A large-scale crop optimization mapping method based on progressive iterative training

Abstract

The present invention discloses a large-scale crop optimization mapping method based on progressive iterative training, comprising the following steps: selecting an initial training area from a rice remote sensing mapping target area, and then obtaining a certain number of discrete samples from the initial training area; using a random forest method to expand the discrete samples into a spatially continuous sample area, and using the sample data set to train a deep learning model to generate a preliminary crop classification map; testing the preliminary crop classification map to find out areas with serious misclassification, and selecting a certain area as a sample supplement area; completing negative sample mining or new continuous sample sorting according to the regional misclassification category selection, and introducing a model for iterative training, and realizing the optimization of the model for large-scale crop mapping. The present invention improves the sensitivity of the model to misclassification through negative sample mining and continuous sample integration, so that it performs stably under different geographical environments and rice planting types, and improves the reliability of the classification results.

Description

Technical Field

[0001] The present invention relates to the field of crop remote sensing mapping using deep learning technology, and in particular to a large-scale crop optimization mapping method based on progressive iterative training. Background Art

[0002] Crop monitoring is of vital importance in modern agricultural management. It not only helps farmers optimize their planting methods, but also supports policymakers in making scientific decisions. With the increasing global concern about food security, accurate crop mapping has become particularly important. Remote sensing technology combined with deep learning methods provides an efficient and extensive means of crop monitoring. Through remote sensing technology, rich information about crop growth and distribution can be extracted from satellite images, while deep learning technology can mine complex patterns and features from these large amounts of data, thereby improving the accuracy of crop classification and prediction. This combination makes it possible to carry out high-resolution crop mapping of large areas, providing accurate crop information under different environmental and climatic conditions, thereby promoting more precise agricultural management and resource allocation.

[0003] In large-scale crop mapping, the complexity of crop conditions is a major problem. For example, in the middle and lower reaches of the Yangtze River, crop planting shows a high degree of spatial and temporal variation due to the diversity of terrain, climate, and soil conditions. This complexity causes crop species, distribution, and growth status to appear differently in remote sensing images, increasing the difficulty of accurate classification. In addition, there are some areas with special geographical environments and unique crop rotation systems in large-scale target areas. If sample data from these areas is lacking, the accuracy of the model will be greatly reduced. In addition, sample collection and acquisition is also a major challenge. Deep learning models require a large amount of labeled data for training, and collecting these samples requires not only extensive field surveys, but also precise labeling, which is time-consuming and laborious over a large area. Summary of the invention

[0004] The first purpose of the present invention is to address the shortcomings of the prior art and provide a large-scale rice remote sensing mapping method based on step-by-step deep learning, which improves the generalization ability and classification accuracy of the model through sample expansion, negative sample mining and iterative training, and is suitable for precise remote sensing monitoring and mapping of large-scale rice-growing areas.

[0005] A large-scale crop optimization mapping method based on progressive iterative training includes the following steps:

[0006] Step 1: Select an initial crop sample area in the crop remote sensing mapping target area;

[0007] Step 2: Obtain discrete samples from the initial crop sample area through visual interpretation and field investigation; the discrete samples include positive samples and negative samples, wherein the positive samples are target crop samples and the negative samples are non-target crop samples;

[0008] Step 3: Use the random forest method to expand discrete samples into spatially continuous samples, use the spatially continuous samples to train the deep learning model, and generate a preliminary crop classification map;

[0009] Step 4: Check the preliminary crop classification map to determine whether there are serious misclassification areas in the preliminary crop classification map. If so, select a part of the serious misclassification areas as sample supplementary areas, and then complete negative sample mining or new spatial continuous sample sorting according to the regional misclassification situation (over-classification, missed classification), and integrate all sample information to iterate the model and optimize the crop classification map; if not, output the current deep learning model;

[0010] Step 5: Use the deep learning model obtained in step 4 to classify crops in the target area of ​​crop remote sensing mapping and complete large-scale crop mapping.

[0011] Furthermore, in step 1, the initial rice sample area is selected based on the terrain, phenological characteristics and planting system; specifically, the selection and conditions of the initial area are as follows: the orientation of the initial area depends on the spatial distribution of the geographical characteristics of the target area, and the area is appropriately large to ensure sufficient rice samples and other ground samples.

[0012] Furthermore, in step 2, in order to enable the model to learn enough sample features so that it has a certain preliminary spatial generalization ability to meet the crop identification requirements of the entire target area, discrete samples include positive samples (single-season rice, double-season early rice, double-season late rice) and negative samples (buildings, water bodies, grasslands, forests, sunny slopes, shady slopes, corn, wheat and other crops), and both the categories and the number of samples must be sufficient.

[0013] Furthermore, in step 3, the random forest method is used to complete sample augmentation based on the discrete samples obtained in step 2 and the Sentinel-1 SAR time series data and Sentinel-2 optical image time series data of the corresponding area. The deep learning model adopts the semantic segmentation model (SFDLM) based on the fusion of optical image and SAR image time series features.

[0014] Furthermore, in step 4, the area of ​​the sample supplement area is close to the initial crop sample area and contains one or more of the misclassified categories. The sample adjustment includes negative sample mining or new continuous sample arrangement; wherein the negative sample mining is specifically: marking the sample supplement area with only multiple classifications and no target crop samples as a negative sample area; the new continuous sample arrangement is specifically: re-obtaining discrete samples from the supplement area with missed classifications, mixed multiple classifications and containing target crop samples according to steps 2 to 3 and expanding them into spatially continuous samples.

[0015] The second object of the present invention is to provide a large-scale crop mapping system for implementing the above method, comprising the following modules:

[0016] A sample acquisition module is used to select an initial crop sample area from a crop remote sensing mapping target area and obtain discrete samples from the initial crop sample area;

[0017] The sample expansion module is used to expand discrete samples into spatially continuous samples and generate a preliminary crop classification map after inputting them into the deep learning model;

[0018] The verification module is used to judge the classification of the preliminary crop classification map and update the deep learning model;

[0019] A mapping module that uses updated deep learning models for crop classification and large-scale crop mapping.

[0020] A third object of the present invention is to provide an electronic device, comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the above method.

[0021] A fourth object of the present invention is to provide a machine-readable storage medium storing machine-executable instructions, which, when called and executed by a processor, prompt the processor to implement the above method.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] The present invention uses a systematic technical process to gradually expand discrete samples and iterative training, reduce the workload of sample collection and annotation, and enhance the representativeness of the data set, significantly improving the accuracy of rice remote sensing classification, and overcoming the problem of insufficient classification accuracy in large-scale crop mapping in traditional methods. Through negative sample mining and continuous sample integration, the sensitivity of the model to misclassification is improved, and the continuous sample integration technology is used for gradual optimization, which enhances the generalization ability of the model, making it stable in different geographical environments and rice planting types, avoiding the poor classification results and low recognition accuracy in some areas with special planting systems in traditional methods of large-scale crop mapping. In summary, the method of the present invention can effectively identify target crops in areas with sufficiently complex climate and terrain, effectively improve the accuracy of crop classification and prediction, and provide reliable remote sensing monitoring means and decision support tools for agricultural management. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flowchart of a large-scale crop optimization mapping method based on progressive iterative training according to the present invention;

[0025] Figure 2 The geographical location of the target area, initial sample area, and supplementary sample area selected for the present invention;

[0026] Figure 3 To output sample area samples and continuous sample extension results;

[0027] Figure 4 The structure diagram of the SFDLM deep learning model used for classification;

[0028] Figure 5 The preliminary classification results of rice in the target area;

[0029] Figure 6 To supplement the multi-point area and map comparison of the sample area; A is area 1 containing multi-point situations, B is the map comparison of area 1, C is area 2 containing multi-point situations, and D is the map comparison of area 2.

[0030] Figure 7 To supplement the comparison between different misclassification situations of sample areas and other results; A is area 3 containing different misclassification situations, B is other result display of area 3, C is area 4 containing different misclassification situations, and D is other result display of area 4.

[0031] Figure 8 An example of negative sample labeling and new continuous sample integration is given, where A is the negative sample labeling result and B is the new continuous sample production result;

[0032] Fig. 9 This is the final rice identification and classification result for the target area. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples.

[0034] like Figure 1 As shown, it is a flow chart of the large-scale crop optimization mapping method based on progressive iterative training of the present invention. In this embodiment, rice is used as an example for large-scale mapping. First, the initial rice sample area is selected based on the terrain, phenological characteristics and planting system. Then, in this area, rice and non-rice samples are selected by visual interpretation and other means. Next, Sentinel-1 SAR time series images and Sentinel-2 optical time series images are used, and the discrete samples are expanded to the continuous sample area through the random forest algorithm. The expanded sample data will be used for the preliminary training of the deep learning model to generate preliminary crop classification results. If there is a large range of misclassified areas in the preliminary classification results, the sample supplement process is entered, and the production of the supplementary sample data set is completed by selecting negative sample mining or a new continuous sample set according to different misclassification situations, and then the model is iteratively trained and the classification results are updated. This process gradually improves the generalization ability of the model through multiple iterative optimizations until a high-precision crop classification map of the target area is obtained.

[0035] A large-scale crop optimization mapping method based on progressive iterative training includes the following steps:

[0036] Step 1: According to the geographical information and rice planting conditions of the target area of ​​rice remote sensing mapping, select an area of ​​appropriate location and size as the initial training area of ​​the deep learning model;

[0037] Specifically, the initial regional orientation depends on the spatial distribution of the geographical features of the target area, and the area should be appropriately large to ensure sufficient rice samples and other ground feature samples. Figure 2 The present invention selects Zhejiang Province, Jiangsu Province, and Shanghai as the target area, with an area of ​​about 210,000 square kilometers; the initial sample area is located in the southern part of Taizhou City, with a side length of about 40km. The area has both single-season rice and double-season rice, and sufficient typical rice samples can be obtained. At the same time, a variety of other crops are also planted, and other types of land features can also meet the needs of sample acquisition.

[0038] Step 2: Obtain a certain number of discrete samples of rice and other features from the initial area selected in step 1 through visual interpretation and field survey;

[0039] Specifically, in order to enable the model to learn enough sample features and have a certain initial spatial generalization ability to meet the crop identification requirements of the entire target area, discrete samples include positive samples (single-season rice, double-season early rice, double-season late rice) and negative samples (buildings, water bodies, grasslands, forests, sunny slopes, shady slopes, corn, wheat and other crops). The categories and number of samples must be sufficient, such as Figure 3 , a total of 3,000 rice samples and 7,000 other samples were selected.

[0040] Step 3: Based on the discrete samples obtained in step 2, the random forest method is used to expand them into spatially continuous sample areas, and the deep learning model is trained using the sample data set to generate a preliminary crop classification map.

[0041] Specifically, when using the random forest method to expand the continuous sample, the time series Sentinel-1 SAR data synthesized every 12 days during the growing season (April to November) and the time series Sentinel-2 optical image data synthesized every month are used. The completed continuous sample is shown in Figure 3 . Train the semantic segmentation model (SFDLM) based on the fusion of optical and SAR image temporal features. The model structure is as follows: Figure 4 , and then complete the preliminary crop classification map of the target area, such as Figure 5 .

[0042] Step 4: Based on the preliminary crop classification map generated in step 3, use the map image, existing rice classification results, and field survey data to verify and find out the areas with serious misclassification, and select a certain size of area as the sample supplement area;

[0043] Specifically, the area of ​​the sample supplement area should be close to the initial area and have representative misclassification situations (missing classification, over-classification). Figure 6 It shows that the areas where only multi-class cases exist are marked as negative samples. Figure 7 It shows that the areas containing other misclassification cases will be made into new continuous sample areas.

[0044] Step 5: Based on the sample supplement area in step 4, select and complete negative sample mining or new continuous sample arrangement according to the misclassified categories in the sample supplement area, and introduce the model for iterative training to achieve model optimization and improve model accuracy;

[0045] Specifically, the sample supplement area with only multiple points and no rice samples is marked as a negative sample area to complete the negative sample mining; the sample supplement area with missed points, multiple points and rice samples needs to repeat steps 2 and 3 to complete the new continuous sample arrangement. Figure 6The two supplementary sample areas shown in the figure mistakenly identify the hillside as rice. According to the map image, there is no rice planted in the area. Therefore, all pixels in these two areas are marked as negative samples. Figure 7 In the example, two sample areas respectively identify some other crops as rice. In this case, the continuous sample areas are recreated according to steps 2 and 3. Then the new sample data marked and created above are introduced into the model for iterative training. Figure 8 The negative sample labeling results are shown in the examples. Figure 8 A) and the new continuous sample integration results ( Figure 8 B in the figure).

[0046] Step 6: Repeat steps 4 and 5 to eventually develop a deep learning model with strong generalization ability, which can then be applied to large-scale crop mapping.

[0047] Specifically, after continuous iterative training, the accuracy of the classification results is also improved, and finally a high-precision rice classification map of the target area is obtained, see Fig. 9 . Fig. 9 The color of the supplementary sample area represents the sample data set added in each iterative training.

Claims

1. A large-scale crop optimization mapping method based on progressive iterative training, characterized in that: The method comprises the following steps: Step 1: Select an initial crop sample area in the crop remote sensing mapping target area; Step 2: Obtain discrete samples from the initial crop sample area through visual interpretation and field investigation; the discrete samples include positive samples and negative samples, wherein the positive samples are target crop samples and the negative samples are non-target crop samples; Step 3: Based on the discrete samples obtained in step 2 and their corresponding feature data, the random forest method is used to expand them into spatially continuous samples; wherein the feature data are the Sentinel-1SAR time series data and the Sentinel-2 optical image time series data of the discrete samples in the initial crop sample area; then the spatially continuous samples are used to train the semantic segmentation model SFDLM based on the fusion of optical image and SAR image time series features to generate a preliminary crop classification map; Step 4: Check the preliminary crop classification map to determine whether there are serious misclassification areas in the preliminary crop classification map. If so, select a part of the serious misclassification areas as sample supplement areas, and then adjust the samples according to the misclassified categories in the sample supplement areas. Combine the initial crop samples and the adjusted sample data to iterate the model and optimize the crop classification map. If not, output the current deep learning model. Step 5: Use the deep learning model obtained in step 4 to classify crops in the target area of ​​crop remote sensing mapping and complete large-scale crop mapping.

2. The large-scale crop optimization mapping method based on progressive iterative training according to claim 1 is characterized in that: In step 1, the initial rice sample area is selected based on topography, phenological characteristics and planting system.

3. The large-scale crop optimization mapping method based on progressive iterative training according to claim 1 is characterized in that: In step 4, the misclassification categories include missed classification and over-classification.

4. The large-scale crop optimization mapping method based on progressive iterative training according to claim 1, characterized in that: In step 4, the adjustment sample includes negative sample mining or new continuous sample arrangement; The negative sample mining is specifically as follows: marking the sample supplement area where only multiple distribution situations exist and no target crop samples exist as a negative sample area; The new continuous sample arrangement is specifically as follows: the sample supplement area with missed samples, multiple mixed samples and containing target crop samples is re-obtained with discrete samples according to steps 2 to 3 and expanded into spatially continuous samples.

5. A large-scale crop mapping system implementing the method according to any one of claims 1 to 4, characterized in that: Includes the following modules: A sample acquisition module is used to select an initial crop sample area from a crop remote sensing mapping target area and obtain discrete samples from the initial crop sample area; The sample expansion module is used to expand discrete samples into spatially continuous samples and generate a preliminary crop classification map after inputting them into the deep learning model; The verification module is used to judge the classification of the preliminary crop classification map and update the deep learning model; A mapping module that uses updated deep learning models for crop classification and large-scale crop mapping.

6. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method according to any one of claims 1 to 4.

7. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the method described in any one of claims 1 to 4.

Images