Method for extracting crop type of agricultural field based on medium time sequence and high semantic image clustering

Abstract

The application discloses an agricultural plot crop type extraction method based on medium-time sequence and high-semantic image clustering, and comprises the following steps: preparing data and models; extracting plot images of high-resolution images as clustering samples A; extracting large-plot pure pixel time sequence information as time sequence training samples B; performing image semantic clustering training on the high-resolution plot images to obtain preliminary reliable labels C, and classifying the time sequence training samples B according to data integrity; performing color-based fine adjustment classification on the reliable labels C to obtain new semi-supervised training samples D; obtaining labels E of all large plots and labels F of all small plots; performing supervised learning training on the time sequence complete data samples, and performing time sequence clustering training among the three time sequence data samples; reclassifying the large-plot crop types; reclassifying the time sequence large-plot prediction results H and the clustering large-plot prediction results F by using a confidence algorithm, and combining the clustering small-plot prediction results G to obtain all-plot prediction results I.

Description

Technical Field

[0001] This invention proposes a method for extracting crop types from agricultural plots based on mid-time series and high-resolution semantic image clustering, involving fields such as remote sensing information processing, image semantic clustering, mid-time series sequences, and semi-supervised learning. Background Technology

[0002] With rapid population growth and urban expansion, ensuring food security has become a fundamental and crucial issue for global sustainable development. Agricultural land, as the primary site for agricultural production activities, relies heavily on accurate extraction of its boundaries and types for applications such as agricultural statistics and planning. It is also a vital foundation for improving agricultural production efficiency and promoting agricultural informatization. With the development of remote sensing technology, utilizing macroscopically covered, periodically updated remote sensing imagery for agricultural information extraction has become an important research direction in agricultural remote sensing.

[0003] With the improvement of spatial resolution of remote sensing imagery, the boundaries of agricultural plots can now be clearly presented and applied. However, crop planting types are complex due to the influence of phenological timing, farming habits, market conditions, and other factors, resulting in significant annual and interannual variations. Since mid-resolution imagery has a relatively short coverage period, it is more conducive to observing crop growth changes and is currently the main data source for extracting crop type information. However, this information often cannot accurately match the agricultural plots presented on high-resolution imagery, making it still difficult to accurately extract crop type information at the plot level.

[0004] Agricultural remote sensing crop type extraction can be divided into pixel-based extraction methods and object-based extraction methods. The former is mostly applicable to medium-resolution remote sensing imagery. It distinguishes crop types by classifying each pixel into different categories, utilizing image features (especially crop temporal variation information) and general classification methods. However, its accuracy is generally low, especially for small plots or areas with mixed pixels around plots, making it difficult to meet the needs of precision agriculture. The latter is mostly used for high-resolution imagery with richer spatial details. It first determines the boundaries of agricultural plots through segmentation and then uses various features of the plots to determine the types of crops planted within each plot. It generally has higher boundary accuracy, but type identification is often inaccurate due to data limitations. With the application of deep learning methods, semantic segmentation and instance segmentation actively learn features through operations such as convolution and pooling, classifying agricultural features on high-resolution imagery by pixel. This has significantly improved plot boundary and type extraction, but the learning and extraction process still targets a specific type of data, making it difficult to comprehensively leverage the advantages of multi-resolution and multi-source data observation.

[0005] In recent years, some scholars have used high-resolution imagery to extract plot boundaries and combined it with mid-resolution temporal imagery to extract plot types to obtain accurate plot information. However, due to differences in spatiotemporal resolution between data sources, high requirements for image registration, incomplete information caused by cloud shadows, and the inability to distinguish small plots caused by mixed pixels on mid-resolution images, problems still exist. Therefore, this technology proposes a method for extracting crop types from agricultural plots by fusing mid-resolution and high-resolution information. It uses the spectral characteristics of high-resolution plots as the medium for type association between different plots, and uses the pure pixels on the mid-resolution imagery corresponding to the center point of large plots as the multi-source association medium. This reduces the requirements for image registration, avoids the influence of mixed pixels, and thus improves the accuracy of crop type extraction for plots (especially small plots). Summary of the Invention

[0006] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a method for extracting crop types of agricultural plots based on mid-time series and high-resolution semantic image clustering.

[0007] This invention aims to improve the accuracy of crop type identification for agricultural plots by better integrating the observational advantages of high-resolution and mid-resolution imagery. It assumes that the same crop exhibits the same plot characteristics in high-resolution imagery, thus semantic clustering in high-resolution imagery can distinguish the similarity between different plots. Assuming that the same crop is planted within each plot, the crop type can be determined by the corresponding pixel of the plot's center in mid-resolution imagery. If the plot is large enough, the pixels surrounding the center will obviously reflect the same crop type characteristics. Therefore, based on the determined plot range, the key to crop type extraction lies in the following process: first, identifying key pure crop pixels based on the plot; second, using mid-resolution temporal features to determine the crop type of each pixel; and finally, combining the similarity between high-resolution plots to determine the crop type of each plot.

[0008] The method for extracting crop types from agricultural plots based on mid-resolution temporal and high-resolution semantic image clustering of the present invention comprises the following steps:

[0009] Step 1: Data and Model Preparation. The data can be divided into mid-resolution time-series infographics and high-resolution agricultural plot images. The models include image semantic clustering models, time-series classification models, and time-series clustering models.

[0010] Step 1.1: Prepare the mid-section time-series information map. The mid-section time-series image contains time-series information of each plot, which reflects the vegetation cover of the plot in each month, and thus indicates the type of agricultural crop of the plot.

[0011] Step 1.2: Prepare high-resolution agricultural plot imagery and vector data. The plot imagery contains complete high-resolution images of the administrative region, and the vector data contains the boundaries of each plot within that administrative region. These can be used to batch extract images of individual plots as a dataset.

[0012] Step 1.3: Prepare image clustering and temporal classification models. Select image clustering, temporal classification, and temporal clustering models suitable for the task objectives.

[0013] Step 2: Extract the plot images from the high-resolution image as clustering samples A; select large plots from the medium-resolution image and extract the pure pixel time series information of the large plots as time series training samples B.

[0014] Step 2.1: Select the center point of all plots in the high-resolution image, crop the remote sensing image of appropriate size with these points as the center, and adjust them all to the cluster input size as cluster sample A.

[0015] Step 2.2: Use the polygon inscribed circle algorithm to select and mark the plots that belong to the large plots.

[0016] Step 2.3: Extract the temporal information of the pure pixels at the center of the large land parcel from the mid-series time-series information map as the temporal training sample B.

[0017] Step 3: Perform image semantic clustering training on high-resolution plot images to obtain preliminary reliable labels C, and classify time-series training samples B according to data completeness.

[0018] Step 3.1: Train the cluster head by inputting cluster sample A into the clustering network.

[0019] Step 3.2: Set the credibility threshold for the credible pseudo-labels, and use the trained clustering head to predict the credible pseudo-labels C for the land parcel classification.

[0020] Step 3.3: Classify the time series training sample B into three types of data samples based on data completeness: complete time series data sample B-1, relatively complete time series data sample B-2, and incomplete time series data sample B-3.

[0021] Step 3.4: Provide true labels for the complete time series data sample B-1 as samples for time series classification.

[0022] Step 4: Perform color-based fine-tuning classification on the trustworthy label C to obtain a new semi-supervised training sample D.

[0023] Step 4.1: Trusted labels are obtained from image semantic clustering. The labels are distinguished based on semantics. Further, each label needs to be manually divided into multiple categories based on color on the basis of each semantic label. After the distinction, a new training sample D can be obtained.

[0024] Step 5: Perform semi-supervised training on the semi-supervised training sample D to obtain the labels E for all large plots and the labels F for all small plots.

[0025] Step 5.1: Retrain the classification and clustering model for sample D using semi-supervised learning, select the clustering model, and obtain the clustering prediction labels for all plots.

[0026] Step 5.2: Selectively merge the cluster prediction labels obtained above based on color, that is, merge labels with similar textures but the same color into one label to obtain the final cluster prediction label E.

[0027] Step 5.3: The cluster prediction labels E obtained above are divided into large plot prediction labels F and small plot labels G based on the large plot markings made by the mid-section image.

[0028] Step 6: Perform supervised learning training on the complete time series data sample B-1, and perform time series clustering training among the three time series data samples.

[0029] Step 6.1: Supervised training is performed on the complete time-series data sample B-1. After the model is trained, predictions are made on the complete time-series data sample B-1 to obtain the prediction results of crop types for large plots in the complete time-series data sample B-1.

[0030] Step 6.2: Perform temporal clustering algorithm analysis on the temporal data of the three time-series training samples to obtain the clustering similarity between the relatively complete temporal data sample B-2 and the incomplete temporal data sample B-3 and the complete temporal data sample B-1.

[0031] Step 7: Using the clustering similarity between the relatively complete time series data sample B-2 and the incomplete time series data sample B-3 and the complete time series data sample B-1 as the similarity judgment algorithm parameter, determine the large plot type of the relatively complete time series data sample B-2 and the incomplete time series data sample B-3, thereby reclassifying the crop type of the large plot.

[0032] Step 7.1: Given the crop type prediction results for plot B-1, use the similarity between the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 and the complete time-series data sample B-1 as parameters for the similarity judgment algorithm. The similarity judgment algorithm will then determine whether the crop types of plots in the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 are the same as those in the complete time-series data sample B-1.

[0033] Step 7.2: Combine the crop types of large plots from complete time-series data sample B-1, relatively complete time-series data sample B-2, and incomplete time-series data sample B-3 into a single prediction result H.

[0034] Step 8: Reclassify the time series large land parcel prediction results H and the clustered large land parcel prediction results F using a confidence algorithm, and then combine them with the clustered small land parcel prediction results G to form a total land parcel prediction result I.

[0035] Step 8.1: Use the confidence algorithm to reclassify the prediction results H and F of the two crop types in the same large plot and merge them into a single prediction result.

[0036] Step 8.2: Combine the merged large plot crop type results and small plot prediction results G into a single prediction result I for all plots.

[0037] Preferably, in step 1, the higher the texture differentiation of each crop type within the high-resolution plot image, the better the clustering effect; the temporal information of the medium-resolution temporal information map can better reflect crop types in large plots, while the temporal information of small plots is affected by vegetation outside the plots; the selected image clustering model needs to be able to differentiate the semantics of the image, thereby distinguishing plot textures; the selected temporal classification model needs to be a supervised model to improve classification performance; the selected temporal clustering model needs to be able to cluster temporal data with different levels of completeness.

[0038] Preferably, in step 2, when cropping the high-resolution image of the plot, a suitable uniform size should be adopted according to the resolution of the plot in the high-resolution image. The plot image of the appropriate size will cover more of its own plot area, and the parts outside the plot should be excluded as much as possible, which will make the effect of clustering and distinguishing texture better. The main goal of the mid-segment time-series information map is to distinguish the crop type of large plots, so an appropriate inscribed circle radius threshold should be selected.

[0039] Preferably, in step 3, to obtain trustworthy labels with obvious texture distinctions, it is necessary to set an appropriate trust threshold. Too high a threshold will result in too few trustworthy labels, while too low a threshold will result in unclear texture distinctions. The time series can reflect the vegetation coverage of the plot in each month, and thus reflect the crop type of the plot. The method of distinguishing the completeness of time series data is based on time series information. If the time series information is available at each time, it is considered complete. The more time series information is missing, the less complete it is considered.

[0040] Preferably, in step 4, the trusted label only distinguishes based on texture, while the type classification also requires color classification, which requires manual color classification of each category of the trusted label.

[0041] Preferably, in step 5, after the semi-supervised learning classification prediction, it is necessary to re-merge the different label classes according to the same semantics and the same color to compress the number of labels.

[0042] Preferably, in step 6, supervised training using only data samples with complete time series can better reflect the accuracy of the prediction.

[0043] Preferably, in step 7, the similarity of time series data samples with different levels of completeness can determine the crop type of the plots in the missing data samples, thereby reducing training costs.

[0044] Preferably, in step 8, the clustering prediction results of crop types for the same large plot and the labels of crop types predicted by time series need to be fused to achieve consistency. Therefore, the quality of the fusion algorithm will affect the accuracy of reclassification.

[0045] This invention utilizes the unsupervised nature of image semantic clustering, using only a small number of real labels, which can greatly reduce the workload of feature engineering, solve the problem of inaccurate temporal judgment when small plots or surrounding areas exist in the form of mixed pixels, and improve the accuracy of crop type extraction for large plots.

[0046] This invention has the following advantages: Crop type for a plot is determined by the texture, color, and time series of the plot image. Compared to mid-time series analysis, high-resolution image semantic clustering more intuitively identifies crop type based on plot texture and color. It also solves the problem of inaccurate time series identification for small plots or areas surrounding plots that exist in a mixed pixel form, and increases the accuracy of crop type extraction for large plots. Therefore, a combination of mid-time series analysis and high-resolution image semantic clustering is used. Furthermore, high-resolution image semantic clustering only requires minor color adjustments to trusted labels, while mid-time series analysis only requires a small number of real labels for large plots, significantly reducing the workload of feature engineering. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method of the present invention;

[0048] Figure 2 This is a high-resolution remote sensing image plot example selected in this invention;

[0049] Figure 3 This is a vector map of NDVI information of a centrally located remote sensing image plot in this invention;

[0050] Figure 4 This is the result of combining mid-range temporal data and high-resolution image semantic clustering prediction. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0052] A method for extracting crop types from agricultural plots based on mid-time and high-resolution semantic image clustering includes the following steps:

[0053] Step 1: Data and Model Preparation Stage. The data can be divided into mid-resolution time-series information maps and high-resolution agricultural plot images. The models are divided into image semantic clustering models, time-series classification models, and time-series clustering models.

[0054] Step 1.1: Prepare the mid-section time-series information map. The mid-section time-series information map contains NDVI information for each plot for 12 months. The NDVI information reflects the vegetation cover of the plot in each month, thus indicating the type of agricultural crop of the plot.

[0055] Step 1.2: Prepare high-resolution agricultural plot imagery and vector data. The plot imagery contains complete high-resolution images of the administrative region, and the vector data contains the boundaries of each plot within that administrative region. These can be used to batch extract images of individual plots as a dataset.

[0056] Step 1.3: Prepare image clustering and temporal classification models. Select image clustering, temporal classification, and temporal clustering models suitable for the task objectives.

[0057] Step 2: Extract the plot images from the high-resolution image as clustering samples A; select large plots from the medium-resolution image and extract the pure pixel time series information of the large plots as time series training samples B.

[0058] Step 2.1: For example, select Zongyang County as the target area, select the center point of all plots in the high-resolution image, crop the remote sensing image to a size of 20×20 with these points as the center, and adjust them to a suitable clustering input size of 32×32 as clustering sample A.

[0059] Step 2.2: In the mid-image, the polygon inscribed circle algorithm is used to select large plots. There are three strategies for determining large plots: First, select plots whose inscribed circle radius exceeds a certain value as large plots; second, sort all plots in descending order of their inscribed circle radii and select the top 10% as large plots; third, use the Jenks natural discontinuity classification algorithm to obtain the last category of radius values ​​for all plots' inscribed circle radii. The lower limit of this last category's radius is the lower threshold for determining a plot. Here, the natural discontinuity classification method is chosen.

[0060] Step 2.3: Extract the temporal information and real labels of these marked large land parcels from the mid-series NDVI information map as temporal training samples B.

[0061] Step 3: Perform image semantic clustering training on high-resolution plot images to obtain preliminary reliable labels C, and classify time-series training samples B according to data completeness.

[0062] Step 3.1: Select a suitable clustering network ResNet18 based on the 32×32 size of cluster sample A and train the cluster heads. Set the number of cluster heads to 6.

[0063] Step 3.2: Set the credibility threshold for the credible pseudo-labels. The credible pseudo-labels C for the land parcel classification are predicted by the trained clustering head. Based on multiple tests, the optimal credibility threshold is set to 85%, so that the number of credible pseudo-labels for each label class is sufficient and there is texture distinction between each label class.

[0064] Step 3.3: Classify the time series training sample B into three types of data samples based on data completeness, such as complete time series data sample B-1, relatively complete time series data sample B-2, and incomplete time series data sample B-3.

[0065] Step 3.4: Provide true labels for the complete time series data sample B-1 as samples for time series classification.

[0066] Step 4: Perform color-based fine-tuning classification on the trustworthy label C to obtain a new semi-supervised training sample D.

[0067] Step 4.1: Trustworthy labels are obtained from image semantic clustering. The labels are distinguished based on semantics. It is necessary to further manually divide each label into multiple classes based on color based on each semantic label. After the classification, new training samples D can be obtained. For example, label 1 can be divided into label 1-1, 1-2, and 1-3 according to color, and label 2 can be divided into 2-1, 2-2, and 2-3 according to color.

[0068] Step 5: Perform semi-supervised training on the semi-supervised training sample D to obtain the labels E for all large plots and the labels F for all small plots.

[0069] Step 5.1: Retrain the classification and clustering model for sample D using semi-supervised learning, select the clustering model ResNet18, and obtain the clustering prediction labels for all plots.

[0070] Step 5.2: Selectively merge the cluster prediction labels obtained above based on color. That is, merge labels with similar textures but the same color into one label to obtain the final cluster prediction label E. For example, label 1-1 and label 3-2 have the same color and similar texture, so they can be merged into a new label.

[0071] Step 5.3: The cluster prediction labels E obtained above are divided into large plot prediction labels F and small plot labels G based on the large plot markings made by the mid-section image.

[0072] Step 6: Perform supervised learning training on the complete time series data sample B-1, and perform time series clustering training among the three time series data samples.

[0073] Step 6.1: Supervised training is performed on the complete time-series data sample B-1. After the model is trained, predictions are made on the complete time-series data sample B-1 to obtain the prediction results of crop types for large plots in the complete time-series data sample B-1.

[0074] Step 6.2: Perform temporal clustering algorithm analysis on the temporal data of the three temporal training samples to obtain the clustering similarity between the relatively complete temporal data sample B-2 and the incomplete temporal data sample B-3 and the complete temporal data sample B1.

[0075] Step 7: Using the clustering similarity between the relatively complete time series data sample B-2 and the incomplete time series data sample B-3 and the complete time series data sample B1 as the similarity judgment algorithm parameter, determine the large plot type of the relatively complete time series data sample B-2 and the incomplete time series data sample B-3, thereby reclassifying the crop type of the large plot.

[0076] Step 7.1: Given the crop type prediction results for plot B-1, use the similarity between the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 and the complete time-series data sample B-1 as parameters for the similarity judgment algorithm. The similarity judgment algorithm will then determine whether the crop types of plots in the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 are the same as those in the complete time-series data sample B-1.

[0077] Step 7.2: Combine the crop types of large plots from complete time-series data sample B-1, relatively complete time-series data sample B-2, and incomplete time-series data sample B-3 into a single prediction result H.

[0078] Step 8: Reclassify the time series large land parcel prediction results H and the clustered large land parcel prediction results F using a confidence algorithm, and then combine them with the clustered small land parcel prediction results G to form a total land parcel prediction result I.

[0079] Step 8.1: Use the confidence algorithm to reclassify the prediction results H and F of the two crop types in the same large plot and merge them into a single prediction result.

[0080] Step 8.2: Combine the merged large plot crop type results and small plot prediction results G into a single prediction result I for all plots.

[0081] This invention provides a method for extracting crop types from agricultural plots based on the fusion of mid-time series and high-resolution semantic image clustering. Since crop type extraction relies on the texture, color, and time series of plot images, unlike mid-time series which only provides time series judgment, this invention combines mid-time series and high-resolution semantic image clustering. This solves the problem of inaccurate time series judgment for small plots or areas surrounding plots where mixed pixels exist, and also increases the accuracy of crop type extraction for large plots. By using a small number of real labels for large plots and fine-tuning the reliable labels, a high-precision type analysis for all plots can be obtained.

Claims

1. A method for extracting crop types from agricultural plots based on mid-resolution temporal images and high-resolution semantic images, characterized in that, Includes the following steps: Step 1: Prepare data and models; the data consists of mid-resolution time-series information maps and high-resolution agricultural plot images; the models consist of image semantic clustering models, time-series classification models, and time-series clustering models. Step 1.1: Prepare the mid-section time-series information map; the mid-section time-series image contains time-series information of each plot, which reflects the vegetation cover of the plot in each month, and thus indicates the type of agricultural crop of the plot; Step 1.2: Prepare high-resolution agricultural plot images and vectors; the plot images contain complete high-resolution plot images of the administrative region, and the vectors contain the boundary range of each plot in the administrative region, which are used to extract images of each plot in batches as a dataset; Step 1.3: Prepare the image semantic clustering model, temporal classification model, and temporal clustering model; Select image semantic clustering model, temporal classification model, and temporal clustering model that are suitable for the task objective; Step 2: Extract the plot images from the high-resolution imagery as clustering samples A; select large plots from the mid-resolution imagery and extract the temporal information of the pure pixels at the center of the large plots as temporal training samples B; Step 2.1: Select the center point of all plots in the high-resolution image, crop the remote sensing image of appropriate size with these points as the center, and adjust them all to the cluster input size as cluster sample A; Step 2.2: Use the polygon inscribed circle algorithm to select and mark the plots that are large plots; Step 2.3: Extract the temporal information of the pure pixels at the center of the large land parcel from the mid-time series information map as temporal training sample B; Step 3: Perform image semantic clustering training on high-resolution plot images to obtain preliminary reliable labels C, and classify time-series training samples B according to data completeness; Step 3.1: Input cluster sample A into the clustering network ResNet18 to train the clustering head; Step 3.2: Set the credibility threshold for the credible pseudo-labels, and predict the credible pseudo-labels C for the land parcel classification using the trained clustering head; Step 3.3: Classify the time series training sample B into three types of data samples based on data completeness, including complete data sample B-1, relatively complete data sample B-2, and incomplete data sample B-3; Step 3.4: Provide true labels for the complete data sample B-1 as samples for time series classification; Step 4: Perform color-based fine-tuning classification on the trustworthy label C to obtain a new semi-supervised training sample D; Step 4.1: Trusted labels are obtained from image semantic clustering. The labels are distinguished according to semantics. It is necessary to further manually divide each label into multiple categories based on color on the basis of each semantic label. After the distinction, a new training sample D is obtained. Step 5: Perform semi-supervised training on the semi-supervised training sample D to obtain the labels E for all large plots and the labels F for all small plots; Step 5.1: Retrain the classification and clustering model using semi-supervised learning on the semi-supervised training samples D, select the clustering model, and obtain the clustering prediction labels for all plots; Step 5.2: Selectively merge the cluster prediction labels obtained above based on color, that is, merge labels with small texture differences but the same color into one label to obtain the final cluster prediction label E; Step 5.3: The cluster prediction labels E obtained above are divided into large plot prediction labels F and small plot labels G based on the large plot markings made in the mid-section image; Step 6: Perform supervised learning training on the complete time series data sample B-1, and perform time series clustering training among the three time series data samples; Step 6.1: Supervised training is performed on the complete time series data sample B-1. After the model is trained, predictions are made on the complete time series data sample B-1 to obtain the prediction results of crop types for large plots in the complete time series data sample B-1. Step 6.2: Perform temporal clustering algorithm analysis on the temporal data of the three time-series training samples to obtain the clustering similarity between the relatively complete temporal data sample B-2 and the incomplete temporal data sample B-3 and the complete temporal data sample B-1; Step 7: Using the clustering similarity between the relatively complete time series data sample B-2 and the incomplete time series data sample B-3 and the complete time series data sample B-1 as the similarity judgment algorithm parameters, determine the large plot types of the relatively complete time series data sample B-2 and the incomplete time series data sample B-3, thereby reclassifying the crop types of the large plots; Step 7.1: Given the crop type prediction results of the complete time-series data sample B-1, use the similarity between the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 and the complete time-series data sample B-1 as parameters for the similarity judgment algorithm. The similarity judgment algorithm will then determine whether the crop types of the large plots in the relatively complete time-series data sample B-2 and the incomplete time-series data sample B-3 are the same as those in the complete time-series data sample B-1. Step 7.2: Combine the crop types of large plots from complete time-series data sample B-1, relatively complete time-series data sample B-2, and incomplete time-series data sample B-3 into a single prediction result H; Step 8: Reclassify the time series large land parcel prediction results H and the clustered large land parcel prediction results F using a confidence algorithm, and then combine them with the clustered small land parcel prediction results G to form a total land parcel prediction result I; Step 8.1: Use the confidence algorithm to reclassify the prediction results H and F of the two crop types in the same large plot and merge them into a single prediction result; Step 8.2: Combine the merged large plot crop type results and small plot prediction results G into a single prediction result I for all plots.

2. The method for extracting crop types from agricultural plots based on mid-resolution temporal and high-resolution semantic image clustering according to claim 1, characterized in that: In step 2, when cropping the high-resolution image of the plot, an appropriate uniform size should be adopted according to the resolution of the plot in the high-resolution image. The plot image of the appropriate size will cover more of its own plot area, and the parts outside the plot should be excluded as much as possible, which will make the effect of clustering and distinguishing texture better. The main goal of the mid-segment time-series information map is to distinguish the crop type of large plots, so an appropriate inscribed circle radius threshold should be selected.

3. The method for extracting crop types from agricultural plots based on mid-resolution temporal and high-resolution semantic image clustering according to claim 1, characterized in that: In step 3, to obtain trustworthy labels with clear texture distinctions, it is necessary to set an appropriate trust threshold. If the trust threshold is too high, there will be too few trustworthy labels, and if it is too low, the texture distinctions will not be obvious. The time series can reflect the vegetation coverage of the plot in each month, and thus reflect the crop type of the plot. The method of distinguishing the completeness of time series data is based on the time series information. If the time series information is available at each time, it is considered complete. The more time series information is missing, the less complete it is considered.

4. The method for extracting crop types from agricultural plots based on mid-resolution temporal and high-resolution semantic image clustering according to claim 1, characterized in that: In step 5, after the semi-supervised learning classification prediction, it is necessary to re-merge the different label classes based on the same semantics and the same color to compress the number of labels.

Images