County-level rice biomass mapping method based on automatic identification of rice field boundary and phenology

Abstract

The invention relates to the technical field of automatic identification of rice field boundaries and phenology, in particular to a county-level rice biomass mapping method based on automatic identification of rice field boundaries and phenology, which comprises the following steps: collecting visible light and multispectral images of an unmanned aerial vehicle in a research area, randomly selecting sampling points to measure rice biomass, and obtaining ground sampling data; adopting a zero sample segmentation, lightweight classification and automatic extraction accurate identification framework to extract a global rice field space boundary and mark a rice growth period of each rice field; on the basis of unmanned aerial vehicle double images and ground sampling data, screening characteristic variables to construct biomass inversion models in different growth periods; a high-precision mask is obtained through a rice field extraction model in the frame, and rice field scale mask classification is completed through a classifier; according to the rice field masks and the growth period labels, pure rice field images in all growth periods are obtained through cutting; and calling a corresponding optimized inversion model, converting the pure image into a biomass distribution diagram, and integrating to form a global complete county rice biomass mapping result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic identification of rice field boundaries and phenology, and more particularly to a county-level rice biomass mapping method based on automatic identification of rice field boundaries and phenology. BACKGROUND

[0002] As one of the most important food crops in the world, accurate estimation of rice biomass is of great significance for food security, agricultural resource management, and climate change research. During key windows such as rice tillering, jointing, and heading, the planting main body needs to obtain real-time biomass maps of the entire county (hundreds to thousands of acres) within 1-2 days to uniformly dispatch unmanned aerial vehicle variable fertilization, plant protection, and water management. However, during continuous rainy weather, satellites are completely ineffective. Even on sunny days, the long revisit period and low resolution of satellites cannot meet the needs of precision agriculture. Research on rice biomass inversion at the field level using unmanned aerial vehicles has matured, but how to extend it to the county level and achieve automatic and rapid generation of rice biomass within the county remains a final hurdle for the implementation of precision agriculture.

[0003] High-precision extraction of rice fields is a prerequisite for automatic mapping of rice biomass. Although traditional classification algorithms such as maximum likelihood and K-Means clustering are widely used, they often fail to achieve satisfactory segmentation results in complex field environments. In recent years, with the rise of deep learning technology, segmentation algorithms based on deep learning have gradually become the mainstream of research. Deep learning models can automatically learn features from images, allowing for more accurate identification of rice fields at the county level. For example, convolutional neural networks (CNNs) have been widely used in remote sensing image classification and segmentation tasks, achieving remarkable results. However, most deep learning models require a large amount of labeled data for training, which is often difficult to obtain in practical applications. In addition, rice field images in different regions have different characteristics, and traditional models need to be retrained and adjusted for each region, increasing the cost and time of model application. Therefore, how to develop a rice field extraction framework that can adapt to different regions and conditions while reducing dependence on labeled data is an important direction for current research.

[0004] Rice biomass not only reflects the growth status of crops, but also is closely related to yield. Traditional biomass estimation methods mainly rely on vegetation indices, such as normalized difference vegetation index and enhanced vegetation index. These indices can reflect the growth status of vegetation, but in complex environments, a single vegetation index often cannot accurately estimate biomass. With the development of machine learning technology, various regression algorithms have been applied to the estimation of rice biomass, such as random forest, support vector regression, etc. However, these methods in the middle and late stages of rice growth, due to high vegetation coverage, the relationship between feature variables and biomass becomes complex, resulting in a decrease in estimation accuracy. How to obtain a high-precision biomass inversion model is also an important task to realize the automatic mapping of rice field biomass. SUMMARY

[0005] Therefore, the present application provides a county-level rice biomass mapping method based on automatic identification of rice field boundaries and phenology, aiming to solve the above technical problems.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions: Comprising the following steps: S1, data acquisition: collecting unmanned aerial vehicle visible light images and unmanned aerial vehicle multispectral images of the study area, and selecting a plurality of sampling points in the study area by random sampling method, and measuring the biomass of rice at each sampling point to obtain ground sampling biomass data; S2, rice field boundary extraction and phenology identification: using a three-stage farmland image precision extraction and phenology identification framework of zero-shot segmentation, lightweight classification and automatic extraction, using the unmanned aerial vehicle visible light images to extract the spatial boundaries of all rice fields in the study area, and marking the growth period (seedling stage, tillering stage, jointing stage, heading stage, mature stage) of rice in each rice field, S3, biomass inversion model construction and biomass calculation: based on the unmanned aerial vehicle multispectral images, unmanned aerial vehicle visible light images and ground sampling biomass data, screening feature variables to construct biomass inversion models for different growth stages of rice, and using the biomass inversion models to calculate the biomass data of each rice field; the feature variables include multispectral vegetation index, visible light vegetation index, visible light texture feature and rice plant height; S4, county-level rice biomass mapping: high-precision masks are obtained by using the rice field extraction model in the framework, and the mask classification at the field scale (such as non-rice field, seedling stage, tillering stage, heading stage, and mature stage) is completed by the classifier; then, based on the rice field mask and the corresponding growth period label, the visible light and multispectral images are cropped to remove the interference of non-rice field areas, and the images are divided into subsets according to the growth period to obtain pure rice field images at each growth period; finally, for rice fields at different growth periods, the corresponding optimized biomass inversion model is called to convert the pure rice field images at each growth period into growth period-specific biomass distribution maps, and finally the complete county-level rice biomass mapping results are integrated.

[0007] Further, the rice field extraction framework of step S2 is specifically as follows, S21, zero-shot mask generation, calling the pre-trained weights of Segment Anything Model (SAM), inputting the UAV visible light image into SAM after tiling, automatically generating the candidate polygon farmland mask of the study area, and without retraining or parameter tuning of SAM; S22, mask filtering and preliminary screening, based on the area threshold and the aspect ratio constraint, removing the fine noise spots with an area much smaller than the real field and the masks not conforming to the conventional shape of farmland field from the candidate polygon farmland mask, to obtain the candidate set of potential farmland masks; S23, lightweight classification and voting decision: a multi-classification data set containing non-farmland categories and five growth periods (seedling stage, tillering stage, jointing stage, heading stage, and mature stage) of rice is constructed, and a MobileNetV3 classifier is trained and selected; each candidate mask in the candidate set of potential farmland masks is randomly window-sampled, and the sample images obtained by sampling are input into the MobileNetV3 classifier to obtain the class probability, and the majority voting method is used to determine the final label of the same candidate mask, and the rice field boundary and phenology recognition is completed.

[0008] S24, extracting all masks labeled as "rice field", performing morphological smoothing and edge regularization, and finally superimposing back to the original UAV image to crop out pure rice field RGB and multispectral images, providing the required data for subsequent biomass inversion.

[0009] Further, before inputting the UAV visible light image into SAM in step S21, the UAV visible light image is first subjected to 10 times down-sampling processing to save computing power and ensure that the segmentation quality of SAM does not decrease significantly.

[0010] Further, the random window sampling in the step S23 satisfies: the pixel size of the sampling window is 256*256; if the area of the candidate mask is less than 4 times the ground area corresponding to the sampling window, the candidate mask is sampled only once; if the area of the candidate mask is greater than or equal to 4 times the ground area corresponding to the sampling window, the sampling times are linearly increased with the increase of the area of the candidate mask, and the sampling times are in the range of 3 times to 10 times.

[0011] Further, the method for obtaining the rice plant height in the step S3 is: generating a digital surface model (DSM) and a digital terrain model (DTM) of the research area through the unmanned aerial vehicle image, and calculating the difference between the DSM and the DTM to obtain the plant height data.

[0012] Further, the characteristic variables used for constructing the biomass inversion model in the step S3 include 4 kinds of multispectral vegetation indexes, 4 kinds of visible light vegetation indexes, 5 kinds of visible light texture characteristics and the plant height, and a total of 14 characteristic variables.

[0013] Further, the plant height is obtained by calculating the difference between the digital surface model and the digital terrain model generated by the unmanned aerial vehicle visible light image, and taking the 95% quantile.

[0014] Further, the plant height variable is included in the biomass prediction model from the rice tillering stage.

[0015] Further, the biomass inversion model is constructed by using the Stacking integrated learning method, and the model training and evaluation are performed through ten-fold cross validation.

[0016] Further, in the step S23, after the multi-classification data set is constructed, the multi-classification data set is divided into a training set and a verification set according to a ratio of 7:3; the training set is used to train the MobileNetV3 classifier, and the verification set is used to verify the accuracy and inference speed of the classifier, so as to ensure that the Top-1 accuracy of the MobileNetV3 classifier is greater than or equal to 0.99 and the parameter amount is less than or equal to 1.6M.

[0017] Compared with the prior art, the present application has the following positive effects: By using the proposed farmland image accurate extraction and phenology identification framework, the present application realizes high-precision extraction and growth period determination of rice fields at a county scale, and completes the automatic mapping of county rice biomass through the constructed biomass prediction model. The present application can realize rapid and automatic monitoring of rice field biomass, and provides technical support for the development of modern agriculture. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only represent the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0019] Figure 1 is the framework of the present application for precise extraction of farmland images and phenology identification.

[0020] Figure 2 is the farmland mask result map of the present application, and the mask label 0 is the seedling stage, and 1 is the tillering stage.

[0021] Figure 3 is the plant height extraction result map of the present application, including seedling stage image plant height calculation, tillering stage image plant height calculation, jointing stage image plant height calculation, and heading stage image plant height calculation.

[0022] Figure 4 is the final result map of the method of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Referring to the accompanying Figures 1-4 , the present embodiment specifically describes a specific implementation of a county-level rice biomass mapping method based on automatic identification of paddy field boundaries and phenology. The county-level rice biomass mapping method based on automatic identification of paddy field boundaries and phenology comprises: S1: Collecting unmanned aerial vehicle visible light images and multispectral images in the study area, and selecting sampling points in the study area by random sampling to measure the rice biomass.

[0025] S2: Using the obtained unmanned aerial vehicle visible light images to extract all paddy fields in the study area and mark the growth period of rice in each paddy field by using the proposed framework for precise extraction of farmland images and phenology identification.

[0026] S3: Constructing a crop biomass inversion model for different growth periods according to the unmanned aerial vehicle images in step S1 and the ground sampling data, so as to obtain rice biomass data for each field according to the crop biomass inversion model for different growth periods.

[0027] S4: Realizing the rice biomass mapping of the study area.

[0028] The rice field extraction framework of the step S2 is specifically as follows: A three-stage framework of "zero-shot segmentation + lightweight classification + automatic extraction" is proposed, which couples the Segment Anything Model (SAM) with a lightweight classifier to realize automatic extraction of rice fields without manual annotation.

[0029] First stage: SAM zero-shot mask generation Directly call the SAM pre-training weight, input the whole study area image patch at one time, and automatically generate candidate polygon mask. This process does not require any retraining or parameter tuning, avoiding the data bottleneck of traditional deep learning models that must collect samples, label, and fine-tune.

[0030] The mask output by SAM contains a large number of fine noise spots (area much smaller than the real field). By area threshold (≥1000 m²) and aspect ratio constraint, obviously non-agricultural objects are quickly removed, and a set of potential farmland mask candidates are obtained, greatly reducing the computational load of subsequent classification.

[0031] Second stage: lightweight classification and voting decision Construct a multi-classification data set of non-agricultural land and different growth stages of rice fields (seedling stage, tillering stage, jointing stage, heading stage, and mature stage), select seven classic networks for training and comparison on the self-built data, and finally select MobileNetV3 with the highest precision and fastest inference speed as the classifier.

[0032] Random window sampling is performed on each candidate mask: window size 256x256 pixels; if the mask area is <4 times the window area, only 1 time is sampled; the sampling times are increased linearly with the increase of the area (minimum 3 times, maximum 10 times). All sampled windows are superimposed on the unmanned aerial vehicle visible light image for sampling, and the collected sample images are named as mask n-m. Where n is the mask number, and m is the first sampling under the same mask. This naming method is convenient for identifying the sample belonging to the mask after classification.

[0033] All collected samples are sent to MobileNetV3, and the class probability is output; the multiple prediction results of the same mask are determined by majority voting to determine the final label.

[0034] Third stage: image automatic cropping Extract all masks labeled as "rice field", perform morphological smoothing and edge regularization, and finally superimpose them back to the original unmanned aerial vehicle image to crop pure rice field RGB and multispectral images, providing the required data for subsequent biomass inversion.

[0035] The rice biomass inversion model of the step S3 is specifically as follows: The variable autocorrelation of multispectral vegetation index and visible light vegetation index was analyzed respectively, and by eliminating the highly autocorrelated vegetation index, 4 multispectral vegetation indexes and 4 visible light vegetation indexes were finally reserved. Combined with 5 visible light texture features and plant height, a total of 14 characteristic variables were used for the construction of the rice biomass prediction model, and the optimal biomass prediction model was trained.

[0036] The biomass prediction of all rice fields was completed through the optimal biomass prediction model, and the results were presented in a visual way, i.e. the final rice biomass distribution map.

[0037] S5: The visible light images of different growth periods in the study area were processed by blocking, and finally 200 images were obtained. In order to save computing power, different degrees of down-sampling operation were performed on the visible light images, and the SAM model was input into the operation to check whether the segmentation quality was affected. After testing, after 10 times of down-sampling, the SAM segmentation result still did not appear obvious decline. Therefore, 10 times down-sampling operation was performed on the 200 images, and then input into the SAM model for segmentation. After obtaining the segmentation result, the non-rice field mask was deleted by comparing the visible light image. At the same time, the rice field mask of each image was drawn as the real mask. Finally, the segmentation effect of SAM model was evaluated through the real mask of visual interpretation, and the average of related evaluation indexes of 200 images was listed in Table 1. From the table, it can be seen that the SAM model performs excellently, and the average of all indexes is close to 1, indicating that the model can accurately segment in most cases.

[0038] Table 1

[0039] S6: The classification data set was made by visual interpretation from the images of different periods in the study area. The sample image size of each class in the data set was 256*256, and the number was more than 500. The data set was divided into training set and validation set according to the ratio of 7 to 3, and the performance of each classification model on the validation set was counted, and the results were shown in Table 2. From the table, it can be seen that the MobileNetV3 model has the best comprehensive performance, and the Top-1 Acc score is only 0.0044 lower than the highest DenseNet model, but its parameter amount is much smaller than other models. Therefore, MobileNetV3 was selected as the classifier of the framework of precise extraction of farmland images and phenology recognition proposed in this method.

[0040] Table 2

[0041] S7: Calculate the 14 feature variables of each sampling point to form a dataset, and train the biomass prediction model using ten-fold cross-validation. The prediction results of the model on the test set are shown in Table 3. As can be seen from the table, the Stacking model is the optimal biomass prediction model, with high prediction accuracy in all four growth stages, indicating that the model can achieve high-precision inversion of rice biomass.

[0042] Table 3

[0043] S8: In this invention, we obtain the plant height of rice by calculating the difference between the digital surface model and the digital terrain model generated from the visible light image of the UAV. To avoid misjudging the maximum value due to outliers such as weeds and shading, we use the 95th percentile to determine the plant height. This method can effectively reflect the actual height of most plants while excluding a very small number of abnormally high values. Figure 3 The correlation analysis results between rice plant height extracted using UAV technology and plant height obtained from field measurements are presented. The figures show that the correlation between plant height extracted at the seedling stage and the measured plant height is low, R0. 2 The correlation coefficient (R0.542) is only 0.542, which may be due to the small size of rice seedlings and the limited resolution of images taken by the drone at an altitude of 80m, resulting in low accuracy in the extracted plant height. However, as the rice enters the tillering stage, the correlation between the extracted plant height and the measured plant height significantly improves, with R0.542. 2 All were above 0.788, and the heading period R 2 The value reached 0.889, indicating that from the tillering stage onwards, the plant height extracted by the drone can accurately reflect the actual plant height and can be used as an effective method for obtaining plant height. Therefore, we also tried to remove the plant height variable and examine the performance of the rice biomass prediction model at different growth stages. The results are shown in Table 4. The contribution of plant height at the seedling stage to rice AGB prediction was relatively low. From the tillering stage onwards, after removing the plant height variable, the prediction accuracy decreased significantly. Therefore, the tillering stage can be used as a dividing point, and from this period onwards, the plant height variable should be added to the rice biomass prediction.

[0044] Table 4

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0046] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A county-level rice biomass mapping method based on automated identification of paddy field boundaries and phenology, characterized in that, Includes the following steps: S1. Data acquisition: Collect UAV visible light images and UAV multispectral images of the study area, and select several sampling points in the study area through random sampling to measure the biomass of rice at each sampling point in the field to obtain ground sampling biomass data. S2. Paddy field boundary extraction and phenological identification: A three-segment farmland image precise extraction and phenological identification framework of zero-sample segmentation, lightweight classification and automatic extraction is adopted. The spatial boundaries of all paddy fields in the study area are extracted using the visible light images of the UAV, and the growth stage of rice in each paddy field is marked. S3. Biomass inversion model construction and biomass calculation: Based on the UAV multispectral imagery, UAV visible light imagery and the ground-sampled biomass data, feature variables are selected to construct a biomass inversion model for different growth stages of rice, and the biomass data of each paddy field is calculated using the biomass inversion model; the feature variables include multispectral vegetation index, visible light vegetation index, visible light texture features and rice plant height. S4. County-level rice biomass mapping: High-precision masks are obtained using the paddy field extraction model in the framework, and the masks are classified at the field scale using a classifier. Then, based on the paddy field mask and the corresponding growth period label, the visible light and multispectral images are cropped to remove interference from non-paddy field areas and divide the images into subsets according to the growth period to obtain pure paddy field images for each growth period. Finally, the corresponding optimized biomass inversion model is called for different growth periods to convert the pure paddy field images for each growth period into biomass distribution maps, and finally integrated to form a complete county-level rice biomass mapping result.

2. The method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 1, characterized in that, The specific framework for paddy field extraction and phenological identification in step S2 is as follows. S21. Zero-sample mask generation: The pre-trained weights of the Segment Anything Model are called, and the UAV visible light image is sliced ​​and input into the Segment Anything Model to automatically generate candidate polygon farmland masks for the study area without retraining or parameter tuning of the Segment Anything Model. S22. Mask filtering and preliminary screening: Based on area threshold and aspect ratio constraints, remove small noise spots with areas much smaller than the real fields and masks that do not conform to the conventional shape of farmland fields from the candidate polygon farmland masks to obtain a potential farmland mask candidate set. S23. Lightweight Classification and Voting Decision: Construct a multi-class dataset containing non-farmland categories and five rice growth stage categories, train and filter to obtain a MobileNetV3 classifier; randomly sample the visible light image corresponding to each candidate mask in the potential farmland mask candidate set, input the sampled image into the MobileNetV3 classifier to obtain the class probability, and use majority voting to determine the final label of the multiple prediction results of the same candidate mask; S24. Extract all masks marked as paddy fields, perform morphological smoothing and edge regularization, and finally overlay them back onto the original UAV image. Crop out the pure RGB and multispectral images of the paddy fields to provide the necessary data for subsequent biomass inversion.

3. The method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 2, characterized in that, In step S21, before inputting the UAV visible light image into SAM, the UAV visible light image is first downsampled by 10 times to save computing power while ensuring that the segmentation quality of Segment Anything Model does not decrease significantly.

4. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 2, characterized in that, The random window sampling in step S23 satisfies the following conditions: the pixel size of the sampling window is 256×256; if the area of ​​the candidate mask is less than 4 times the ground area corresponding to the sampling window, the candidate mask is sampled only once; if the area of ​​the candidate mask is ≥ 4 times the ground area corresponding to the sampling window, the number of samplings increases linearly with the increase of the candidate mask area, and the number of samplings ranges from 3 to 10.

5. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 1, characterized in that, The method for obtaining rice plant height in step S3 is as follows: a digital surface model (DSM) and a digital terrain model (DTM) of the study area are generated using UAV imagery, and the difference between the DSM and DTM is calculated to obtain initial plant height data.

6. The method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 1, characterized in that, The feature variables used in step S3 to construct the biomass inversion model include 4 multispectral vegetation indices, 4 visible light vegetation indices, 5 visible light texture features, and plant height, totaling 14 feature variables.

7. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 5, characterized in that, The plant height was obtained by calculating the difference between the digital surface model and the digital terrain model generated from the visible light image of the UAV, and then taking the 95th percentile.

8. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 7, characterized in that, Plant height was incorporated into the biomass prediction model starting from the rice tillering stage.

9. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology as described in claim 6, characterized in that, The biomass inversion model was constructed using the Stacking ensemble learning method and trained and evaluated using ten-fold cross-validation.

10. A method for county-level rice biomass mapping based on automated identification of paddy field boundaries and phenology, as described in claim 2, is characterized in that... In step S23, after constructing the multi-class dataset, the multi-class dataset is divided into a training set and a validation set in a 7:3 ratio; the MobileNetV3 classifier is trained using the training set, and the accuracy and inference speed of the classifier are verified using the validation set, ensuring that the Top-1 accuracy of the MobileNetV3 classifier is ≥0.99 and the number of parameters is ≤1.6M.

Images