A grassland coverage change detection method based on unmanned aerial vehicle images

Abstract

The application discloses a grassland coverage change detection method based on unmanned aerial vehicle images, comprising the following steps: collecting real-time and historical remote sensing images of a grassland area to be detected by using a remote sensing image acquisition unmanned aerial vehicle; performing multi-feature extraction and fusion on the real-time remote sensing images and the historical remote sensing images respectively by using an image processing technology, obtaining a data set based on the fused features, and dividing the data set into a training set and a test set; constructing a BP neural network model; training the BP neural network model by using the training set; and detecting the grassland coverage change by using the test set and the trained BP neural network model. The application can solve the backward mode for grassland vegetation coverage monitoring at present, adopts a ground monitoring device and an air monitoring device, is flexible in networking, has a long transmission distance and high reliability, and realizes rapid, efficient, convenient and accurate ground-air integrated grassland vegetation monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grassland detection, and particularly relates to a grassland coverage change detection method based on images of unmanned aerial vehicles. BACKGROUND

[0002] At present, there are still many problems in the use and management of grassland resources. For example, there are heavy losses every year due to grassland fires, diseases and insect pests; the problem of random reclamation and illegal occupation of grassland land is serious in some local areas; grassland overloading grazing, unreasonable utilization mode, unscientific livestock structure and production mode, and insufficient disaster prevention and mitigation measures. Therefore, how to effectively detect the coverage of grassland vegetation is an urgent problem at present.

[0003] On the other hand, with the rapid development of the Internet and the Internet of Things, at present, when used for grassland detection, only limited technical means such as cameras, fences and other devices can be used to provide video detection, which cannot meet the efficient detection requirements of grassland vegetation. How to more quickly, efficiently and conveniently realize the detection of grassland vegetation is a problem that practitioners urgently need to solve. SUMMARY

[0004] The present application aims to solve the problems of the prior art and provides a grassland coverage change detection method based on images of unmanned aerial vehicles, which can solve the backward way of monitoring the coverage of grassland vegetation at present. The ground monitoring device and the air monitoring device are used, the network is flexible, the transmission distance is far, and the reliability is high; the rapid, efficient, convenient and accurate ground-air integrated grassland vegetation monitoring is realized.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A grassland coverage change detection method based on images of unmanned aerial vehicles, comprising the following steps:

[0007] An unmanned aerial vehicle is used to collect real-time remote sensing images of the grassland area to be detected;

[0008] Using image processing technology, the real-time remote sensing images and the preset stored historical remote sensing images are respectively subjected to multi-feature extraction and fusion, based on the fused features, a data set is obtained, and the data set is divided into a training set and a test set;

[0009] A BP neural network model is constructed;

[0010] The training set is used to train the BP neural network model;

[0011] The test set and the trained BP neural network model are used to realize the detection of the grassland coverage change.

[0012] Preferably, the remote sensing image collecting UAV is equipped with a CCD camera; and the CCD camera photographs the grassland area to be detected.

[0013] Preferably, the multi-feature extraction method for the real-time remote sensing image and the preset stored historical remote sensing image respectively comprises:

[0014] The real-time remote sensing image and the historical remote sensing image are subjected to background modeling respectively to extract foreground images;

[0015] The foreground images are subjected to color normalization processing;

[0016] The foreground images subjected to the color normalization processing are segmented into a plurality of blob sub-regions by using a segmentation algorithm, and the segmented blob sub-regions are fitted into preset regular geometric shapes by using computational geometry methods;

[0017] Color and position features corresponding to each fitted blob sub-region are extracted.

[0018] Preferably, the color normalization processing method for the foreground images comprises: normalizing pixel color values with respect to the sum of R, G and B channel brightness or the maximum value of the three channels.

[0019] Preferably, the multi-feature fusion method comprises:

[0020] In a keras deep learning framework, inverse convolution operation is introduced to fuse feature information at different levels.

[0021] Preferably, the BP neural network model construction method comprises:

[0022] The number of input layer nodes of the BP neural network model is set to be equal to the number of fused features, and the number of output layer nodes of the BP neural network model is set;

[0023] The number of hidden layer nodes of the BP neural network model is determined by using a trial-and-error method according to the number of input layer nodes and the number of output layer nodes;

[0024] The BP neural network model is constructed according to the determined number of input layer nodes, the number of output layer nodes and the number of hidden layer nodes.

[0025] Preferably, the BP neural network model construction according to the determined number of input layer nodes, the number of output layer nodes and the number of hidden layer nodes comprises:

[0026] The BP neural network model is constructed by using a Sig-moid function as an activation function according to the determined number of input layer nodes, the number of output layer nodes and the number of hidden layer nodes.

[0027] Preferably, the BP neural network model training method using the training set comprises:

[0028] 80% of the image samples with fused features are randomly selected as training samples to train the BP neural network model; normalization is used when inputting data to accelerate the convergence speed of the BP neural network model during learning.

[0029] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention utilizes a remote sensing image acquisition drone to collect real-time and historical remote sensing images of the grassland area to be detected; it employs image processing technology to extract and fuse multiple features from the real-time and historical remote sensing images respectively; based on the fused features, a dataset is obtained, and the dataset is divided into a training set and a test set; a BP neural network model is constructed; the BP neural network model is trained using the training set; and the test set and the trained BP neural network model are used to detect grassland cover changes. This invention can solve the problem of outdated methods currently used for grassland vegetation cover monitoring, employing a combination of ground monitoring devices and aerial monitoring devices, resulting in flexible networking, long transmission distances, and high reliability; achieving rapid, efficient, convenient, and accurate integrated ground-air grassland vegetation monitoring. Attached Figure Description

[0030] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the grassland cover change detection method based on UAV images according to the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1

[0035] like Figure 1 As shown, this invention provides a grassland cover change detection method based on UAV images, comprising the following steps:

[0036] The unmanned aerial vehicle for collecting remote sensing images is used to collect real-time remote sensing images of the grassland area to be detected.

[0037] The image processing technology is used to perform multi-feature extraction and fusion on the real-time remote sensing images and the preset stored historical remote sensing images respectively, to obtain a data set based on the fused features, and to divide the data set into a training set and a test set.

[0038] A BP neural network model is constructed.

[0039] The BP neural network model is trained by using the training set.

[0040] The test set and the trained BP neural network model are used to realize detection of the grassland coverage change.

[0041] In the embodiment, the unmanned aerial vehicle for collecting remote sensing images is equipped with a CCD camera; the CCD camera takes pictures of the grassland area to be detected.

[0042] The unmanned aerial vehicle for collecting remote sensing images can adopt a common four-rotor or six-rotor unmanned aerial vehicle on the market, which has the advantages of stable flight, easy operation, flexible action, and can be used for hovering. For example, the unmanned aerial vehicle can use a 11.2V, 3300mAh lithium battery as a power source.

[0043] The unmanned aerial vehicle for collecting remote sensing images is also equipped with a processing module, a communication module and a storage module.

[0044] When the unmanned aerial vehicle works, it flies to the sky according to the grassland area to be detected, and collects remote sensing images according to the pre-set path. In addition, the processing module can assign geographical parameters to the remote sensing images according to the built-in GPS system and inertial navigation system. The processing module is connected with the communication module and the storage module; the storage module is used to store remote sensing images with geographical parameters, and the processing module sends the remote sensing images to the radio base station through the communication module according to the received instructions (from the ground control center or the built-in preset instructions).

[0045] In the embodiment, the method for performing multi-feature extraction on the real-time remote sensing images and the preset stored historical remote sensing images respectively is as follows:

[0046] The background modeling is performed on the real-time remote sensing images and the historical remote sensing images respectively, so as to extract a foreground image; wherein the background model is a background model based on MRF.

[0047] The foreground image is subjected to color normalization processing, which specifically includes: the pixel color value is normalized to the sum of R, G and B three channel brightness or to the maximum value of the three channels.

[0048] The normalized foreground image is segmented into multiple blob sub-regions by using a segmentation algorithm, specifically including: clustering each pixel point of the image by using an image segmentation algorithm to segment the image into multiple sub-regions, specifically including: establishing an undirected weighted graph G(V, E) according to the image to be segmented, wherein the pixel points are the vertices of the graph, the edges connect adjacent pixel points, and the weight of the edge is the dissimilarity measure of the two pixel points, and iteratively merging the clusters adaptively.

[0049] The segmented blob sub-regions are fitted into a preset regular geometric shape by using a computational geometry method, wherein each blob block corresponds to RGB color and spatial coordinate information; the image segmentation algorithm is based on the GrabCut algorithm;

[0050] The color and position features corresponding to each blob sub-region are extracted.

[0051] In this embodiment, the method for fusing multiple features is:

[0052] In the keras deep learning framework, an inverse convolution operation is introduced to fuse different levels of feature information.

[0053] In this embodiment, the feature map information fusion formula of different resolutions is:

[0054] S ij =X i +δ c {T(Y j )} i≠j,i,j∈N

[0055] Wherein, S ij represents the fused feature map, X i represents the high-resolution feature map, Y j represents the low-resolution feature map or the feature map fused last time, T is an inverse convolution transform, δ c is a channel transform, and N represents a feature map set.

[0056] For example, the 38x38x512 feature map of the extraction network is fused with the 19x19x1024 feature map, first the 19x19x1024 feature map is up-sampled by using inverse convolution to increase the resolution to 38x38x1024 feature map, then a 1x1 convolution is used to convert the channel number of the 38x38x1024 feature map to the same dimension as the 38x38x512 feature map; Finally, the two feature maps are added in the corresponding channels to realize feature fusion. The fused feature map can be repeatedly operated as described above to continue to be fused with the high-resolution feature map.

[0057] In this embodiment, the method for constructing a BP neural network model is:

[0058] The input layer node number of the BP neural network model is set as the number of fusion features, and the output layer node of the BP neural network model is set;

[0059] According to the input layer node number and the output layer node data, the number of hidden layer nodes of the BP neural network model is determined by using a trial and error method;

[0060] According to the determined input layer node number, output layer node number and hidden layer node number, a BP neural network model is constructed.

[0061] In the embodiment, the BP neural network model is constructed according to the determined input layer node number, output layer node number and hidden layer node number.

[0062] According to the determined input layer node number, output layer node number and hidden layer node number, a BP neural network model is constructed.

[0063] Specifically, the design of the BP neural network of the present application is as follows:

[0064] 1. The number of input nodes is determined. There are 15 color feature parameters and 16 position features, and a total of 31 features. Preferably, the number of input nodes is 31.

[0065] 2. The number of hidden layer nodes is determined. The least square fitting value is used as the number of hidden layer nodes, and the number of hidden layer nodes is adjusted on the basis of experiments to determine the number of hidden layer nodes. The calculation formula is as follows:

[0066]

[0067] In the formula, the result is rounded, wherein N represents the number of hidden layer nodes, I represents the number of input layer nodes, and 0 represents the number of output layer nodes.

[0068] Through calculation, preferably, the number of hidden layer nodes is 3.

[0069] 3. The number of output nodes is determined. There are 6 kinds of conventional weeds in grassland, which are Alternanthera philoxeroides, Lamiophlomis rotata, Eclipta prostrata, Sagittaria trifolia, Echinochloa crus-galli and Euphorbia lathyris. Preferably, the number of output nodes is 6.

[0070] 4. The learning rate is selected. The learning rate determines the weight change of each cycle training. Preferably, the initial learning rate is set to 0.001, and an adaptive learning rate method is used during training to improve the training speed.

[0071] 5. The expected error is selected. The expected error value is determined according to the number of hidden layer nodes. Preferably, the target error value is selected to be 0.005.

[0072] In the embodiment, the method for training the BP neural network model by using the training set is as follows:

[0073] 80% of the image samples of the fusion features are randomly extracted as training samples, and the BP neural network model is trained; when the data is input, normalization processing is used to accelerate the convergence speed of the BP neural network model during learning.

[0074] The remaining 20% of the samples are used for neural network simulation test in the embodiment. After the feature parameters are extracted, the trained BP neural network is input, and the average test accuracy and the average simulation test time are determined. If the recognition rate does not meet the requirement, the training samples are increased; if the recognition speed does not meet the requirement, the BP neural network is optimized; if both meet the requirement, the BP neural network model test is completed.

[0075] The above-described embodiments only describe the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for detecting grassland cover change based on images of unmanned aerial vehicles, characterized in that, Includes the following steps: A drone was used to acquire real-time remote sensing images of the grassland area being monitored. Using image processing technology, multiple features are extracted and fused from the real-time remote sensing images and the pre-stored historical remote sensing images. Based on the fused features, a dataset is obtained, and the dataset is divided into a training set and a test set. Construct a BP neural network model; The BP neural network model is trained using the training set; Using the test set and the trained BP neural network model, grassland cover change can be detected. The method for performing multi-feature extraction on the real-time remote sensing images and the preset stored historical remote sensing images is as follows: Background modeling is performed on the real-time remote sensing image and the historical remote sensing image respectively, and foreground image is extracted; Perform color normalization on the foreground image; The normalized foreground image is segmented into multiple b1ob sub-regions using a segmentation algorithm, and the segmented b1ob sub-regions are fitted into geometric shapes according to a preset rule using computational geometry. Extract the color and position features corresponding to each fitted blob sub-region; The method for fusing multiple features is as follows: On the Keras deep learning framework, inverse convolution is introduced to fuse feature information from different levels.

2. The UAV image-based rangeland cover change detection method of claim 1, wherein, The remote sensing image acquisition drone is equipped with a CCD camera; the CCD camera takes pictures of the grassland area being detected. 3.The UAV image-based grassland cover change detection method of claim 1, wherein, The method for color normalization of the foreground image is to normalize the pixel color value to the sum of the brightness of the R, G, and B channels or to the maximum value of the three channels.

4. The UAV image based rangeland cover change detection method of claim 1, wherein, The method for constructing a BP neural network model is as follows: The number of input layer nodes of the BP neural network model is set to the number of fused features, and the number of output layer nodes of the BP neural network model is also set. The number of hidden layer nodes in the BP neural network model is determined by trial and error based on the number of input layer nodes and the output layer node data. Based on the determined number of input layer nodes, output layer nodes, and hidden layer nodes, a BP neural network model is constructed.

5. The grassland cover change detection method based on UAV imagery according to claim 1, characterized in that, The construction of the BP neural network model based on the determined number of input layer nodes, output layer nodes, and hidden layer nodes includes: Based on the determined number of input layer nodes, output layer nodes, and hidden layer nodes, a BP neural network model is constructed using the Sig-moid function as the activation function.

6. The grassland cover change detection method based on UAV imagery according to claim 1, characterized in that, The method for training the BP neural network model using the training set is as follows: 80% of the image samples with fused features are randomly selected as training samples to train the BP neural network model; normalization is used when inputting data to accelerate the convergence speed of the BP neural network model during learning.

Images