Unmanned aerial vehicle self-adaptive fertilization method for crops

Abstract

The invention discloses an unmanned aerial vehicle self-adaptive fertilization method for crops. The method comprises the following steps: acquiring a multispectral image of a to-be-fertilized area; background pixels of the multispectral image are removed, and vegetation indexes, texture features and reflectivity are extracted after dimensionality reduction; after screening, determining a first feature set and a second feature set; generating a weed distribution map according to the first feature set; generating a crop growth distribution diagram I according to the second feature set; adjusting the brightness and color of the multispectral image to obtain a color enhancement image; generating a crop growth distribution diagram II according to the color enhancement diagram; the weed distribution diagram, the first crop growth distribution diagram and the second crop growth distribution diagram are fused, and self-adaptive fertilization is conducted through an unmanned aerial vehicle according to a fusion result diagram. According to the application, crop growth difference and weed interference in the region are comprehensively considered, and personalized fertilization can be carried out for different regions (a crop region, a weed region and a region with different crop growth vigor).

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) agricultural application, and more particularly to an adaptive fertilization method for crops using UAV. Background Art

[0002] With the development of modern agricultural technology, drone fertilization has been widely used as a new agricultural management method. Drone fertilization has the advantages of high efficiency, precision, reduced labor costs, and can fertilize according to the fertilizer requirements of crops. However, the existing drone fertilization methods still have certain limitations, including:

[0003] Most existing methods rely on preset fertilization schemes and fail to fully consider the individual differences in crop growth. That is, during the growth process of crops, they are affected by multiple factors such as soil, climate, and water, which lead to different crop growth and different fertilizer requirements. However, existing methods fail to fully consider the actual needs of crops during growth and cannot make personalized adjustments based on crop growth, resulting in poor fertilization effects.

[0004] At the same time, during the growth of crops, weeds compete with crops for nutrients, water and light, affecting crop growth. However, existing drone fertilization systems often ignore the interference of weeds in the crop area during the fertilization process and are unable to effectively identify and deal with weed interference, resulting in reduced fertilization effects.

[0005] Therefore, how to solve the above technical problems is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, in order to at least partially solve the above technical problems, the present invention provides a method for adaptive fertilization of crops by drone.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for adaptive fertilization of crops by drone, comprising the following steps:

[0009] Acquire a multispectral image of the area to be fertilized;

[0010] Remove the background pixels of the multispectral image, perform dimensionality reduction, and extract vegetation index, texture features, and reflectance;

[0011] Performing feature screening on vegetation index, texture feature and reflectance to obtain a first feature set and a second feature set;

[0012] generating a weed distribution map using a first classification supervision model according to the first feature set;

[0013] Generate crop growth distribution using the second classification supervision model based on the second feature setFigure 1 ;

[0014] Adjust the brightness and color of the multispectral image to obtain a color enhancement map; generate crop growth distribution using a deep learning network based on the color enhancement map Figure 2 ;

[0015] Weed distribution map and crop growth distribution Figure 1 and crop growth distribution Figure 2 The fusion is performed, and the UAV is used for adaptive fertilization according to the fusion result map.

[0016] Preferably, removing background pixels of the multispectral image includes:

[0017] Use the vegetation index DVI grayscale image corresponding to the multispectral image to create a mask file and mask the multispectral image;

[0018] The image segmentation threshold is determined by maximizing the variance between image pixel classes;

[0019] Image segmentation threshold is used to remove background pixels from multispectral images.

[0020] Preferably, principal component analysis is used for dimensionality reduction, and three principal components are retained according to the amount of information.

[0021] Preferably, the vegetation indices include NDVI, DVI, RVI, SAVI, OSAVI, GNDVI and EVI; the texture features include: mean, variance, synergy, contrast, dissimilarity, information entropy, second-order moment and correlation.

[0022] Preferably, feature screening is performed on vegetation index, texture feature and reflectivity; including:

[0023] The first feature set is obtained by using the filtering feature selection method to reduce the feature dimension, and the first feature set includes: the first principal component synergy, information entropy, correlation, the second principal component mean, information entropy, correlation, the third principal component mean, variance, correlation, blue low-pass reflectivity, green low-pass reflectivity, red edge low-pass reflectivity, blue high-pass reflectivity, and red edge high-pass reflectivity;

[0024] Recursive feature elimination is used to reduce the dimension of features and obtain the second feature set, which includes: mean of the first principal component, variance, synergy, dissimilarity, second-order moment, correlation, mean of the second principal component, information entropy, correlation, mean of the third principal component, difference vegetation index, ratio vegetation index, enhanced vegetation index, and red-edge high-pass reflectivity.

[0025] Preferably, the first classification supervision model / the second classification supervision model is a support vector machine, a random forest or an artificial neural network.

[0026] Preferably, the brightness and color of the multispectral image are adjusted to obtain a color enhancement image, including:

[0027] Convert the multispectral image from RGB color space to HSV color space;

[0028] Color enhancement based on saturation and brightness in HSV color space;

[0029] The enhanced image is normalized, and brightness adjustment and denoising are performed using weighted guided filtering and global adaptive logarithmic mapping adjustment methods.

[0030] Preferably, color enhancement is performed based on the saturation and brightness of the HSV color space according to the following formula:

[0031]

[0032] Where I represents the calculated color enhancement map, S and V represent the saturation and brightness of the multispectral image converted to the HSV color space, respectively.

[0033] Preferably, the deep learning network structure includes a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, an Inception block, a connection layer, a fifth convolution block, a global pooling layer and a Softmax classifier connected in sequence; wherein,

[0034] The first convolution block, the second convolution block and the fourth convolution block include a convolution layer, a pooling layer and a normalization layer, respectively;

[0035] The third convolution block and the fifth convolution block include a convolution layer and a normalization layer, respectively.

[0036] Preferably, the convolutional layer in the first convolutional block is expanded.

[0037] It can be seen from the above technical solutions that the present invention discloses a method for adaptive fertilization of crops by drones. Compared with the prior art, the present application comprehensively considers the differences in crop growth and the interference of weeds in the region, and can perform personalized fertilization for different areas (crop areas, weed areas, and areas with different crop growth) to achieve personalized, precise and intelligent drone fertilization, which helps to improve fertilizer utilization, reduce costs, and promote crop growth, increase crop yield and quality, increase agricultural income, and improve agricultural production efficiency.

[0038] Specifically, the beneficial effects also include:

[0039] 1) By combining vegetation index, texture characteristics and reflectivity to identify weed interference, the accuracy of weed identification results can be guaranteed;

[0040] 2) By determining the difference in crop growth based on vegetation index, texture characteristics and reflectivity, combined with the results of auxiliary identification prediction based on color enhancement images, the comprehensiveness, accuracy and real-time nature of identification can be effectively guaranteed;

[0041] 3) Color enhancement images can eliminate the impact of different lighting conditions on image quality, highlight the key information in the image, and make the contrast between different crops more obvious; thereby better revealing the growth status of crops and improving the recognition rate, accuracy and reliability of crop growth distribution maps. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 This is a flow chart of the crop drone adaptive fertilization method of the present invention;

[0044] Figure 2 This is an example diagram of the structure of the deep learning network of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The embodiment of the present invention discloses a method for adaptive fertilization of crops by drone, which can fully consider the growth differences of crops and the interference of weeds during fertilization. That is, the present application can achieve a more accurate fertilization effect, has high practical value and broad application prospects.

[0047] In one embodiment, a method for adaptive fertilization of crops by drones, such as Figure 1 , including the following steps:

[0048] Acquire a multispectral image of the area to be fertilized;

[0049] Remove the background pixels of the multispectral image, perform dimensionality reduction, and extract vegetation index, texture features, and reflectance;

[0050] Performing feature screening on vegetation index, texture feature and reflectance to obtain a first feature set and a second feature set;

[0051] generating a weed distribution map using a first classification supervision model according to the first feature set;

[0052] Generate crop growth distribution using the second classification supervision model based on the second feature set Figure 1 ;

[0053] Adjust the brightness and color of the multispectral image to obtain a color enhancement map; generate crop growth distribution using a deep learning network based on the color enhancement map Figure 2 ;

[0054] Weed distribution map and crop growth distribution Figure 1 and crop growth distribution Figure 2 The fusion is performed, and the UAV is used for adaptive fertilization according to the fusion result map.

[0055] According to an embodiment of the present invention,

[0056] 1. Obtain multispectral images of the area to be fertilized to provide data support for subsequent crop growth analysis and fertilization decisions.

[0057] In this embodiment, a multispectral camera can be mounted on a drone to capture multispectral images of the desired area.

[0058] That is, firstly, the flight route of the drone is planned according to the shape and size of the area to be fertilized to ensure that the entire area can be covered;

[0059] Then determine the drone’s flight altitude to ensure the image has adequate resolution and coverage; and choose a time when there is plenty of sunlight and low wind speed to shoot and scan to reduce the impact of shadows and motion blur.

[0060] Second, remove the background pixels of the multispectral image and extract the vegetation index, texture features and reflectance after dimensionality reduction;

[0061] Since vegetation and non-vegetation have obvious color differences, this application removes non-vegetation background pixels from multispectral images to improve the efficiency of weed recognition and save running time.

[0062] In one embodiment, removing background pixels of a multispectral image includes:

[0063] Use the vegetation index DVI grayscale image corresponding to the multispectral image to create a mask file and mask the multispectral image;

[0064] The image segmentation threshold is determined by maximizing the variance between image pixels. The maximum variance between classes method is a threshold selection method based on the image grayscale histogram. It divides image pixels into two categories. If the difference between the background and the target in the image increases, the difference between the two categories of image pixels will also increase. Therefore, the threshold that maximizes the variance between image pixels in this application is the optimal threshold for segmenting the image.

[0065] Image segmentation threshold is used to remove background pixels from multispectral images.

[0066] Furthermore, in order to effectively remove redundant information between bands of multispectral images, reduce the correlation between bands, concentrate effective information into fewer conversion bands, and characterize the original data information to the maximum extent, the principal component analysis method is used in this embodiment to reduce the dimension of multispectral data, thereby eliminating redundant information in the multispectral data.

[0067] In one embodiment, the transformed multi-band image has 5 principal components, and the information content thereof is 80.42%, 13.93%, 5.00%, 0.48% and 0.17% respectively. Since the first 3 principal components add up to 99.35%, only the first 3 principal components with the most information content are retained, and the remaining 2 principal components are discarded.

[0068] To further optimize the above technical solution, vegetation index, texture features and reflectivity are extracted; among them,

[0069] Vegetation indices include NDVI, DVI, RVI, SAVI, OSAVI, GNDVI and EVI;

[0070] Texture features include: mean, variance, coherence, contrast, dissimilarity, information entropy, second-order moment, and correlation;

[0071] And the reflectivity is obtained by low-pass and high-pass filtering.

[0072] In order to improve the model operation efficiency and reduce the dimension of the data set, this application uses two feature selection methods to reduce the dimension of the original data set; including

[0073] The first feature set is obtained by using the filtering feature selection method to reduce the feature dimension, and the first feature set includes: the first principal component synergy, information entropy, correlation, the second principal component mean, information entropy, correlation, the third principal component mean, variance, correlation, blue low-pass reflectivity, green low-pass reflectivity, red edge low-pass reflectivity, blue high-pass reflectivity, and red edge high-pass reflectivity;

[0074] Recursive feature elimination is used to reduce the dimension of features and obtain the second feature set, which includes: mean of the first principal component, variance, synergy, dissimilarity, second-order moment, correlation, mean of the second principal component, information entropy, correlation, mean of the third principal component, difference vegetation index, ratio vegetation index, enhanced vegetation index, and red-edge high-pass reflectivity.

[0075] In the second feature set, variance describes the degree of discreteness of the data and can reflect the uniformity of crop growth; synergy, dissimilarity, second-order moment, correlation: these statistical features can provide detailed information about the internal relationship of spectral data, which helps to distinguish different crop states; information entropy represents the uncertainty of data, which can reflect the complexity and health status of crop growth; difference vegetation index (DVI), ratio vegetation index (RVI), enhanced vegetation index (EVI), these indices can be used to assess the health and growth status of vegetation, and are highly correlated with crop growth; red edge high-pass reflectance, the spectral reflectance of the red edge area is related to the chlorophyll content and vegetation biomass, which is conducive to identifying crop growth.

[0076] Compared with the first feature set, the second feature set contains multiple indices directly related to vegetation health and growth conditions (such as DVI, RVI, EVI and red-edge high-pass reflectance). Therefore, crop growth is identified based on the second feature set, and weed distribution is identified based on the first feature set.

[0077] In a specific embodiment,

[0078] 3. Generate a weed distribution map using the first classification supervision model according to the first feature set;

[0079] Generate crop growth distribution using the second classification supervision model based on the second feature set Figure 1 ;

[0080] The first classification supervision model or the second classification supervision model is any one of a support vector machine, a random forest or an artificial neural network. In addition, the data set is normalized before classification, and the model is trained using a 10-fold cross validation method.

[0081] 4. Use deep learning networks based on multispectral images to re-identify crop growth to improve the accuracy of growth identification.

[0082] In order to enhance the natural color and texture details of the image, the brightness and color of the multispectral image are first adjusted to obtain a color enhancement map, including:

[0083] Convert the multispectral image from RGB color space to HSV color space;

[0084] Color enhancement is performed based on the saturation and brightness of the HSV color space; the enhancement method is:

[0085]

[0086] Where I represents the calculated color enhancement map, S and V represent the saturation and brightness of the multispectral image converted to the HSV color space, respectively;

[0087] Then, the enhanced image is normalized, and brightness adjustment and denoising are performed using weighted guided filtering and global adaptive logarithmic mapping adjustment methods in sequence.

[0088] In one embodiment, the normalization process and the weighted guided filtering formula used are:

[0089] G csi =Norm(I c )

[0090] L w (x,y)=W GuidF (τ,ε)(I,G csi )

[0091] In the formula, I c is the enhanced image, G c is the normalized representation, Norm is the normalization operation, W GuidF is a weighted guided filter function, τ, ε are the parameters used by the guided filter function, and I is the brightness value of the input enhanced image. In this application, I is calculated as follows:

[0092] I=0.299R+0.587G+0.114B

[0093] Furthermore, the calculation formula of the global adaptive logarithmic mapping adjustment method applied is:

[0094]

[0095] Where I c (x, y) is the brightness value after the initial noise reduction adjustment process, L wmax Indicates the maximum value of the input image brightness, L′ w Represents the logarithmic mean of the brightness values ​​of the input image.

[0096] In a preferred embodiment, the deep learning network structure constructed in this application refers to Figure 2 , including: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, an Inception block, a connection layer, a fifth convolution block, a global pooling layer and a Softmax classifier connected in sequence; wherein,

[0097] The first convolution block, the second convolution block and the fourth convolution block include a convolution layer, a pooling layer and a normalization layer, respectively;

[0098] The third convolution block and the fifth convolution block include a convolution layer and a normalization layer, respectively.

[0099] The present invention is composed of a combination of multiple convolutional layers, pooling layers and normalization layers, which together form a complex feature extraction and classification system.

[0100] The convolution layer is used to extract local features in the image; by using different convolution kernels, the edges, textures and other features in the image can be accurately identified; the pooling layer is used to reduce the feature dimension and reduce the amount of calculation while maintaining important feature information; the normalization layer is used to standardize the input data so that it has zero mean and unit variance, thereby helping to improve the training speed and stability of the model.

[0101] Furthermore, the Inception block extracts features of different scales by combining convolution kernels of different sizes (1x1, 3x3, and 5x5). Finally, these features are fused together through a "connection layer". The design allows the network to capture image features at multiple scales, improving the performance of the model.

[0102] In this embodiment, the global pooling layer is used to convert the feature map into a vector of fixed length, reducing the number of parameters of the model while maintaining the feature information. The Softmax classifier is used to classify the input data and output the probability distribution of each category, thereby obtaining a crop growth distribution map.

[0103] In order to further optimize the above technical solution, the convolution of the first convolutional block is expanded to increase the receptive field, so that the network can capture a wider range of image information, and reduce the number of parameters in the network, reducing the computational complexity and memory requirements. In addition, by expanding the convolution, the network can better capture the detailed features in the image, thereby improving the classification performance of the model, and making the network have better results in processing multispectral images and generating crop growth distribution maps.

[0104] 5. Weed distribution map and crop growth distribution Figure 1 and crop growth distribution Figure 2 The fusion is performed, and the UAV is used for adaptive fertilization according to the fusion result map.

[0105] In an optional embodiment, first, the weed distribution map and the crop growth distribution map are fused by taking a union, that is, all pixels in the two maps are merged together to form an image containing all information.

[0106] Then, adaptive fertilization is implemented based on the fusion results. One of the fertilization strategies includes:

[0107] Combined with the drone's positioning system, when the drone flies to a weed area, it stops fertilizing; when it flies to an area with weaker crop growth, it increases the amount of fertilizer; when it flies to an area with normal growth, it fertilizes according to the pre-set fertilization strategy.

[0108] The present invention uses drones to monitor farmland and can adjust fertilization strategies in real time according to the growth of crops. Through this adaptive fertilization strategy, farmland can be managed more accurately and crop yield and quality can be improved. At the same time, this method can also reduce fertilizer waste, reduce agricultural costs, and achieve sustainable development.

[0109] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0110] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one 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 present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for adaptive fertilization of crops by drone, characterized in that: Acquire a multispectral image of the area to be fertilized; Remove the background pixels of the multispectral image, perform dimensionality reduction, and extract vegetation index, texture features, and reflectance; Performing feature screening on vegetation index, texture feature and reflectance to obtain a first feature set and a second feature set; generating a weed distribution map using a first classification supervision model according to the first feature set; Generate crop growth distribution map 1 using the second classification supervision model according to the second feature set; The brightness and color of the multispectral image are adjusted to obtain a color enhancement map; based on the color enhancement map, a deep learning network is used to generate a second crop growth distribution map; The weed distribution map is fused with crop growth distribution map 1 and crop growth distribution map 2, and adaptive fertilization is performed using drones based on the fusion result map.

2. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: Remove background pixels from multispectral images, including: Use the vegetation index DVI grayscale image corresponding to the multispectral image to create a mask file and mask the multispectral image; The image segmentation threshold is determined by maximizing the variance between image pixel classes; Image segmentation threshold is used to remove background pixels from multispectral images.

3. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: The principal component analysis method was used for dimensionality reduction, and three principal components were retained according to the amount of information.

4. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: Vegetation indices include NDVI, DVI, RVI, SAVI, OSAVI, GNDVI and EVI; Texture features include: mean, variance, coherence, contrast, dissimilarity, information entropy, second-order moment and correlation.

5. The method for adaptive fertilization of crops by drone according to claim 4, characterized in that: Feature screening of vegetation index, texture characteristics and reflectance; including: The first feature set is obtained by using the filtering feature selection method to reduce the feature dimension, and the first feature set includes: the first principal component synergy, information entropy, correlation, the second principal component mean, information entropy, correlation, the third principal component mean, variance, correlation, blue low-pass reflectivity, green low-pass reflectivity, red edge low-pass reflectivity, blue high-pass reflectivity, and red edge high-pass reflectivity; Recursive feature elimination is used to reduce the dimension of features and obtain the second feature set, which includes: mean of the first principal component, variance, synergy, dissimilarity, second-order moment, correlation, mean of the second principal component, information entropy, correlation, mean of the third principal component, difference vegetation index, ratio vegetation index, enhanced vegetation index, and red-edge high-pass reflectivity.

6. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: The first classification supervised model / second classification supervised model is a support vector machine, a random forest or an artificial neural network.

7. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: Adjust the brightness and color of the multispectral image to obtain a color enhancement map, including: Convert the multispectral image from RGB color space to HSV color space; Color enhancement based on saturation and brightness in HSV color space; The enhanced image is normalized, and brightness adjustment and denoising are performed using weighted guided filtering and global adaptive logarithmic mapping adjustment methods.

8. The method for adaptive fertilization of crops by drone according to claim 7, characterized in that: Color enhancement is performed based on the saturation and brightness of the HSV color space according to the following formula; Where I represents the calculated color enhancement map, S and V represent the saturation and brightness of the multispectral image converted to the HSV color space, respectively.

9. The method for adaptive fertilization of crops by drone according to claim 1, characterized in that: The deep learning network structure includes a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, an Inception block, a connection layer, a fifth convolution block, a global pooling layer and a Softmax classifier which are connected in sequence; wherein, The first convolution block, the second convolution block and the fourth convolution block include a convolution layer, a pooling layer and a normalization layer, respectively; The third convolution block and the fifth convolution block include a convolution layer and a normalization layer, respectively.

10. The method for adaptive fertilization of crops by drone according to claim 9, characterized in that: Dilate the convolutional layer in the first convolutional block.

Images