Lodging area segmentation method based on Improved_Unet network

Abstract

The present invention particularly relates to a lodging area segmentation method based on an Improved_Unet network, comprising the following steps: using an unmanned aerial vehicle (UAV) to acquire an image of a wheat field to be inspected at a specified height and growth period; geometrically correcting and splicing the image of the wheat field to be inspected to obtain a large RGB image to be inspected; and inputting the large RGB image to be inspected into a trained Improved_Unet network to obtain a lodging area segmentation map, wherein the Improved_Unet network uses the Involution operator to replace the convolution operation at the backbone position of the Unet network. The Improved_Unet network combines the advantages of the Segnet and Unet networks, using the Involution operator to replace the convolution operation at the backbone position, achieving spatial specificity and applying different operations to different pixels, ensuring maximum information extraction and significantly improving segmentation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of lodging detection, and in particular to a lodging area segmentation method based on an Improved_Unet network. Background Art

[0002] Wheat is the world's second-largest food crop. Wind and rain, pests and diseases, high-density planting, or high nitrogen fertilizer application often reduce plant photosynthesis, leading to lodging. This makes harvesting difficult and reduces yield and quality. Lodging is one of the most important factors affecting wheat yield, reducing yield by up to 45%, posing a serious threat to global food security. China is one of the most important wheat-producing countries, with an annual wheat planting area of ​​approximately 24,000 hectares. Lodging causes yield reductions of over 10% annually. Anhui Province is China's third-largest wheat production base. Ensuring the healthy production of wheat in Anhui Province is not only beneficial to food safety and economic development in the five provinces of East my country, but also plays an important role in ensuring national food security. Currently, there is an urgent need for new technologies and methods for rapid and non-destructive assessment of wheat lodging.

[0003] Traditional lodging monitoring methods require on-site agricultural technicians to conduct sampling and analysis of lodging areas, which is time-consuming, labor-intensive, and inefficient. When large-scale lodging occurs, traditional detection methods are time-consuming and inaccurate, often leading to insurance claims disputes and undermining the production enthusiasm of farmers and businesses. Remote sensing, a remote information acquisition method, can accurately capture the spatiotemporal changes of crops, providing powerful technical support for non-destructive and rapid lodging monitoring. Common remote sensing methods include satellite remote sensing, aerial remote sensing, and near-ground remote sensing. Satellite remote sensing, due to its large-scale nature, is increasingly being used for crop lodging detection. However, due to limitations in spatiotemporal resolution and meteorological conditions (southern China is prone to cloudiness and rain, resulting in limited available data), satellite remote sensing has low lodging detection accuracy. Furthermore, due to the long satellite return period, it struggles to meet the demand for rapid response. Aerial remote sensing offers the advantage of greater mobility and speed compared to space-based remote sensing, making it suitable for monitoring large-scale lodging caused by conditions such as strong winds and floods. However, it is further constrained by China's airspace policies and flight costs, making scientific research and technological dissemination difficult. Near-ground (or airborne) remote sensing uses small drones equipped with sensors to acquire a variety of high-resolution data, such as high-definition RGB images, multispectral images, hyperspectral images, and thermal imagers. Its low cost, high flexibility, and ability to provide high spatial resolution have made it a hot topic in crop lodging monitoring research. For example, Moon, Liu, and Chu have used drones for lodging monitoring in various crops, including rice, wheat, and corn. However, due to the large amount of image data captured by drones, data processing methods have become a bottleneck for their further application. Therefore, new methods for processing drone-based lodging monitoring data are urgently needed.

[0004] Many researchers have been researching new methods for drone-based crop lodging monitoring. Ding et al. used YCbCr transformation and support vector machines to extract wheat lodging area, achieving an accuracy of 92%. Liu et al. used visible light and thermal infrared drone imagery as classification data. By combining color, texture, and temperature features, they employed the Optimization of Support Vector Machine (SVM) Multiclass by Particle Swarm (PSO-SVM) method. The correlation coefficient between the obtained lodging percentage and the actual lodging percentage was 0.9415. Chauhan et al. used drone multispectral imagery and employed a multi-resolution segmentation (MRS) algorithm and a nearest neighbor classification algorithm to classify wheat lodging severities. They found that red edge and near-infrared band data could effectively distinguish different classes, achieving an overall accuracy of 90%. Cao et al. proposed a hybrid algorithm based on a watershed algorithm and adaptive threshold segmentation for wheat lodging extraction, which achieved better segmentation results than a single watershed algorithm. However, these methods require technical personnel to screen and identify features, making them difficult to implement and promote in business systems. The emergence of deep learning provides an important opportunity to address these issues. In recent years, with the continuous research of deep learning algorithms, deep learning has achieved significant breakthroughs in classification, detection, and segmentation tasks. Some scholars have conducted preliminary research using deep learning in the field of crop lodging. In the paper "Use of Unmanned Aerial Vehicle Imagery and Deep Learning UNet to Extract Rice Lodging," Zhao et al. used a UNet network to extract areas of rice lodging in late maturity, achieving a Dice coefficient of 0.9442. In the paper "Semantic Segmentation Using Deep Learning with Vegetation Indices for Rice Lodging Identification in Multi-date UAV Visible Images," Yang et al. used an FCN-AlexNet network trained on visible light image data of lodging rice, achieving a lodging extraction accuracy of 0.9443. In the paper "Identifying sunflower lodging based on image fusion and deep semantic segmentation with UAV remote sensing imaging," Song et al. used a custom Improved_SegNet method to extract sunflower lodging areas from fused images, achieving an accuracy 15% to 20% higher than that of traditional support vector machine (SVM) methods.In the paper "Automatic extraction of wheat lodging area based on transfer learning method and deeplabv3+ network," Ding et al. used transfer learning to train the Deeplabv3+ model to segment wheat lodging areas at different growth stages. The results showed that the transfer learning method was more effective than the traditional Unet method. Compared with traditional machine learning methods, deep learning methods have greater potential and space for wheat lodging monitoring research. However, many of the methods mentioned above cannot balance accuracy and efficiency, and lack research at different flight altitudes, which means that the robustness and applicability of the algorithms still need to be verified. Summary of the Invention

[0005] The purpose of the present invention is to provide a lodging area segmentation method based on Improved_Unet network with high extraction accuracy and good robustness.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a lodging area segmentation method based on Improved_Unet network, comprising the following steps: using a drone to obtain an image of a wheat field to be detected at a specified height and a specified growth period; geometrically correcting and splicing the image of the wheat field to be detected to obtain a large RGB image to be detected; inputting the large RGB image to be detected into a trained Improved_Unet network to obtain a lodging area segmentation map, wherein the Improved_Unet network is obtained by using the Involution operator to replace the convolution operation at the backbone position of the Unet network; the Improved_Unet network includes an encoder and a decoder; the encoder includes: the input image is first processed by Conv+BN+Relu to obtain a first feature map; the first feature map is sequentially processed by Dense block and Maxpool to obtain a second feature map; the second feature map is subjected to three Involution+BN+Relu and Dense The third feature map is obtained by block and Maxpool processing; the third feature map is processed by Involution+BN to obtain the fourth feature map; the decoder includes: performing an upsampling operation on the fourth feature map through four consecutive Conv+BN+Relu processes, and using a jump connection in the upsampling process to splice the intermediate feature map generated by the encoder to the feature map generated by the decoder, and obtaining the fifth feature map after upsampling; the fifth feature map is subjected to a 1*1 convolution operation to obtain a segmentation map of the falling area; the Dense block processing includes the following steps: the input feature map is processed by BN+Relu+Involution to obtain a feature map Figure 1; Input feature map and feature Figure 1 After splicing, the features are obtained after BN+Relu+Involution processing Figure 2 ; Input feature map, feature Figure 1 and features Figure 2 The feature map obtained after splicing and BN+Relu+Conv processing is used as the output feature map of Dense block processing; the Conv is the convolution layer, BN is the batch normalization layer, Relu is the activation layer, and Maxpool is the maximum pooling layer.

[0007] Compared with the prior art, the present invention has the following technical effects: using the Improved_Unet network in the present invention, compared with the existing methods, the results on the four evaluation indicators are significantly better than the classic Unet network, Segnet, and DeeplabV3+ network, not only with higher accuracy but also stronger applicability; using the involution operator instead of the convolution operation at the backbone position can obtain a larger receptive field and retain more information, thereby improving the segmentation accuracy; ordinary convolution operations achieve channel specificity, that is, sharing the convolution operator on the same channel and using different operators on different channels. The construction method of the Involution operator is opposite to that of the convolution operation. The Improved_Unet network combines the advantages of the Segnet and Unet networks, achieves spatial specificity, expands the receptive field on the basis of reducing the amount of calculation, and at the same time reduces information redundancy between channels, takes different operations for different pixels, and ensures maximum information extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is the Improved_Unet network structure diagram and its annotation diagram;

[0009] Figure 2 This is the structural diagram of the Dense block;

[0010] Figure 3 It is a schematic diagram of involution;

[0011] Figure 4 It is the loss convergence graph of Unet network and Improved_Unet network;

[0012] Figure 5 It is an arbitrarily cropped image of the study area and its annotation map, as well as the prediction maps of the two network models;

[0013] Figure 6 This is a comparison chart of the processing results of different network models; DETAILED DESCRIPTION

[0014] The following combination Figures 1 to 6 , the present invention is described in further detail.

[0015] See Figure 1 A lodging area segmentation method based on the Improved_Unet network includes the following steps: using a drone to obtain an image of a wheat field to be detected at a specified height and a specified growth period; geometrically correcting and splicing the image of the wheat field to be detected to obtain a large RGB image to be detected; inputting the large RGB image to be detected into a trained Improved_Unet network to obtain a lodging area segmentation map, wherein the Improved_Unet network is obtained by using the Involution operator instead of the convolution operation at the backbone position of the Unet network. Compared with existing methods, the Improved_Unet network in the present invention is not only more accurate but also more applicable; the Improved_Unet network combines the advantages of the Segnet and Unet networks, uses the involution operator instead of the convolution operation at the backbone part, realizes spatial specificity, takes different operations on different pixels, and ensures maximum information extraction. Therefore, the results on the four evaluation indicators are better than those of the other three classic deep learning networks.

[0016] Classic Unet, Segnet, and DeeplabV3+ networks have been used by some researchers for crop lodging extraction, achieving high model accuracy. Segnet, unlike traditional FCN networks, employs a different feature acquisition strategy, using encoding and decoding techniques and pooling indices to preserve pooling position information. The DeepLabV3+ network architecture also consists of an encoding and decoding structure. The encoding layer includes Atrous Spatial Pyramid Pooling (ASPP) and an improved Xception module, which has input, intermediate, and output streams. Previous studies have used transfer learning to reduce the network's requirement for wheat lodging data, but selecting an appropriate public dataset remains challenging. The Unet network employs an encoding and decoding structure, with valid padding applied to convolutions to ensure that contextual features are not missing. In the decoding stage, the feature map obtained in the encoding phase is concatenated with that in the decoding phase, and pixel-level classification is achieved through 1x1 convolution. Compared to the Segnet network, the Unet network features a skip connection structure, which alleviates the vanishing gradient problem during training to some extent. This simplifies the model and makes it more adaptable to the complex information in wheat lodging images, the varying lodging conditions at different growth stages, and the vastly different lodging areas. This also allows for the fusion of low-level and high-level information to capture more information. Compared to DeepLabV3+, the Unet network has significantly fewer layers, requiring fewer training samples and less training time. It also eliminates the need for transfer learning, lowering the standard for use.

[0017] In view of the differences in the above-mentioned network structures and taking into account the real scenes in the field, the present invention is modified based on the lightweight network Unet, which not only ensures the accuracy and reliability of the model, but also reduces the time consumption of model training. In addition, in order to obtain a larger receptive field and retain more information to improve the segmentation accuracy, the present invention uses the involution operator to replace the convolution operation of the backbone position on the basis of ensuring the network structure. Ordinary convolution operations achieve channel specificity, that is, the convolution operator is shared on the same channel, and different operators are used on different channels. The construction method of the Involution operator is opposite to that of the convolution operation, which achieves spatial specificity, thereby expanding the receptive field on the basis of reducing the amount of calculation, while reducing the information redundancy between channels. The schematic diagram of Involution is shown as follows Figure 3 shown.

[0018] H is the involution kernel, calculated by the function Φ. Involution uses grouped convolution, that is, convolution is shared on the channel. For example, if the channel is 32 and 16 channels share the convolution, the grouping is 2. For ease of presentation, the diagram uses a grouping of 1. This construction method achieves spatial specificity and channel sharing.

[0019] Furthermore, the Unet network is very common. The present invention transforms the Unet network to obtain an Improved_Unet network. Specifically, the Improved_Unet network includes an encoder and a decoder; the encoder includes: the input image is first processed by Conv+BN+Relu to obtain a first feature map; the first feature map is sequentially processed by Dense block and Maxpool to obtain a second feature map; the second feature map is processed by Involution+BN+Relu, Denseblock and Maxpool three times to obtain a third feature map; the third feature map is processed by Involution+BN to obtain a fourth feature map; the decoder includes: performing an upsampling operation on the fourth feature map through four consecutive Conv+BN+Relu processes, and using a jump connection to splice the intermediate feature map generated by the encoder onto the feature map generated by the decoder during the upsampling process, and obtaining a fifth feature map after upsampling; the fifth feature map is subjected to a 1*1 convolution operation to obtain a segmentation map of the falling area; the Conv is a convolution layer, BN is a batch normalization layer, Relu is an activation layer, and Maxpool is a maximum pooling layer. During the encoding phase, a dense block module is added to retain more information by combining previous and next feature maps. Max-pooling is used to preserve the maximum pooling position during pooling. Unpooling is then used during decoding to amplify the feature map using the maximum pooling position information recorded during encoding. The final layer of the network uses a 1x1 convolution kernel and a convolution method that limits the number of feature maps to achieve pixel-level classification.

[0020] Further, see Figure 2 The Dense block processing includes the following steps: the input feature map is processed by BN+Relu+Involution to obtain the feature Figure 1 ; Input feature map and feature Figure 1 After splicing, the features are obtained after BN+Relu+Involution processing Figure 2 ; Input feature map, feature Figure 1 and features Figure 2The concatenated feature maps are processed by BN+Relu+Conv and serve as the output feature maps of the Dense block. The idea behind Dense block processing is that for each layer, the feature maps of all previous layers serve as the input of the current layer, while its own feature map serves as the input of subsequent layers, forming a fully interconnected network. The feature maps extracted from each layer can be used by subsequent layers.

[0021] Furthermore, the Improved_Unet network is trained according to the following steps: using a drone to obtain wheat field images at a specified height and a specified growth period; geometrically correcting and splicing the wheat field images to obtain a large RGB image; with the help of expert experience, manually marking the RGB large image to obtain a label image; cropping the RGB large image and the label image by sliding window, the size of each sliding window is 256*256, and the size of the cropped training image is 256*256, and the cropped image is subjected to data augmentation operation to obtain a sample image, for example, data augmentation can be performed by rotation, filtering, image flipping, and adding noise, and the sample images are divided into a training set and a test set according to a set ratio. For example, in this embodiment, 15,000 training images and their corresponding label images are obtained after data augmentation, 12,000 images are randomly selected as the training set of the model, and the remaining 3,000 images are used as the test set of the model; the purpose of random selection is to balance the wheat lodging area and inconsistent lighting in different plots.

[0022] The Improved_Unet network was trained using the training set, employing the Tversky loss function. To compare the networks' ability to identify fallen areas, multiple networks were trained simultaneously: the Segnet network, the Unet network, the Improved_Unet network, and the DeeplabV3+ network. Training was performed on Windows using the Keras deep learning framework and the Python 3.7 programming language, with TensorFlow as the backend. The Adam optimizer was used, with an initial learning rate of 0.001, as per the original settings. The batch size for the Segnet, Unet, and Improved_Unet networks was set to 10, and the number of epochs was set to 100. The final models were obtained after training. The Deeplab V3+ network, using transfer learning, uses pre-trained network parameters derived from training on the public PASCAL VOC 2012 dataset. Given the depth of the network, only the final layer was unfrozen for training on the custom dataset. Because the pre-trained parameters were used, the network quickly reached stability. Therefore, the epoch count was set to 60, and the batch size was also set to 10, using the same training environment as the previous networks. To balance the uneven distribution of lodging and non-lodging samples, all network training used the Tversky loss function.

[0023] Substitute the test set into the trained Improved_Unet network and calculate one or more evaluation metrics: Precision, Dice coefficient, Recall, and Accuracy. If the evaluation metric meets the set threshold, the trained Improved_Unet network is output as the final Improved_Unet network. Otherwise, re-sample and retrain. Using these metrics, the advantages and disadvantages of the four networks can be more objectively and accurately evaluated.

[0024] Specifically, the loss function and evaluation index are calculated according to the following formula:

[0025]

[0026] Where TP represents the division of the fallen area into the fallen area, TN represents the division of the non-fallen area into the non-fallen area, FP represents the division of the non-fallen area into the fallen area, FN represents the division of the fallen area into the non-fallen area, TL is the network loss function, α and β are hyperparameters, α, β∈[0,1] and α+β=1. By adjusting α and β, the trade-off between false positives and false negatives can be controlled. ε is a preset constant. Using the above formula, Tversky loss and the four evaluation indicators can be conveniently calculated.

[0027] Furthermore, the method of using a drone to obtain images of a wheat field at a specified height and growth period includes the following steps: flying the drone to a specified height, using software to plan a route, and flying over the wheat field with an 80% forward overlap, an 80% side overlap, and a flight speed of 1 to 4 m / s; during flight, using the drone's built-in camera to capture images of the wheat field at a set interval; and in the step of geometrically correcting and stitching the wheat field images to obtain a large RGB image, using a Trimble R2 to set ground control points on the ground for geometric correction of the drone images, and using Photoscan software to stitch the wheat field images taken by the drone to obtain a large RGB image. In this embodiment, the drone used is a DJI Phantom 4 Pro, and its built-in camera has an image resolution of 5472×3648 pixels.

[0028] The drone's flight altitude and growth period affect the results of lodging segmentation, so the optimal altitude and growth period should be selected. In this invention, the designated altitude is preferably 20-40 meters, and the designated growth period is the filling or maturity period. Considering that factors such as the monitoring area, operation time, and weather during operation can affect the effectiveness of drone monitoring of crop lodging, and flight altitude is key to solving these problems, exploring flight altitude is of great production significance. In this invention, the evaluation indicators at different flight altitudes were compared and the results are shown in the table below.

[0029] Table 1: Comparison of evaluation indicators at different flight altitudes

[0030] Flight altitude Precision Dice Recall Accuracy 20m 0.907 0.929 0.884 0.933 40m 0.900 0.918 0.871 0.922 80m 0.860 0.887 0.852 0.896 120m 0.845 0.864 0.841 0.881

[0031] As can be seen from Table 1, with the increase of flight altitude, the segmentation accuracy of lodging wheat gradually decreases. However, it should be noted that this is only the result in a specific embodiment of the present invention. The different monitoring location environments will affect the selection of the optimal altitude. Therefore, for a specific monitoring point, data at multiple altitudes can be collected at the same time, and the Improved_Unet networks corresponding to multiple altitudes can be trained. Then, the trained Improved_Unet network at the optimal altitude can be selected according to the evaluation parameters for lodging recognition, thereby obtaining more accurate segmentation accuracy.

[0032] Figure 4The following is a loss convergence plot for the Unet network and the Improved_Unet network proposed in this paper. As can be seen from the plot, the Improved_Unet network converges faster and has a smoother curve than the Unet network. While the Unet network's loss fluctuates dramatically at the beginning, the proposed network's loss changes more gradually. In regions with smaller loss fluctuations, the training and validation losses of the Improved_Unet network are closer, and the curves fluctuate less, indicating that the Improved_Unet network is a more stable model.

[0033] Figure 5 is an arbitrarily cropped image of the study area and the prediction results of the two network models, where Figure 5 a is the original map of any study area, Figure 5 b is a manually labeled image. Figure 5 c is the Unet network prediction graph, Figure 5 d is the prediction image of the Improved_Unet network. From the overall prediction and segmentation effect, the Improved_Unet network segmentation results are better in segmenting large and small fallen areas, and its results are closer to manual annotation; the Unet network segmentation results have some noise points, and the prediction of some small fallen areas is not successful, or the prediction has a larger deviation. These two points can be seen in Figure 5 The area selected in c.

[0034] In order to more precisely display the verification and analysis results of the four networks, the data at a flight altitude of 20m is taken as an example to obtain the corresponding evaluation indicators of the four networks, as shown in Table 2 below.

[0035] Table 2: Comparison of evaluation indicators corresponding to the four networks

[0036]

[0037] Table 2 shows that the Improved_Unet network outperforms the Unet network in all four evaluation metrics. The Improved_Unet network's Accuracy and Dice coefficient are both greater than 0.923, demonstrating excellent segmentation results. The Accuracy for the mature stage is 0.933, the highest of all results. Furthermore, the Improved_Unet network achieves the highest performance across all indices in both the filling and mature stages.

[0038] Although the Improved_Unet network model proposed in this paper has higher accuracy than other classic network models, its applicability and robustness across different years and regions require further verification. This paper uses images of wheat lodging taken at other farms in Shucheng County, Anhui Province, China in 2019 for illustration. The image acquisition platform was a DJI Phantom 4 Pro, and the footage was taken on May 10, 2019, at an altitude of 20 meters. The data processing process is the same as previously described. Figure 6 It is the result of arbitrary cropping of the original map of the study area obtained in 2019, manual annotation, and four network processing. Figure 6 a is the original picture of fallen wheat taken by a drone. Figure 6 b is the result of manual annotation. Figure 6 c, 6d, 6e, and 6f are the prediction results of Improved_Unet, Segnet, Unet, and DeeplabV3+ networks, respectively.

[0039] Table 3: Robustness verification results of different networks

[0040] Network Name Precision Dice Recall Accuracy Segnet 0.768 0.823 0.802 0.840 Unet 0.812 0.855 0.827 0.861 Improved_Unet 0.851 0.892 0.844 0.885 DeeplabV3+ 0.827 0.874 0.816 0.869

[0041] from Figure 6 Visual inspection of the results shows that the lodged wheat areas segmented by the Improved_Unet network are highly consistent with the manually labeled results, regardless of whether the lodged areas are large or small. Furthermore, the segmentation results were quantified and compared using four evaluation metrics: Precision, Dice, Recall, and Accuracy. Table 3 shows that the segmentation performance of the Improved_Unet network proposed in this paper is significantly superior to that of the other three networks, with the results for the four metrics being 0.851, 0.892, 0.844, and 0.885, respectively. Among them, the Segnet network achieved the worst segmentation results, with a Precision of 0.768 and an Accuracy of 0.840, and exhibited the most severe noise and misclassification. The Improved_Unet network outperformed these results by 0.083 and 0.045, respectively. Therefore, the proposed method achieved the best prediction results among the various networks, with smooth edges in the overall segmentation results. The model's robustness, validated using data from different years and regions, demonstrates acceptable persuasiveness.

[0042] The Improved_Unet network was used to study wheat lodging segmentation. Compared with the classic Segnet, Unet, and Deeplabv3+ networks, the Improved_Unet network outperformed the other networks in four evaluation metrics: Precision, Dice, Recall, and Accuracy. Each metric was at least 2 to 3 percentage points higher than the other networks, with a maximum difference of 9 percentage points. Furthermore, the segmentation results were the best when validated with data from different years and regions. This demonstrates that the proposed method is not only more accurate but also more applicable, making it more advantageous for extracting wheat lodging segmentation.

[0043] The present invention also discloses a computer-readable storage medium and an electronic device. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the aforementioned method for segmenting a lodging area based on an Improved_Unet network. The electronic device includes a memory, a processor, and the computer program stored in the memory. When the processor executes the computer program, it implements the aforementioned method for segmenting a lodging area based on an Improved_Unet network.

Claims

1. A lodging area segmentation method based on Improved_Unet network, characterized by: The steps include: Use drones to obtain images of the wheat field to be inspected at a specified height and growth period; Perform geometric correction and splicing on the wheat field image to be detected to obtain a large RGB image to be detected; The RGB image to be detected is input into the trained Improved_Unet network to obtain the segmentation map of the lodging area. The Improved_Unet network is obtained by using the Involution operator to replace the convolution operation at the backbone position of the Unet network. The Improved_Unet network includes an encoder and a decoder; The encoder comprises: The input image is first processed by Conv+BN+Relu to obtain the first feature map; The first feature map is processed by Dense block and Maxpool in sequence to obtain the second feature map; The second feature map is processed three times with Involution+BN+Relu, Dense block and Maxpool to obtain the third feature map; The third feature map is processed by Involution+BN to obtain the fourth feature map; The decoder comprises: After four consecutive Conv+BN+Relu processes, the fourth feature map is upsampled. During the upsampling process, the intermediate feature map generated by the encoder is spliced ​​to the feature map generated by the decoder using skip connections, and the fifth feature map is obtained after upsampling. The fifth feature map is subjected to a 1*1 convolution operation to obtain a segmentation map of the lodging area; The Dense block processing includes the following steps: The input feature map is processed by BN+Relu+Involution to obtain feature map 1; The input feature map and feature map 1 are concatenated and processed by BN+Relu+Involution to obtain feature map 2; The input feature map, feature map 1, and feature map 2 are concatenated and processed by BN+Relu+Conv to obtain the feature map as the output feature map of the Dense block. The Conv is a convolutional layer, BN is a batch normalization layer, Relu is an activation layer, and Maxpool is a maximum pooling layer.

2. The lodging area segmentation method based on the Improved_Unet network according to claim 1, characterized in that: The Improved_Unet network is trained as follows: Use drones to obtain images of wheat fields at a specified height and growth period; The wheat field image is geometrically corrected and stitched to obtain a large RGB image; With the help of expert experience, the RGB image is manually labeled to obtain the label image; The RGB large image and label image are cropped by sliding window, and the cropped images are augmented to obtain sample images. The sample images are divided into training set and test set according to the set ratio. Use the training set to train the Improved_Unet network, and use the Tversky loss function during training; Substitute the test set into the trained Improved_Unet network and calculate one or more evaluation indicators including Precision, Dice coefficient, Recall, and Accuracy. If the evaluation indicators meet the set thresholds, the trained Improved_Unet network is output as the final Improved_Unet network. Otherwise, re-sample and then train again.

3. The lodging area segmentation method based on the Improved_Unet network according to claim 2, characterized in that: The loss function and evaluation index are calculated according to the following formula: Where TP represents the division of the fallen area into the fallen area, TN represents the division of the non-fallen area into the non-fallen area, FP represents the division of the non-fallen area into the fallen area, FN represents the division of the fallen area into the non-fallen area, TL is the network loss function, α and β are hyperparameters, α, β∈[0,1] and α+β=1, and ε is a preset constant.

4. The lodging area segmentation method based on the Improved_Unet network according to claim 2, characterized in that: The method of using a drone to obtain wheat field images at a specified height and in a specified growth period includes the following steps: Fly the drone to a specified altitude. Use software to plan a flight path, flying over the wheat field with a forward overlap of 80%, a side overlap of 80%, and a flight speed of 1 to 4 m / s; During flight, use the drone's built-in camera to capture images of the wheat field at the set intervals. In the step of geometrically correcting and stitching the wheat field images to obtain a large RGB image, Trimble R2 is used to arrange ground control points on the ground for geometric correction of the drone images, and Photoscan software is used to stitch the wheat field images taken by the drone to obtain a large RGB image.

5. The lodging area segmentation method based on the Improved_Unet network according to claim 2, characterized in that: The designated height is 20 to 40 meters, and the designated growth period is the grouting period or the maturity period.

6. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the method for segmenting a lodging area based on an Improved_Unet network as described in any one of claims 1 to 5 is implemented.

7. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory, wherein when the processor executes the computer program, the method for segmenting a lodging area based on the Improved_Unet network as claimed in any one of claims 1 to 5 is implemented.

Images