Method and device for calculating land parcel area in land parcel image and unmanned aerial vehicle
By combining deep learning neural networks and inertial navigation data, the accuracy and efficiency issues of plot area calculation from the perspective of drones are solved, and efficient and accurate plot area calculation in complex scenarios is achieved.
Patent Information
- Application Number
- CN202510754060.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-05
- Publication Date
- 2025-09-16
AI Technical Summary
Traditional methods have problems of low accuracy and low efficiency when calculating the area of land from the perspective of drones. Especially in complex land scenes and with many noisy areas, it is difficult to achieve efficient and accurate segmentation and calculation.
A deep learning-based neural network is used for land parcel segmentation. The U-Net structure and the adaptive spatial pyramid pooling convolution layer are combined, and the model is trained using a hybrid loss function. Land parcel images are acquired through image acquisition equipment and noise areas are filtered. Inertial navigation data is used to construct a ground resolution map and calculate the land area.
It achieves accurate segmentation of plot boundaries in complex plot scenarios, improves the accuracy and efficiency of plot area calculation, is suitable for resource-constrained embedded devices, and has high robustness and real-time performance.
Smart Images
Figure CN120655696A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of vegetation protection in the agricultural field, and in particular to a method, device and drone for calculating the area of a plot in a plot image. Background Art
[0002] Drones (UAVs) are widely used in agriculture, forestry, land management, and other fields. They are particularly useful for monitoring agricultural plots, where they can quickly acquire high-resolution images of land. However, efficiently and accurately calculating plot areas from drone-perspective images is a challenging problem. This challenge manifests itself in many ways. For example, traditional image segmentation methods (such as those based on thresholding and edge detection) often struggle to handle complex plot scenarios, such as those with blurred boundaries between plots and complex surface textures. Furthermore, after segmentation, the image often contains a high level of noise, which can affect computational accuracy. Summary of the Invention
[0003] In view of the above analysis, the embodiments of the present invention aim to provide a method, device and drone for calculating the area of a plot in a plot image, so as to solve the problems of low accuracy and low efficiency of traditional plot area calculation methods in the field of agricultural vegetation protection.
[0004] In a first aspect, an embodiment of the present invention provides a method for calculating the area of a plot in a plot image, comprising the following steps:
[0005] Using an image acquisition device to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its location information and a second target point and its location information;
[0006] Using the plot image as an input image of a trained neural network-based image segmentation method, the output image of the trained neural network-based image segmentation method being a plot segmentation map;
[0007] Determining a ground resolution of the land segmentation map based on the first target point and its position information and the second target point and its position information in the land segmentation map; and
[0008] The area of the land parcel in the land parcel segmentation map is determined according to the ground resolution.
[0009] Based on a further improvement of the above calculation method, determining the ground resolution of the land segmentation map according to the first target point and its position information and the second target point and its position information in the land segmentation map includes:
[0010] constructing a mapping from pixel coordinates in the land segmentation map to pixel position information based on the inertial navigation data of the image acquisition device, the pixel coordinates of the first target point and the second target point in the land segmentation map, and the position information of the first target point and the second target point; and
[0011] Based on the mapping, a ground resolution of the land parcel segmentation map is determined.
[0012] Based on a further improvement of the above calculation method, the first target point is the take-off point of the drone, and the second target point is the center point of the image.
[0013] Based on a further improvement of the above calculation method, before determining the ground resolution of the land segmentation map, the calculation method further includes:
[0014] determining a noise area in the land segmentation map according to the position information of the second target point; and
[0015] The noise area is removed.
[0016] Based on a further improvement of the above calculation method, determining the noise area in the land segmentation map according to the position information of the second target point includes:
[0017] Determine in the land parcel segmentation map that a land parcel with an area smaller than 1000 / n pixels is a noise area, where n is the altitude of the second target point.
[0018] Based on a further improvement of the above-mentioned calculation method, a neural network-based image segmentation method includes encoding and decoding, wherein a U-Net is included between the encoding and decoding, and the U-Net includes an adaptive spatial pyramid pooling convolution layer. The adaptive spatial pyramid pooling convolution layer includes a spatial pyramid pooling layer and a dilated convolution layer. The encoded image is input to the spatial pyramid pooling layer, and the image output by the dilated convolution layer is decoded.
[0019] Based on a further improvement of the above calculation method, the encoding module includes a first encoder, a second encoder, a third encoder, a fourth encoder and a fifth encoder, wherein the first encoder is used to extract edge information of the plot image, the second encoder is used for texture information of the plot image, the third encoder is used to extract semantic information of the plot image, and the fourth encoder and the fifth encoder are used to extract multi-scale features of the plot image.
[0020] Based on a further improvement of the above calculation method, the first encoder, the second encoder, the third encoder, the fourth encoder, and the fifth encoder are respectively the first to fifth convolutional layers of the ResNet 34 network.
[0021] Based on the further improvement of the above calculation method, the training process of the segmentation neural network model includes:
[0022] Constructing a land parcel image set, wherein the land parcel images in the land parcel image set include annotations representing land parcel segmentation information;
[0023] The segmentation neural network model is trained on the land image set based on a hybrid loss function; wherein,
[0024] The hybrid loss function is a weighted combination of the cross entropy loss function and the Dice coefficient loss function.
[0025] In a second aspect, an embodiment of the present invention provides a device for calculating the area of a land parcel in a land parcel image, comprising:
[0026] An image acquisition device is used to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its position information and a second target point and its position information;
[0027] An inference module is used to input the land parcel image into a pre-trained segmentation neural network model to obtain a land parcel segmentation map;
[0028] a first determining module, configured to determine a ground resolution of the land segmentation map based on the first target point and its position information and the second target point and its position information in the land segmentation map; and
[0029] The second determining module is configured to determine the area of the land parcel in the land parcel segmentation map according to the ground resolution.
[0030] In a third aspect, an embodiment of the present invention provides a drone, including:
[0031] An image acquisition device is used to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its position information and a second target point and its position information;
[0032] An embedded device communicatively connected to the image acquisition device; wherein the embedded device is used to execute the method for calculating the area of a plot in a plot image according to any one of the first aspects of the present invention.
[0033] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0034] 1. The solution of the present invention adopts a deep learning-based neural network for plot segmentation, which can adapt to complex plot scenarios, accurately segment the plot boundaries, and accurately calculate the actual area of each plot based on the resolution.
[0035] 2. The solution of the present invention can effectively filter the noise areas in the image, thereby further improving the accuracy of calculating the land area.
[0036] 3. The deep learning-based neural network model in the solution of the present invention has a streamlined structure and can achieve efficient computing even when running on resource-constrained embedded devices.
[0037] 4. The deep learning-based neural network model in the solution of the present invention adopts a hybrid loss function to complete the model training. This hybrid loss function can significantly improve the ability to capture boundary details while ensuring the overall segmentation accuracy, providing higher robustness for the plot area calculation task.
[0038] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.
[0040] Figure 1 The figure shows a flow chart of a method for calculating the area of a land parcel in a land parcel image according to an embodiment of the present invention.
[0041] Figure 2 A schematic diagram of the structure of U-Net in an embodiment of the present invention is shown.
[0042] Figure 3 A structural block diagram of a segmentation neural network model according to an embodiment of the present invention is shown.
[0043] Figure 4 A structural block diagram of a segmentation neural network model according to another embodiment of the present invention is shown.
[0044] Figure 5 Figure 3 shows the overall structure of the ResNet 34 network.
[0045] Figure 6 An exemplary block diagram of a device for calculating the area of a land parcel in a land parcel image according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0046] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.
[0047] Figure 1 FIG. 1 is a flow chart showing a method for calculating the area of a land parcel in a land parcel image according to an embodiment of the present invention. Figure 1 As shown, the method for calculating the area of the plot in the plot image includes the following steps:
[0048] Step S100: using an image acquisition device to acquire a plot image from the perspective of a drone.
[0049] In this embodiment, a camera carried by a drone can be used to take a bird's-eye view of the land parcel to obtain a high-definition image of the land parcel.
[0050] In this embodiment, the land parcel image includes a first target point. Exemplarily, the first target point is the take-off point of the drone. In embodiments where the first target point is the take-off point of the drone, the drone needs to capture the take-off point when photographing the land parcel. In this case, the first target point is the pixel corresponding to the take-off point in the land parcel image.
[0051] In this embodiment, the land parcel image includes a second target point. For example, the second target point may be the location of the drone when the image was taken. In the embodiment where the second target point is the drone's shooting point, the second target point is the center pixel of the land parcel image.
[0052] In this embodiment, the location information of the first target point and the second target point can be obtained by the navigation system onboard the drone. The location information here can be coordinates consisting of longitude, latitude, and altitude. After the location information is obtained, it can be annotated at the corresponding pixel point in the land parcel image.
[0053] It should be noted that, although the above embodiment gives an example in which the first target point is a take-off point or the second target point is a shooting point, the present invention is not limited thereto. The first target point and the second target point can be any two points in space, as long as the first target point and the second target point have corresponding pixel content in the plot image.
[0054] Step S200: using the plot image as an input image of the trained neural network-based image segmentation method, and the output image of the trained neural network-based image segmentation method is a plot segmentation map.
[0055] In this embodiment, the plot segmentation map is used to distinguish the plot from the background in the plot image. In this embodiment, a segmentation neural network model can be used to generate the plot segmentation map. The segmentation neural network model is a neural network-based image segmentation method. In this embodiment, the segmentation neural network model includes an encoding module and a decoding module, which are interconnected using a U-Net. Figure 2 The schematic diagram of the structure of the encoding module and the decoding module using U-Net interconnection is shown. Figure 2 As shown in Figure 1, the encoding module includes five encoders, labeled Encoder 1, Encoder 2, Encoder 3, Encoder 4, and Encoder 5, respectively. The decoding module includes five decoders, labeled Decoder 1, Decoder 2, Decoder 3, Decoder 4, and Decoder 5, respectively. The decoders in the decoding module not only gradually upsample image features but also fuse image features from the encoders using skip connections. For example, Decoder 2 not only upsamples the image features output by Decoder 3 but also fuses image features from Encoder 2.
[0056] In this embodiment, an adaptive spatial pyramid pooling convolution layer is further included between the encoding module and the decoding module. Figure 3 FIG. 4 shows a structural block diagram of a segmentation neural network model according to an embodiment of the present invention. Figure 3 As shown, the segmentation neural network model includes an encoding module, a decoding module, and an adaptive spatial pyramid pooling convolution layer. The adaptive spatial pyramid pooling convolution layer includes a spatial pyramid pooling layer and a dilated convolution layer. The output of the encoding module is the input of the spatial pyramid pooling layer, and the output of the dilated convolution layer is the input of the decoding module.
[0057] The Spatial Pyramid Pooling Layer (SPP) can divide the input feature map into multiple grids of different sizes to form a pyramid structure. The size of each grid determines the receptive field of the pooling operation, and pooling operations (such as maximum pooling) can be performed on each grid.
[0058] Dilated Convolution (DC) is a variant of the convolution operation in convolutional neural networks, which expands the receptive field by introducing an interval (dilation rate) within the convolution kernel without adding additional parameters or computation.
[0059] Figure 4 FIG. 4 shows a structural block diagram of a segmentation neural network model according to another embodiment of the present invention. Figure 4As shown, the segmentation neural network model includes an encoding module, an adaptive spatial pyramid pooling convolutional layer, a decoding module, and a feature fusion layer. The encoding module includes five convolutional layers, each of which is an encoder. The first convolutional layer is used to extract edge information of the image, the second convolutional layer is used to extract texture information of the image, the third convolutional layer is used to extract semantic information of the image, and the fourth and fifth convolutional layers are used to extract multi-scale features of the image. The decoding module includes a central decoding block and five decoding blocks. The first to fifth decoding blocks and the first to fifth convolutional layers are interconnected using U-Net. For example, the fourth decoding block upsamples the feature map output by the fifth decoding block while also fusing the feature map of the fourth convolutional layer. The central decoding block is used to perform preliminary decoding on the feature map output by the adaptive spatial pyramid pooling convolutional layer. The feature fusion layer is a fully connected layer used to integrate multi-scale image features.
[0060] In some embodiments, Figure 4 The first to fifth convolutional layers in are the first to fifth convolutional layers of the ResNet 34 network. Figure 5 The overall structure of the ResNet 34 network is shown in FIG. Figure 4 The first to fifth convolutional layers in Figure 5 The five convolutional layers are outlined by the dashed boxes. Since the ResNet 34 network belongs to the existing technology, it will not be described here.
[0061] It should be noted that the encoding module in this embodiment uses the design of the first five layers of the ResNet 34 network, which is beneficial. For resource-constrained embedded devices, if the neural network model running on the device is too large, its inference speed is slow and it is difficult to meet the real-time requirements. By streamlining the inference, efficient inference can be achieved and the real-time requirements of low-cost embedded devices can be met.
[0062] In some embodiments, the training process of the segmentation neural network model includes: constructing a set of plot images, wherein the plot images in the set contain annotations representing plot segmentation information; and training the segmentation neural network model on the plot image set based on a hybrid loss function, wherein the hybrid loss function is a weighted combination of a cross entropy loss function and a Dice coefficient loss function. The annotations of the plot images can be binary masks. A binary mask is a data representation method that uses two colors (usually black and white) to represent different areas in an image, where one color is used to mark the target area and the other color is used for the background area.
[0063] The cross entropy loss function is a loss function used in deep learning to measure the difference between the probability distribution of the model output and the probability distribution of the true label. The cross entropy loss function is defined as follows:
[0064]
[0065] in:
[0066] N represents the number of categories, that is, the total number of categories that the model needs to predict.
[0067] ·y i Indicates the true label, usually represented by one-hot encoding.
[0068] In encoding, for a given sample, there is only one category y i The value is 1, and the y of all other categories i The values are all 0.
[0069] · represents the probability distribution of the i-th category predicted by the model.
[0070] The Dice coefficient is a statistical method commonly used to measure the degree of overlap between two sample sets. In deep learning, the Dice coefficient can be used as a loss function to train models. The definition of the Dice coefficient loss function is as follows:
[0071]
[0072] in:
[0073] ·p i is the i-th pixel value of the predicted segmentation map.
[0074] ·g i is the i-th pixel value of the true segmentation map.
[0075] In response to the common class imbalance problems and boundary fuzzy problems in land segmentation tasks, this embodiment designs a hybrid loss function. Specifically, cross entropy loss is used as the basic loss to measure the accuracy of pixel-level classification. At the same time, Dice coefficient loss is introduced to optimize the segmentation effect of land boundaries, especially in scenes with small-area land and complex boundaries. Finally, the cross entropy loss and Dice coefficient loss are weightedly combined with a weight ratio of 1:1 or 1:2 to form a hybrid loss function. This loss function can significantly improve the ability to capture boundary details while ensuring the overall segmentation accuracy, providing higher robustness for land segmentation tasks.
[0076] In some embodiments, the trained segmentation neural network model can be exported to ONNX format, then converted to RKNN format using RKNN-Toolkit, adapted to the Rockchip RK3588 chip, and finally deployed on the drone. In this embodiment, the ONNX to RKNN model conversion fully utilizes the hardware acceleration capabilities of the Rockchip RK3588 chip to achieve efficient inference and meet the real-time requirements of low-cost embedded devices.
[0077] In this embodiment, before inputting the land parcel image into the segmentation neural network model, the land parcel image may be normalized. Normalization refers to mapping the pixel values of the land parcel image to the range required by the segmentation neural network model. For example, normalization may involve mapping the pixel value range of the land parcel image from 0-255 to 0-1.
[0078] Step S300: determining the ground resolution of the land parcel segmentation map according to the first target point and its position information and the second target point and its position information in the land parcel segmentation map.
[0079] Ground resolution refers to the actual distance on the ground represented by each pixel in an image, typically measured in meters. It measures the image's ability to distinguish adjacent features. For example, a ground resolution of 1 meter means that one pixel corresponds to a 1×1 meter area on the ground. In this embodiment, the ground resolution of the land segmentation map can be generated using a ground resolution generation algorithm. The ground resolution generation algorithm uses the pixel coordinates and position information of the first target point and the pixel coordinates and position information of the second target point in the land segmentation map as input to generate the ground resolution of the land segmentation map.
[0080] In some embodiments, the ground resolution generation algorithm includes the following steps:
[0081] Step S310: Construct a mapping from pixel coordinates in the land segmentation map to pixel position information based on the inertial navigation data of the image acquisition device, the pixel coordinates of the first target point and the second target point in the land segmentation map, and the position information of the first target point and the second target point.
[0082] Step S320: Determine the ground resolution of the land segmentation map according to the mapping.
[0083] Steps S310 to S320 are described below with reference to a specific embodiment.
[0084] If the first target point is the take-off point of the drone and the second target point is the shooting point of the drone (i.e. the center pixel of the image), then the ground resolution can be calculated according to the following steps:
[0085] 1. Calculate the geographic offset between the take-off point and the shooting point
[0086] The position information of the take-off point and the shooting point is known, and the geographical offset of the shooting point relative to the take-off point can be calculated by the difference between the position information of the take-off point and the shooting point.
[0087] It should be noted that during the flight, inertial navigation data is also used to adjust the attitude of the image acquisition device.
[0088] Specifically, inertial navigation data provides the pitch, yaw, and roll angles of the drone's image acquisition device during capture. These angles describe the tilt of the drone's image acquisition device and affect the geographic mapping of the land parcel segmentation map. Using these angles, the image acquisition device's orientation can be corrected to ensure that images are captured parallel to the ground.
[0089] 2. Constructing the mapping matrix
[0090] The mapping matrix is a mathematical model used to convert the pixel coordinates in the land segmentation map into the actual coordinates on the ground. The mapping matrix can be obtained by the following formula:
[0091]
[0092] in, is the ground coordinate to be calculated; R imu is a rotation matrix generated by the pitch, yaw, and roll angles of the image acquisition device; is the pixel coordinate; is the geographical offset of the shooting point relative to the take-off point.
[0093] Based on the pixel coordinates of the first target point and the second target point respectively, the ground coordinates of the first target point and the second target point are obtained by using the above mapping matrix.
[0094] 3. Calculate ground resolution
[0095] First, calculate the number of pixels m in the horizontal direction between the takeoff point and the shooting point in the segmented plot. Then, calculate the actual horizontal distance S between the takeoff point and the target point based on the ground coordinates of the first and second target points. Finally, calculate the ground resolution GSD using the following formula:
[0096]
[0097] Step S400: Determine the area of the land parcel in the land parcel segmentation map according to the ground resolution.
[0098] In this embodiment, the number of pixels within each parcel outline in the parcel segmentation map can be calculated, and then the number of pixels within each parcel outline can be converted into the actual parcel area based on the ground resolution. Specifically, the actual parcel area within each parcel outline is equal to the number of pixels within the parcel outline multiplied by the ground resolution.
[0099] Back to Figure 1 .
[0100] In some embodiments, upon entering Figure 1 Before step S300, Figure 1 The method also includes:
[0101] Step S210: determining a noise area in the land segmentation map according to the position information of the second target point.
[0102] Step S22: removing the noise area.
[0103] In this embodiment, for a noise region with a smaller area in the land segmentation map, parameter setting can be performed based on the position information of the second target point, for example, based on the altitude value of the second target point.
[0104] In some embodiments, parameterizing the second target point according to its altitude includes proportionally setting the second target point according to its altitude. For example:
[0105] At a height of 10 meters, all connected regions with an area less than 100 pixels are filtered out;
[0106] At an altitude of 100 meters, all connected regions with an area smaller than 10 pixels are filtered out;
[0107] At an altitude of 200 meters, all connected regions with an area smaller than 5 pixels are filtered out;
[0108] …
[0109] At a height of n meters, all connected regions with an area smaller than 10*100 / n pixels are filtered out.
[0110] Figure 6 FIG. 1 shows an exemplary block diagram of a device for calculating the area of a land parcel in a land parcel image according to an embodiment of the present invention. Figure 6As shown, the device 600 for calculating the area of a plot in the plot image includes: an image acquisition device 601, used to acquire a plot image from the perspective of a drone, wherein the plot image includes a first target point and its position information and a second target point and its position information; an inference module 602, used to use the plot image as an input image of a trained neural network-based image segmentation method, and the output image of the trained neural network-based image segmentation method is a plot segmentation map; a first determination module 603, used to determine the ground resolution of the plot segmentation map based on the first target point and its position information and the second target point and its position information in the plot segmentation map; and a second determination module 604, used to determine the area of the plot in the plot segmentation map based on the ground resolution.
[0111] It should be understood that Figure 6 The modules of the apparatus 600 shown in FIG. 6 can be used in conjunction with the reference Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features and advantages described above for the method are also applicable to the device 600 and the modules included therein. For the sake of brevity, some operations, features and advantages are not repeated here.
[0112] An embodiment of the present invention provides an unmanned aerial vehicle (UAV), comprising: an image acquisition device for acquiring a plot image from the perspective of the UAV, wherein the plot image includes a first target point and its position information and a second target point and its position information; an embedded device communicatively connected to the image acquisition device; wherein the embedded device is used to execute the method for calculating the plot area in the plot image according to any one of the first aspects of the present invention.
[0113] Compared with the prior art, the embodiments of the present invention can achieve at least one of the following beneficial effects:
[0114] 1. The solution of the present invention adopts a deep learning-based neural network for plot segmentation, which can adapt to complex plot scenarios, accurately segment the plot boundaries, and accurately calculate the actual area of each plot.
[0115] 2. The solution of the present invention can effectively filter the noise areas in the image, thereby further improving the accuracy of calculating the land area.
[0116] 3. The deep learning-based neural network model in the solution of the present invention has a streamlined structure and can achieve efficient computing even when running on resource-constrained embedded devices.
[0117] 4. The deep learning-based neural network model in the solution of the present invention adopts a hybrid loss function to complete the model training. This hybrid loss function can significantly improve the ability to capture boundary details while ensuring the overall segmentation accuracy, providing higher robustness for the plot area calculation task.
[0118] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily conceived by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.
Claims
1. A method for calculating the area of a plot in a plot image, characterized in that: The steps include: Using an image acquisition device to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its location information and a second target point and its location information; Using the plot image as an input image of a trained neural network-based image segmentation method, the output image of the trained neural network-based image segmentation method being a plot segmentation map; Determining a ground resolution of the land segmentation map based on the first target point and its position information and the second target point and its position information in the land segmentation map; and The area of the land parcel in the land parcel segmentation map is determined according to the ground resolution.
2. The calculation method according to claim 1, characterized in that Determining the ground resolution of the land segmentation map according to the first target point and its position information and the second target point and its position information in the land segmentation map includes: constructing a mapping from pixel coordinates in the land segmentation map to pixel position information based on the inertial navigation data of the image acquisition device, the pixel coordinates of the first target point and the second target point in the land segmentation map, and the position information of the first target point and the second target point; and Based on the mapping, a ground resolution of the land parcel segmentation map is determined.
3. The calculation method according to claim 1 or 2, characterized in that: The first target point is the take-off point of the drone, and the second target point is the center point of the image.
4. The calculation method according to claim 3, characterized in that Before determining the ground resolution of the land segmentation map, the calculation method further includes: determining a noise area in the land segmentation map according to the position information of the second target point; and The noise area is removed.
5. The calculation method according to claim 2, characterized in that Determining the noise area in the land segmentation map according to the position information of the second target point includes: Determine in the land parcel segmentation map that a land parcel with an area smaller than 1000 / n pixels is a noise area, where n is the altitude of the second target point.
6. The calculation method according to claim 1, characterized in that The neural network-based image segmentation method includes encoding and decoding, wherein a U-Net is included between the encoding and decoding, and the U-Net includes an adaptive spatial pyramid pooling convolution layer. The adaptive spatial pyramid pooling convolution layer includes a spatial pyramid pooling layer and a dilated convolution layer. The encoded image is input to the spatial pyramid pooling layer, and the image output by the dilated convolution layer is decoded.
7. The method according to claim 6, characterized in that The encoding is performed by an encoding module, which includes a first encoder, a second encoder, a third encoder, a fourth encoder, and a fifth encoder. The first encoder is used to extract edge information of the land parcel image, the second encoder is used to extract texture information of the land parcel image, the third encoder is used to extract semantic information of the land parcel image, and the fourth encoder and the fifth encoder are used to extract multi-scale features of the land parcel image.
8. The method according to claim 7, characterized in that The first encoder, the second encoder, the third encoder, the fourth encoder, and the fifth encoder are respectively the first to fifth convolutional layers of the ResNet 34 network.
9. The method according to claim 6, characterized in that The training process of the segmentation neural network model includes: Constructing a land parcel image set, wherein the land parcel images in the land parcel image set include annotations representing land parcel segmentation information; The segmentation neural network model is trained on the land image set based on a hybrid loss function; wherein, The hybrid loss function is a weighted combination of the cross entropy loss function and the Dice coefficient loss function.
10. A device for calculating the area of a land parcel in a land parcel image, characterized in that: These include: An image acquisition device is used to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its position information and a second target point and its position information; an inference module, configured to use the land parcel image as an input image for a trained neural network-based image segmentation method, wherein the output image of the trained neural network-based image segmentation method is a land parcel segmentation map; a first determining module, configured to determine a ground resolution of the land segmentation map based on the first target point and its position information and the second target point and its position information in the land segmentation map; and The second determining module is configured to determine the area of the land parcel in the land parcel segmentation map according to the ground resolution.
11. A drone comprising: An image acquisition device is used to acquire a land image from the perspective of the drone, wherein the land image includes a first target point and its position information and a second target point and its position information; An embedded device communicatively connected to the image acquisition device; wherein the embedded device is used to execute the method for calculating the area of a plot in a plot image according to any one of claims 1-9.