A method for remote sensing classification of large-scale rice areas under complex conditions

Abstract

The application discloses a large-scale regional rice remote sensing classification method under complex conditions. The method is as follows: 1, selecting a source region in the target region; 2, screening and obtaining optical images and SAR images of the target region during the rice growth period, and processing to obtain a composite image; 3, obtaining label data; 4, identifying the spatial distribution map of the rice in the source region based on the training data set; 5, using the composite image in the training data set, taking the spatial distribution map of the rice in the source region obtained in step 3 as the corresponding label data, training the rice recognition model based on deep learning by using the cross-validation method, and obtaining the final rice remote sensing recognition model; 6, inputting the composite image of the target region obtained in step 2 into the rice remote sensing recognition model to obtain the spatial distribution map of the rice planting area of the target region. The application solves the problem that the rice spatial distribution recognition is inaccurate due to the unstable acquisition of time sequence sunny sky images.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of crop remote sensing recognition using machine learning and deep learning technology, and in particular to a large-scale regional rice remote sensing classification method under complex conditions. BACKGROUND

[0002] Rice is an extremely important staple crop, as well as an important source of greenhouse gas emissions and artificial wetlands. Monitoring the planting area and spatial distribution of rice is of great importance to maintaining food security, monitoring greenhouse gas emissions, and protecting the ecological environment. Satellite remote sensing technology, as a new and rapidly developing high-tech technology, has become the most important and feasible way to obtain high-resolution mapping results of large regional rice planting areas due to its low cost, high efficiency, high timeliness, and high resolution. Research on rice remote sensing classification methods using optical satellites and SAR satellites can quickly, accurately, and cost-effectively obtain high-resolution rice planting areas in a large area, which is an important research direction.

[0003] However, due to various factors in some complex classification conditions, current conventional rice remote sensing recognition faces great challenges. Taking the Yangtze River Delta, China's major rice-producing region, as an example, the rice planting patterns in this region are complex (early rice, late rice, single-crop rice), the land is extremely fragmented (for example, in Zhejiang, the per capita cultivated land area is only about 0.5 mu), the background land types are complex (large areas of various economic crops), and the climate is complex and variable (many cloudy and rainy days, especially during the plum rain season, it is difficult to obtain clear optical images). In terms of data sources, conventional rice recognition methods based on time-series remote sensing image data are limited by the long revisit period of medium and high-resolution remote sensing images, making it difficult to obtain stable and continuous clear optical images under complex climate conditions, which makes it difficult to apply methods based on time-series optical remote sensing images. On the other hand, although SAR data can be continuously, stably, and cloud-free, the land in the study area is fragmented and the background is complex, making the noise of SAR images more significant, and the time-series features are easily overwhelmed by the complex background noise, resulting in a significant decrease in classification accuracy. In terms of rice recognition methods, current deep learning-based rice remote sensing recognition can achieve high rice recognition accuracy, but it is still limited by the difficulty, high cost, and low efficiency of obtaining large-scale regional sample data. How to train and obtain a rice recognition model that can be applied in a large-scale area at a relatively low cost is an important research direction.

[0004] In view of this situation, the present application proposes a new method that can be applied to large-scale area rice remote sensing mapping under complex conditions. Through the construction of single-scene clear-sky optical mean value and vegetation index maximum and minimum value images during the whole growth period of rice, the instability of optical satellite data is solved. By using machine learning and deep learning technology, a small number of high-precision point samples are selected in a small range, the source area rice classification result is constructed as the training area of the deep learning model, and then the deep learning model with strong generalization ability is expanded and applied to the whole large-scale area, so as to quickly and accurately extract the high-resolution rice planting area in the target area under large-scale and complex conditions. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies of data and methods in large-scale area rice high-precision recognition under complex conditions. The construction of single-scene clear-sky optical mean value and vegetation index maximum and minimum value images during the whole growth period of rice solves the problem of unstable optical data, and the high-precision recognition of representative source area rice in a small range solves the defect of insufficient sample data of rice recognition model. A large-scale area rice remote sensing classification method under complex conditions is provided.

[0006] A large-scale area rice remote sensing classification method under complex conditions, comprising the following steps:

[0007] Step 1, selecting a source area in the target area.

[0008] Step 2, screening and obtaining optical images and SAR images of the target area including the source area during the growth period of rice, and pre-processing to obtain optical composite images and time-series SAR images. The time-series SAR images and the optical composite images are superimposed to obtain a composite image.

[0009] Step 3, obtaining label data of part or all of the source area; the obtained label data combined with the composite image of the source area obtained in step 2 are used as a training data set.

[0010] Step 4, based on the training data set, using the random forest method, classifying and recognizing the rice to obtain the spatial distribution map of the source area rice;

[0011] Step 5, using the composite image in the training data set, taking the spatial distribution map of the source area rice obtained in step 4 as the corresponding label data, using the cross-validation method, training the rice recognition model based on deep learning to obtain the final rice remote sensing recognition model.

[0012] Step 6, inputting the composite image of the target area obtained in step 2 into the rice remote sensing recognition model to obtain the spatial distribution map of the rice planting area in the target area.

[0013] As preferred, the specific process of obtaining the optical composite image by preprocessing the optical image is as follows: cloud mask processing, NDVI index and LSWI index calculation are performed on the optical image; all cloud mask processed optical images are combined into one clear sky optical image by the mean value synthesis method; according to the NDVI and LSWI indexes, four images of NDVI maximum value, NDVI minimum value, LSWI maximum value and LSWI minimum value are formed and superimposed into the synthesized clear sky optical image as four bands, thereby forming one optical composite image. The NDVI index is the normalized difference vegetation index (Normalized difference vegetation index), and the LSWI index is the land surface water index (Land surface water index).

[0014] As preferred, the time series SAR image is obtained by 12-day mean value synthesis of the SAR image.

[0015] As preferred, the rice recognition model is obtained by combining the LSTM model and the UNET model; the LSTM model processes the time series SAR image to obtain time series features, and the UNET model processes the optical composite image and the time series features to obtain the spatial distribution result of the rice.

[0016] As preferred, the tree number parameter of the random forest classification method in step 4 is set to 100, and the feature number is set to the square root of the total feature number.

[0017] As preferred, in step 3, the label data is divided into rice samples and other ground object samples. The rice samples need to cover early rice, late rice, single-crop rice and other rice types, and the other ground object samples need to cover built-up areas, other vegetation, water bodies and other typical ground objects.

[0018] As preferred, the area of the source region is less than or equal to 5% of the area of the target region.

[0019] As preferred, the rice type in the source region can represent all rice types in the target region.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] This invention is a large-scale rice remote sensing classification method under complex conditions based on a model with strong generalization ability. It addresses the problem of inaccurate rice spatial distribution identification caused by unstable acquisition of time-series clear-sky images by synthesizing the mean clear-sky optical data and the maximum and minimum vegetation index images of the entire rice growth period in the target area. A high-precision rice distribution map of a small-scale representative source area is created and used as training data to train a high-generalization deep learning rice remote sensing identification model, thereby reducing the difficulty of acquiring massive sample data for deep learning models. Finally, based on clear-sky optical data and maximum and minimum vegetation index images of the entire rice growth period in the target area, as well as time-series SAR data and the trained deep learning rice identification model, the method can quickly and accurately obtain large-scale, high-precision results of rice spatial distribution in the target area. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention;

[0023] Figure 2 This is a geographic location map of the target area and source area selected in this embodiment of the invention;

[0024] Figure 3 This is an image showing the source region sample and rice identification results in an embodiment of the present invention;

[0025] Figure 4 This is a remote sensing map of the spatial distribution of rice in the target area in 2022 at a resolution of 10 meters, as shown in this embodiment of the invention.

[0026] Figure 5 This is a map showing the location of verification points for the rice distribution mapping results in this embodiment of the invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] like Figure 1As shown, a large-scale rice remote sensing classification method under complex conditions is provided. First, according to the target area range and the rice growth period, optical and SAR image data of the research area are obtained, and preprocessing steps such as cloud mask, NDVI and LSWI index calculation, SAR image 12-day mean synthesis, optical image full growth period mean synthesis, and NDVI and LSWI maximum and minimum synthesis are performed. Select samples in the source area, perform high-precision rice identification based on all images and the random forest method, and obtain the spatial distribution results of the rice planting area in the source area. On this basis, use the clear sky optical mean and vegetation index maximum and minimum synthesis images of the rice full growth period in the source area and the time series SAR images to train a deep learning rice remote sensing identification model. Use the deep learning rice remote sensing identification model obtained by training, and the optical mean and vegetation index maximum and minimum synthesis images of the entire target area rice full growth period to obtain high-resolution mapping results of the entire target area rice planting area.

[0029] A large-scale rice remote sensing classification method under complex conditions includes the following steps:

[0030] Step 1, according to the rice distribution profile of the target area of the rice remote sensing mapping, select a suitable source area with a proper position and size as the training area of the rice remote sensing identification model;

[0031] Specifically, the source area is inside the target area and has an area much smaller than the target area. The rice and other land cover types in the source area are representative in the target area. For example, Figure 2 As shown, the target area in this embodiment is the entire Zhejiang Province with an area of about 100,000 square kilometers, and the source area is the southern part of Taizhou Wenhuan Plain with an area of only 1600 square kilometers, accounting for about 1.6% of the target area. The source area rice includes different types such as early rice, late rice, and single-season rice, and other land covers include built-up areas, water bodies, other vegetation, facility agriculture (greenhouses), etc. The target area has good representativeness.

[0032] Step 2, screen and obtain all optical images and SAR images of the target area during the rice growth period, and perform preprocessing;

[0033] Specifically, high-resolution satellite images are selected for optical images, such as Sentinel-2 satellite images, which have a resolution of 10 meters; and high-resolution satellite images are selected for SAR images, such as Sentinel-1 satellite images, which have a resolution of 10 meters. In an embodiment, the rice growth period of the source area and the target area lasts from early April to November, so all Sentinel-1 and Sentinel-2 satellite images from April 1 to November 3 are selected; the preprocessing of optical images includes cloud mask processing, NDVI and LSWI index calculation, and the cloud mask processing is performed using the QA band of the Sentinel-2 satellite image, and the NDVI and LSWI index calculation formulas are shown in formulas 1 and 2:

[0034]

[0035]

[0036] In the formula, NIR is the near-infrared band, RED is the red band, and SWIR is the short-wave infrared band.

[0037] All cloud-masked optical images are synthesized into a clear-sky optical image using a mean synthesis method; for NDVI and LSWI indexes, maximum and minimum synthesis methods are used to form four images of NDVI maximum, NDVI minimum, LSWI maximum, and LSWI minimum, which are added as four bands into the synthesized clear-sky optical image to form an optical synthesis image (which contains data of clear-sky optical mean and vegetation index maximum and minimum values during the whole growth period of rice).

[0038] The SAR images obtained by screening are subjected to 12-day mean synthesis to obtain a time-series SAR image; the time-series SAR image is superimposed with the optical synthesis image to obtain a composite image.

[0039] Step 3: Ground surveys, visual interpretation, and other methods are used to obtain label data of typical rice and other main ground object types in the source area. The obtained label data, combined with the composite image of the source area obtained in step 2, are used as a training data set.

[0040] Specifically, the samples are divided into rice samples and other ground object samples, and the rice samples cover early rice, late rice, single-crop rice, and other rice types, and the other ground object samples cover built-up areas, other vegetation, water bodies, and other typical ground objects, as shown in FIG. 2. Figure 3 As shown, a total of 2992 rice sample pixels and 6067 other category sample pixels are selected.

[0041] Step 4: Based on the composite image in the training data set, a random forest method is used to perform high-precision classification and identification of rice to obtain a high-precision spatial distribution map of rice in the source area.

[0042] Specifically, the tree number parameter of the random forest classification method is set to 100, and the feature number is set to the square root of the total number of features. The rice planting area spatial distribution mapping result of the source area obtained by the embodiment is shown in Figure 3 The out-of-bag accuracy of the random forest classification is highest at 99.4039%, lowest at 96.8853%, and average accuracy is 99.0608%.

[0043] Step 5, using the composite image in the training data set, the high-precision spatial distribution map of rice obtained in step 4 is used as the corresponding label data, and the deep learning-based rice recognition model is trained by using the cross-validation method.

[0044] Specifically, the eight-fold cross-validation method is used to train the model, that is, one-eighth of the data is selected from the total samples without replacement each time as the validation data, and the remaining seven-eighths of the data is used as the training data, and the process is repeated eight times. The rice recognition model uses a combination of LSTM model and UNET model, wherein the LSTM model is used to process the time-series SAR data to obtain time-series features, and the UNET model is used to process the optical composite image and the time-series features obtained from the time-series SAR image to finally obtain the spatial distribution result of the rice. In the eight-fold cross-validation, the average F-score of rice recognition is 0.8722, and the standard deviation is 0.0118, indicating that the training accuracy is high.

[0045] Step 6, according to the multiple accuracy verification results of the deep learning rice recognition model obtained by the cross-validation method in step 5, the model with the highest F-score is selected as the final rice remote sensing recognition model. The composite image of the target area obtained in step 2 is input into the rice remote sensing recognition model to obtain the spatial distribution map of the rice planting area in the target area.

[0046] Specifically, in this embodiment, the model with the highest F-score has a verification F-score value of 0.8980, a Kappa coefficient of rice recognition of 0.8897, a mapping accuracy of 90.24%, and a user accuracy of 89.37%, indicating that the model has high accuracy. The model is applied to the entire target area, and the rice recognition result of Zhejiang Province in 2022 is successfully obtained, as shown in Figure 4 From the figure, it can be seen that the rice planting in Zhejiang Province in 2022 is mainly concentrated in the Hangjiahu Plain in the north and the Shaoxing and Ningbo areas, as well as the coastal plain areas of Taizhou and Wenzhou, and the inland areas are mainly distributed in the Jinhua-Quzhou Basin. According to the accuracy verification result of 3895 randomly selected classification result verification points Figure 5 , the overall accuracy of the classification result is 98.36%, the Kappa coefficient is 0.9118, the user accuracy is 92.79%, and the mapping accuracy is 91.42%, indicating the feasibility of the method.

Claims

1. A method for remote sensing classification of rice in large-scale areas under complex conditions, characterized in that: Includes the following steps: Step 1: Select the source region within the target region; Step 2: Select and obtain optical and SAR images of rice during its growth period in the target area, and preprocess them to obtain optical composite images and time-series SAR images; overlay the time-series SAR images with the optical composite images to obtain a composite image. Step 3: Obtain label data for part or all of the source region; combine the obtained label data with the composite image of the source region obtained in Step 2 to form a training dataset; Step 4: Based on the composite images in the training dataset, use the random forest method to classify and identify rice, and obtain the spatial distribution map of rice in the source region. Step 5: Using the composite imagery in the training dataset, the rice spatial distribution map of the source region obtained in Step 3 is used as the corresponding label data. Using the cross-validation method, a deep learning-based rice identification model is trained to obtain the final rice remote sensing identification model. Step 6: Input the composite image of the target area obtained in Step 2 into the rice remote sensing recognition model to obtain the spatial distribution map of the rice planting area in the target area.

2. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: The specific process of obtaining optical composite images through optical image preprocessing is as follows: cloud masking, NDVI index, and LSWI index are calculated on the optical images; all cloud-masked optical images are combined into a clear sky optical image using the mean composite method; based on the NDVI and LSWI indices, four images with the maximum NDVI value, minimum NDVI value, maximum LSWI value, and minimum LSWI value are generated and superimposed as four bands into the composite clear sky optical image to form a single optical composite image.

3. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: Temporal SAR images are obtained by combining SAR images with 12-day averages.

4. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: The rice identification model is obtained by combining the LSTM model and the UNET model; the LSTM model processes the temporal SAR image to obtain temporal features, and the UNET model processes the optical synthetic image and temporal features to obtain the spatial distribution results of rice.

5. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: In step 4, the number of trees in the random forest classification method is set to 100, and the number of features is set to the square root of the total number of features.

6. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: In step 3, the label data is divided into rice samples and other land cover samples.

7. The method for large-scale rice remote sensing classification under complex conditions according to claim 1, characterized in that: The area of ​​the source region is less than or equal to 5% of the area of ​​the target region.

8. The method for large-scale regional rice remote sensing classification under complex conditions according to claim 1, characterized in that: The rice type in the source region can represent all rice types in the target region.

Images