Estimation Scheme of Rice Yield Based on UAV Hyperspectral Images

FT-NIR Spectrometer,fourier transform spectrometer,ftir spectrometer,ftir analysis,Fourier Transform Infrared Spectroscopy,spectral analysis,quantitative analysis,qualitative analysis,diffuse reflection,specular reflection,integrating sphere,FT-IR Spectrometer-Advanced Manufacturer,FT-NIR Spectrometers；ftir spectrometer,Optosky is a manufacturer of advanced FT-IR spectrometers, with products in stock and at affordable prices.,lntegrating Sphere,integrating sphere,specular reflection,diffuse reflection,qualitative analysis,quantitative analysis,spectral analysis,Fourier Transform Infrared Spectroscopy,ftir analysis,ftir spectrometer,fourier transform spectrometer,FT-IR Spectrometer-Advanced Manufacturer,FT-NIR Spectrometers；ftir spectrometer,Optosky is a manufacturer of advanced FT-IR spectrometers, with products in stock and at affordable prices.

author: sherry

2023-01-17

Yield Estimation Scheme of Rice Based on UAV Hyperspectral Images
I Introduction
Rice is currently the most important food crop in China. Monitoring the planting area and yield of rice has always been highly valued. Both the government and scientific researchers are committed to studying how to understand and accurately grasp the information of rice planting area and yield in a timely manner. At the same time, the timely acquisition of rice planting information can provide a scientific basis for the government to formulate agricultural production policies, which is of great significance for ensuring food security. With the rapid development of social economy, the requirements of various government decision-making departments for the cost and accuracy of production estimation have been further increased. How to effectively absorb and utilize advanced technology to gradually improve the production estimation method is an urgent problem to be solved at present.

Real map of paddy fields
The growth process of rice is affected by many factors, such as cultivation mode, variety, natural factors, irrigation and fertilization treatments, etc. Chlorophyll can indicate the physiological and biochemical state of crops during growth, and it can be used to evaluate the degree of stress (such as pests and diseases, heavy metal stress on crops), growth cycle, productivity, photosynthetic capacity, etc. of crops. Not only that, the chlorophyll content is also closely related to the gross primary production (Gross Primary Production) and crop yield. Leaf area index (LAI), chlorophyll (Chl) content, and gross primary productivity (GPP) are the most common parameters in the study of vegetation biochemical parameters retrieved by remote sensing technology. The nutrient element status of crops affects the crop canopy morphology, leaf area and intrinsic physiological characteristics, so it is closely related to the spectral characteristics. In this way, massive hyperspectral data can be used to quickly, conveniently and non-destructively diagnose nutritional status in the field.
Traditional crop yield estimation models are established based on physical processes, so they have clear biological significance. However, for agricultural problems involving large areas, it is difficult to deal with the distribution characteristics of the surface, and physical yield estimation models can only select representative points for simulation. Remote sensing technology can monitor the ground object environment in a large area, and can obtain remote sensing images of a unified area under continuous event status to realize dynamic monitoring. In addition, compared with traditional manual measurement, remote sensing technology will not destroy the vegetation structure and save time and effort. Nowadays, with the rapid development of remote sensing technology, the ground resolution, spectral resolution and time resolution are constantly improving, and the quantity and quality of remote sensing data are also gradually improving. The future application space is very large.
UAV Remote Sensing (UAVRS) combines unmanned aerial vehicle technology with remote sensing technology, POS positioning technology and communication technology. And analysis to provide real-time and rapid technical solutions, with strong advantages of automation, intelligence and specialization. The ATH9010NY UAV crop health index inspection system developed by our company can obtain real-time high spatial resolution and high spectral resolution remote sensing image data maps, and has the irreplaceable advantages of satellite remote sensing and ground remote sensing technology. It can provide multi-angle high-resolution images, which have obvious advantages compared with the limitations of the narrow field of view and small working range of ground remote sensing.

Schematic diagram of ATH9010 UAV hyperspectral flight
II Technical ideas and main content
Use the near-ground UAV platform to obtain imaging hyperspectral data, and carry out rice yield estimation based on machine learning and deep convolutional neural network models. First, hyperspectral image preprocessing is performed based on hyperspectral image data processing software; secondly, deep convolutional neural network is used. The model accurately extracts rice fields, and then analyzes the spectral information and spatial texture information of rice based on the extracted distribution and area information of rice fields. Finally, the machine learning algorithm is used to achieve accurate evaluation of rice yields, thereby providing a basis for rice planning and management. For reference, the specific technical route is shown in Figure 3 below.

图3 水稻产量估产技术流程图
What can be achieved:
Ø1) Acquire large-scale UAV hyperspectral image data, and based on our self-developed UAV hyperspectral image data processing software, the preprocessing of hyperspectral data can be quickly realized.
Ø2) Based on the preprocessed UAV hyperspectral image, using the deep convolutional network model, it can quickly, accurately and automatically obtain a wide range of rice distribution information and rice planting area information, so as to realize real-time monitoring of rice planting information , to provide a reasonable basis for rice planting planning.
Ø3) Based on the extracted rice distribution information, screen the sensitive bands that are highly correlated with rice yield, and then use machine learning algorithms to realize accurate estimation of rice yield and obtain the results of rice yield distribution, so as to provide yield quality management and planting planning for the agricultural sector for reference.

Ø4) Integrate and deploy all algorithms into hyperspectral image processing software to realize algorithm reuse and one-piece operation.

III Application
1 Automatic extraction of rice fields
With the development of computer science and artificial intelligence technology, convolutional network has become a research hotspot in image segmentation and pattern recognition. It learns from abundant training samples autonomously through continuous convolutional calculations of each convolutional layer. This end-to-end processing mode effectively improves the accuracy and speed of image segmentation. The convolutional neural network is also more advantageous in solving the discontinuity problem caused by the block boundary, which can better solve the problem of insufficient discrimination between the rice spectrum and the background spectrum, and it still has a good performance in more complex scenes. Image Interpretation Capability and Robustness. Therefore, the automatic extraction of rice fields this time is implemented using a convolutional neural network model. The specific schematic diagram is shown in Figure 4 below.

Rice automatic extraction technology roadmap

Convolutional neural network is a feed-forward neural network with convolution calculation and deep structure. It is one of the most commonly used representative algorithms for deep learning. It uses deep convolution to simulate the hierarchical perception and local receptive field in the human visual perception mechanism to process some unstructured data and integrate low, medium and high levels of features in an end-to-end manner, thus Obtain rich feature information and improve the accuracy of semantic segmentation of remote sensing images. The basic structure of a convolutional neural network generally includes a data input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Its basic model framework is shown in Figure 5.

The basic model framework of convolutional neural network

When using the convolutional neural network for different ground object classification tasks, you can build a suitable network model according to specific purposes and requirements, or you can use the existing classic network model at this stage. This section will discuss the classic U-Net and residual The difference network structure is introduced to lay the foundation for subsequent model optimization.

Net：

The U-Net convolutional neural network was proposed by Ronneberger et al. in 2015. It is based on the expansion and modification of the full convolutional neural network. It was first applied to cell cutting in the biomedical field and has a high performance for small sample data. Accuracy. The U-Net convolutional network has a clear structure and a symmetrical U-shape. It is a left-encoding-right-decoding structure. The encoding part gradually increases the number of channels through convolution operations, and uses the maximum pooling operation to gradually reduce the spatial dimension. In the decoding part, the number of channels is gradually reduced through the deconvolution operation, and at the same time, the details of the image and the restoration of the spatial dimension are realized. The most critical part is to add the feature map after the convolution operation of the encoding part to the decoding part through a cascading method in the middle of encoding and decoding, so as to realize the effective combination of deep and shallow feature information and increase the detailed information of the segmentation results. It makes the target details of the decoding part better restored, and the specific structure is shown in Figure 6.

Net convolutional neural network structure

The encoding part of U-Net in the figure contains 4 downsampling modules, each downsampling block consists of a convolutional layer and a pooling layer, the number of feature channels is doubled during each downsampling process, and the feature map The size of the feature map is reduced to half of its original size. After four times of downsampling, the number of channels of the feature map finally becomes 1024, while the length and width of the original image are reduced to 1/16 of the original. Due to the loss of edge pixels during the convolution operation, in the upsampling part, a shortcut connection is used to copy and crop the feature map on the left to the upsampling part of the corresponding level to better preserve the details of the image. In the upsampling (decoding) part, it also contains 4 upsampling blocks. The image size is restored by deconvolution operation, and the number of channels is reduced. Finally, a 1*1 convolution is used to map the 64 component element vectors. To the required number of categories, the image segmentation is completed, and the entire network has a total of 23 convolutional layers.
2 Rice Yield Estimation
1) Vegetation Index Construction and Vegetation Index Sensitive Band Screening
Vegetation indices such as normalized difference vegetation index NDVI, enhanced vegetation index EVI2, ratio vegetation index RVI, difference vegetation index DVI, soil-adjusted vegetation index SAVI, OSAVI and MSAVI have been widely used in the growth monitoring of different crops and the inversion of chlorophyll content , Nitrogen diagnosis and yield estimation research. Therefore, this time I tried to use the existing bands to establish the following common vegetation indices for rice yield estimation research. The specific calculation method of the vegetation index is shown in the formula.

In the formula, NIRi is the reflectance value of a certain band i in the near-infrared band range (760nm-900nm), and REDj is the reflectance value of a certain band j in the red light band range (620nm-760mn).
The NDVI, RVI, DVI and EVI2 vegetation indices under different combinations of red light and near-infrared bands were calculated by exhaustive method. Correlation analysis was carried out between the vegetation index and yield data calculated by all band combinations in all key growth periods, and the band combination with the highest Pearson correlation coefficient was selected as the best sensitive band for each vegetation index in the corresponding growth period.

2)Rice Yield Estimation Model Establishment

The rice production estimation modeling method obtains the corresponding optimal yield estimation vegetation index by performing correlation analysis on the output of each selected sensitive vegetation index, and then uses the multiple linear regression modeling method to obtain the combined six Rice Yield Estimation Model. Afterwards, the experimental data is used for modeling training and verification, and the best production estimation model in the combination is screened out through parameters such as model evaluation index R 2 , MAPE and RMSE. The specific formula is as follows.