.7, the adaptive threshold for the current training cycle is set to 0.65, and when the 10th percentile value is greater than or equal to 0.7, the adaptive threshold for the current training cycle is set to 0.7.

[0108] If the decrease in the number of identified pseudo-labels is greater than 20%, calculate the 10th percentile value (Q10) of the predicted probabilities of the pseudo-labels in the previous training cycle; when the 10th percentile value is in the range of 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 value 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 value 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 value 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 the actual training performance enables the model to autonomously select the optimal threshold and achieve a dynamic balance between the quality and representativeness of pseudo-label generation.

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

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

[0111]

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

[0113] The loss function calculated based on the labeled samples and the crop categories is based on the following formula:

[0114]

[0115] where represents the total number of labeled samples in each batch of training in each training cycle, Denote the th labeled sample predicted by the backbone network model, denote the label of the th labeled sample; denote the calculation of the cross-entropy loss value.

[0116] The loss function calculated from the unlabeled data and pseudo-labels as described above is based on the following formula:

[0117] ;

[0118] where denotes the total number of labeled samples in each batch of training in each training epoch, denotes the total number of unlabeled samples in each batch of training in each epoch, is a multiple; denote the th unlabeled sample obtained in step 2.3 of the prediction by the backbone network model, denote the adaptive threshold for judging pseudo-labels; denote the th maximum probability value of the unlabeled sample, is the th pseudo-label corresponding to the maximum probability value of the unlabeled sample, denote the phenological prior crop type of the unlabeled sample , == is the equality operator. 1() is the indicator function, which returns 1 when the condition set in the parentheses is satisfied and 0 when the condition set in the parentheses is not satisfied, denote the calculation of the cross-entropy loss value.

[0119] Step S5, Obtain the temporal VH data and temporal 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 of ordinary skill in the art can understand that all or part of the processes in implementing the method of the above embodiments can be completed by instructing relevant 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 above method embodiments.

[0121] The final crop categories of all cultivated land pixels in the area to be predicted are compared with the crop recognition results of traditional machine learning models (random forest RF model) and backbone network models based on supervised learning (DCM model). Metrics such as F1-score, User’s Accuracy, and Producer’s Accuracy are used to evaluate the accuracy of the results, and the evaluation results are as Figures 3 - 5 shown.

[0122] As can be seen from the figure, for the F1-score metric ( Figure 3 ), the accuracy evaluations of single-season rice, winter wheat-rice, winter rapeseed-rice, winter rapeseed-other crops, rice-shrimp fields, and other crops of the method of the present invention (PhenoSSL DCM model) are higher than those of the RF model (pink column) and the DCM model (blue column). The accuracy evaluation of winter wheat-other crops of the PhenoSSL DCM model (green column) is the same as that of the RF model and the DCM model; for the User’s Accuracy metric ( Figure 4 ), the accuracy evaluations of single-season rice, winter wheat-rice, winter rapeseed-rice, winter wheat-other crops, and other crops of the method of the present invention (PhenoSSL DCM model) are higher than those of the RF model and the DCM model. The accuracy evaluations of winter rapeseed-other crops and shrimp paddy fields of the method of the present invention (PhenoSSL DCM model) are between the RF model and the DCM model; for the Producer’s Accuracy metric ( Figure 5 ), the accuracy evaluations of single-season rice, winter wheat-rice, winter rapeseed-rice, winter rapeseed-other crops, and rice-shrimp fields of the method of the present invention (PhenoSSL DCM model) are higher than those of the RF model and the DCM model. The accuracy evaluation of winter wheat-other crops of the method of the present invention (PhenoSSL DCM model) is weaker than that of the RF model and is the same as that of the DCM model. The accuracy evaluation of other crops of the method of the present invention (PhenoSSL DCM model) 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 synchronous recognition effect of multiple types of crops in a small-sample scenario, has higher accuracy and better recognition effect compared with the random forest RF and the base DCM model, and greatly alleviates the high dependence of deep learning crop remote sensing recognition on field samples.

[0124] Example 2:

[0125] In this embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[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 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 only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. A small-sample crop mapping method based on phenological prior knowledge-guided semi-supervised learning, characterized in that Including the following steps: Step S1: Obtain Sentinel-1 data of the study area and generate time-series VH data and time-series VV data, obtain Sentinel-2 surface reflectance data of the study area, perform cloud removal on the Sentinel-2 surface reflectance data, and calculate Sentinel-2 vegetation index data; obtain the crop categories and geographical 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 categories 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: Construct a semi-supervised learning network, and the semi-supervised learning network 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 under the condition of minimizing the total loss function, train the backbone network model based on the pseudo-label semi-supervised training strategy and the adaptive threshold adjustment strategy in Step S3; 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 categories.

2. The small-sample crop mapping method based on phenological prior knowledge-guided semi-supervised learning according to claim 1, wherein the extraction of the phenological prior knowledge of each phenological prior crop type in the study area in Step S2 is generated based on the following steps: Confirm the phenological prior crop subtypes in the study area according to the phenological prior crop types in the study area, confirm the phenological prior classification rules corresponding to the phenological prior crop subtypes, and calculate the phenological prior knowledge corresponding to the phenological prior crop subtypes by combining the time-series VH data, time-series VV data, and Sentinel-2 vegetation index data obtained in Step 1; Obtain the phenological prior knowledge of each phenological prior crop type in the study area based on the phenological prior knowledge corresponding to the phenological prior crop subtypes; The construction of unlabeled samples based on the overall phenological prior knowledge in Step S2 is based on the following steps: Randomly select multiple pixels from the overall phenological prior knowledge, and use the time-series VH data and time-series VV data corresponding to each pixel as unlabeled samples.

3. The small-sample crop mapping method based on phenological prior knowledge-guided semi-supervised learning according to claim 2, wherein 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.

4. The small-sample crop mapping method based on phenological prior knowledge-guided semi-supervised learning according to claim 3, wherein The phenological prior classification rules corresponding to the phenological prior crop subtypes are as follows: The phenological prior classification rule for rice is as follows: the pixel is an arable land pixel, and the time proportion of the first flooding signal of the pixel is greater than 10%, and the phenological prior type of the arable land pixel is not a rice-crayfish field; 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) < LSWI; The phenological prior classification rule for the rice-crayfish field is as follows: the pixel is an arable 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 for the first flooding signal of the pixel is: NDVI < LSWI or EVI < LSWI, and the judgment basis for the vegetation signal of the pixel is: NDVI > LSWI or EVI > LSWI; The phenological prior classification rule for non-rice crops is as follows: the pixel is an arable land pixel, there is more than 1 scene of cloud-free Sentinel-2 surface reflectance data in the pixel from May to August, and the time proportion of the second flooding signal of the pixel = 0; the judgment basis for the second flooding signal of the pixel is: NDVI < LSWI + 0.05 or EVI < LSWI + 0.05; The phenological prior classification rule for winter wheat is as follows: the winter crop index of the pixel is greater than 0.3 and the median value of the VH data in April is less than -17; The phenological prior classification rule for winter rapeseed is as follows: the winter crop index of the pixel is greater than 0.3 and the median value of the VH data in April is greater than -15; The phenological prior classification rule for non-winter crops is as follows: the winter crop index of the pixel is greater than 0 and less than 0.1; The PW1 period, PW2 period, and PW3 period are all specified periods.

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

6. The small-sample crop mapping method 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. 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; The pseudo-label semi-supervised training strategy in step S3 is specifically: select the crop category corresponding to the maximum probability value of the unlabeled sample, select the unlabeled samples whose crop category corresponding to the maximum probability value is consistent with the phenological prior crop type, further select the unlabeled samples 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; The adaptive threshold adjustment strategy in step S3 is specifically: the adaptive threshold corresponding to the current training cycle is dynamically adjusted based on the change trend of the number of pseudo-labels generated in the previous two training cycles and the 10th percentile of the pseudo-label prediction probability distribution in the previous training cycle.

7. The small-sample crop mapping method guided by phenological prior knowledge according to claim 6, characterized in that, The corresponding relationship between the adaptive threshold corresponding to the current training cycle and the change trend of the number of pseudo-labels generated in the previous two training cycles and the 10th percentile of the pseudo-label prediction probability distribution in the previous training cycle is as follows: When the number of found pseudo-labels increases, or the decrease amplitude of the number of found pseudo-labels is less than or equal to 20%, set continuous Q10 distribution intervals. 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 decrease amplitude of the number of found pseudo-labels is greater than 20%, the adaptive threshold corresponding to the Q10 distribution interval is smaller than the adaptive threshold corresponding to the same Q10 distribution interval in the case where the increase or decrease amplitude of the number of pseudo-labels is 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.

8. The small-sample crop mapping method based on phenological prior knowledge-guided semi-supervised learning according to claim 7, characterized in that The calculation of the total loss function in step S4 is based on the following formula: ； ； ； Among them, is the total loss function, is the loss function calculated based on the labeled samples and the predicted crop categories, is the loss function calculated based on the unlabeled samples and the pseudo-labels; represents the total number of labeled samples in each batch of training in each training cycle, represents the th labeled sample predicted by the backbone network model probability value distribution result, represents the th label of the labeled sample; represents calculating the cross-entropy loss value; represents the total number of unlabeled samples in each batch of training in each round, is a multiple; represents the th unlabeled sample predicted by the backbone network model probability value distribution result; represents the adaptive threshold for judging pseudo-labels; represents the th maximum probability value of the unlabeled sample, is the pseudo-label corresponding to the th maximum probability value of the unlabeled sample, represents the phenological prior crop type of the unlabeled sample . == is the equality operator, and 1() is the indicator function, which returns 1 when the conditions set in the parentheses are met and returns 0 when the conditions set in the parentheses are not met.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the crop mapping method described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the crop mapping method described in any one of claims 1 to 8.

Images