Satellite remote sensing rice lodging detection method and system based on deep learning

Abstract

The invention relates to the technical field of rice lodging detection, in particular to a satellite remote sensing rice lodging detection method and system based on deep learning, and the method comprises the steps: carrying out the clustering and partitioning of a target research region, selecting a training sample in each region, carrying out the refining preprocessing of the classified training samples, and carrying out the recognition of the training samples; correcting the classification condition in combination with spatial region monitoring features to obtain a rice lodging data set; performing dynamic processing on the rice lodging data set by adopting a preset data enhancement strategy to obtain diversified training samples; sequentially performing feature extraction and spatial detail recovery on the training sample based on a preset deep learning detection model to obtain a rice lodging region segmentation result; and the loss function of the deep learning detection model is optimized by adopting FocalLoss and DiceLoss. According to the method, the problems of relatively large error, relatively low robustness and need of manual correction in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rice lodging detection, and particularly relates to a satellite remote sensing rice lodging detection method and system based on deep learning. BACKGROUND

[0002] At present, most of the main rice lodging detection technologies are to use unmanned aerial vehicle images for detection, but the image coverage range is small, and the rice lodging detection is often focused on the field level, and it is difficult to carry out large-scale rice lodging detection. A small part of people use satellite remote sensing to solve this problem, but rice disasters often occur in the rainy season, and are affected by the resolution of remote sensing images. The work of rice lodging detection on the meter-level remote sensing image often focuses on the use of traditional machine learning algorithms. However, due to the influence of sensor models and collection time, different remote sensing images often present different characteristics (mainly in color), and therefore the robustness of the traditional machine learning is low, and it is difficult to adapt to complex application scenarios. SUMMARY

[0003] The purpose of the present application is to provide a satellite remote sensing rice lodging detection method and system based on deep learning, which solves the problems of large error, low robustness and the need for manual correction in the prior art.

[0004] To achieve the above purpose, the present application provides a satellite remote sensing rice lodging detection method based on deep learning, comprising the following steps: S1, clustering and partitioning the target research area, selecting training samples in each area, refining and preprocessing the classified training samples, and correcting the classification by combining the spatial area monitoring features, to obtain a rice lodging data set; S2, dynamically processing the rice lodging data set by using a preset data enhancement strategy to obtain diversified training samples; S3, based on the preset deep learning detection model, sequentially performing feature extraction and spatial detail recovery on the training samples to obtain a rice lodging area segmentation result; S4, using FocalLoss and DiceLoss to optimize the loss function of the deep learning detection model.

[0005] In some embodiments of the present application, in S1, the clustering and partitioning of the target research area and the selection of training samples in each area include: S11, using spectral features, texture features and vegetation indexes to cluster and partition the target research area, and dividing the fields with similar planting structures and management methods into the same area; S12, selecting training samples for different ground object types, and classifying the rice lodging disaster situation; The ground object types at least include normal crops, disaster-affected crops, forests, houses, and roads.

[0006] In some embodiments of the present application, in S1, the classified training samples are subjected to refinement preprocessing, and the classification is corrected in combination with the spatial region monitoring features to obtain the rice lodging dataset. S13, the classified training samples are subjected to K-means clustering optimization, so that the crop boundaries of the training samples are further refined. S14, the crop boundaries of the optimized training samples are further corrected and verified in combination with the field investigation data, the cultivated land spatial distribution information, and the artificial visual interpretation results to obtain the rice lodging dataset.

[0007] In some embodiments of the present application, in S14, the further correction and verification of the crop boundaries of the optimized training samples in combination with the field investigation data, the cultivated land spatial distribution information, and the artificial visual interpretation results include: For places where the classification result is not clear, field investigation, cultivated land spatial distribution information, and online maps of other periods are obtained, the differences between the classification in the training samples and the actual investigation are judged based on human eye contrast analysis, and the classification in the training samples is modified.

[0008] In some embodiments of the present application, in S2, the rice lodging dataset is dynamically processed by using a preset data enhancement strategy to obtain diversified training samples, including: The original images and the corresponding label images in the rice lodging dataset are dynamically processed by using an online data enhancement strategy; Before each round of training starts, the system randomly performs rotation, translation, scaling, brightness change, and light disturbance operations to generate diversified training samples.

[0009] In some embodiments of the present application, in S3, the preset deep learning detection model adopts an encoder-decoder structure, wherein the encoder is used for feature extraction of different ground objects in the input training samples, and the decoder is used for spatial detail recovery.

[0010] In some embodiments of the present application, in S3, the encoder adopts a convolutional neural network (CNN) to extract features of normal rice and lodging rice of different ground objects in the input remote sensing images to obtain multi-scale feature representations containing local textures, edge information, and high-level semantics. The decoder performs spatial detail recovery on the deep feature maps output by the encoder, gradually recovers the spatial resolution and detail structure, and obtains a lodging area segmentation result with consistent spatial scale, clear boundaries, and complete textures.

[0011] In some embodiments of the present application, in S3, the deep learning detection model further comprises a Transformer module, which captures long-distance dependencies through a self-attention mechanism; and the decoder adopts a step-by-step upsampling and jump connection mechanism to fuse deep semantic features and shallow spatial features.

[0012] In some embodiments of the present application, a deep learning-based satellite remote sensing rice lodging detection system is also disclosed, comprising: A data acquisition module is configured to cluster and partition a target research area, select training samples in each region, perform fine preprocessing on the classified training samples, and correct the classification by combining spatial region monitoring features to obtain a rice lodging data set. A dynamic processing module is configured to dynamically process the rice lodging data set using a preset data enhancement strategy to obtain diversified training samples. A model detection module is configured to sequentially perform feature extraction and spatial detail recovery on the training samples based on a preset deep learning detection model to obtain a rice lodging region segmentation result. A model optimization module is configured to optimize the loss function of the deep learning detection model using FocalLoss and DiceLoss.

[0013] The advantages and beneficial effects of the present application over the prior art are: The present application uses deep learning to detect rice lodging, extracts deeper features of rice lodging, and can more accurately extract the spatial distribution of rice lodging, overcoming the feature difference problem caused by remote sensing images from different devices and different collection times, effectively reducing the probability of missed detection and false detection, and reducing the involvement of artificial.

[0014] The technical solutions of the present application will be further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a deep learning-based satellite remote sensing rice lodging detection method in an embodiment of the present application is shown. Figure 2 A structural diagram of a deep learning-based satellite remote sensing rice lodging detection system in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] In the description of the present application, it should be pointed out that the terms "upper", "lower", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly used when the product of the present application is used, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description of the present application, it should be pointed out that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0017] The present application proposes a deep learning model combining convolutional neural network (CNN) and Transformer for extracting rice lodging features from satellite remote sensing images. The model uses CNN to capture local texture details of rice lodging, and combines Transformer to extract global features, organically fuses texture and geometric information, and realizes accurate discrimination of rice lodging state. This method can extract lodging spatial distribution with high precision in complex meter-level remote sensing images, adapt to images of different resolutions, and has strong robustness.

[0018] The embodiments of the present application will be described in detail below with reference to the drawings.

[0019] As Figure 1 shown, the present application provides a satellite remote sensing rice lodging detection method based on deep learning, comprising the following steps: S1, clustering and partitioning the target research area, selecting training samples in each area, refining and preprocessing the classified training samples, and correcting the classification by combining spatial area monitoring features, to obtain a rice lodging data set.

[0020] S2, using a preset data enhancement strategy to dynamically process the rice lodging data set to obtain diversified training samples.

[0021] S3, based on the preset deep learning detection model, the training samples are sequentially subjected to feature extraction and spatial detail recovery to obtain rice lodging region segmentation results.

[0022] S4, using FocalLoss and DiceLoss to optimize the loss function of the deep learning detection model.

[0023] This invention uses deep learning to detect rice lodging, extracting deeper features of rice lodging and more accurately extracting the spatial distribution of rice lodging. It overcomes the feature differences caused by remote sensing images from different devices and different acquisition times, effectively reducing the probability of missed and false detections and reducing human intervention.

[0024] Due to the massive amount of high-resolution remote sensing imagery and the high cost of manual annotation, this study adopts a machine learning-based regional annotation strategy. Specifically, the study area is first clustered and partitioned using spectral features, texture features, and vegetation indices (such as NDVI and SAVI), grouping plots with similar planting structures and management practices into the same region. Then, training samples are selected within each region to classify rice lodging damage. During sample selection, training samples are chosen for different land cover types—including normal crops, damaged crops, forests, houses, and roads. Introducing multi-class samples effectively reduces the probability of misclassification of target crops, thereby improving the accuracy of the dataset. Subsequently, K-means clustering is used to optimize the classification results, further refining crop boundaries. Finally, the plot boundaries are further corrected and validated by combining field survey data, spatial distribution information of cultivated areas, and visual interpretation results, ultimately obtaining a rice lodging dataset that accurately reflects the actual situation. This method fully leverages the consistency of crop features within the region, improving the spatial consistency and precision of the labeled data, and providing high-quality samples for model training.

[0025] In some embodiments of this application, in S1, clustering and partitioning the target study area, and selecting training samples within each area, includes: S11. Cluster and partition the target study area using spectral features, texture features and vegetation indices, and divide plots with similar planting structures and management methods into the same area; S12. Select training samples for different land cover types and classify the lodging and damage of rice. Among them, the land cover types include at least normal crops, disaster-stricken crops, forests, houses, and roads.

[0026] In this stage of processing the dataset, training samples were selected from normal crops, disaster-stricken crops, forests, houses, roads, etc. Therefore, using these training samples will yield all the above distributions. Only normal crops and disaster-stricken crops are needed. The other samples, such as forests, houses, and roads, are used to assist in classification and make the classification of disasters more accurate.

[0027] In some embodiments of this application, in S1, the classified training samples undergo refined preprocessing, and the classification is corrected by combining spatial region monitoring features, resulting in a rice lodging dataset including: S13. Perform K-means clustering optimization on the classified training samples to further refine the crop boundaries of the training samples; S14. Combining field survey data, spatial distribution information of cultivated areas, and results of manual visual interpretation, the crop boundaries of the optimized training samples are further corrected and verified to obtain the rice lodging dataset.

[0028] It is important to understand that during the dataset construction phase, all land cover types are first classified, then optimized, and finally, all categories except crops and disaster-affected crops are removed, resulting in the rice lodging dataset. Therefore, the training samples after classification in this embodiment are the classification results of all land cover types.

[0029] K-means is an unsupervised clustering method based on iterative optimization. Its goal is to minimize intra-cluster variance given a number of clusters K. Specifically, the algorithm first randomly initializes cluster centers, then assigns pixels to the nearest cluster based on the distance between pixels and each center (such as Euclidean distance). Subsequently, it updates the cluster centers based on the pixel mean of each cluster, and iterates through the "assignment-update" process until convergence.

[0030] In some embodiments of this application, S14, further refining and validating the crop boundaries of the optimized training samples by combining field survey data, spatial distribution information of cultivated areas, and results of manual visual interpretation includes: For locations with unclear classification results, obtain field surveys, spatial distribution information of cultivated areas, and online maps from other time periods. Based on human visual comparison analysis, determine the difference between the classification in the training samples and the actual survey results, and modify the classification of the training samples accordingly.

[0031] In some embodiments of this application, in S2, a preset data augmentation strategy is used to dynamically process the rice lodging dataset to obtain diverse training samples, including: An online data augmentation strategy was used to dynamically process the original images and corresponding labeled images in the rice lodging dataset. Before each training round begins, the system randomly performs rotation, translation, scaling, brightness changes, and lighting perturbation operations to generate diverse training samples.

[0032] In some embodiments of this application, in S3, the preset deep learning detection model adopts an encoder-decoder structure, wherein the encoder is used to extract features from different land features in the input training samples, and the decoder is used to restore spatial details.

[0033] In some embodiments of this application, in S3, the encoder uses a convolutional neural network (CNN) to extract features from normal rice and lodged rice in different land features in the input remote sensing image, and obtain a multi-scale feature representation that includes local texture, edge information and high-level semantics. The decoder performs spatial detail restoration on the deep feature map output by the encoder, gradually restoring the spatial resolution and detailed structure, and obtaining the segmentation result of the fallen area with consistent spatial scale, clear boundaries and complete texture with the original image.

[0034] On the test set, the precision of the detection result was 0.8174, while in the actual application of this invention, the precision after testing can reach more than 0.85.

[0035] In some embodiments of this application, in S3, the deep learning detection model further includes a Transformer module, which captures long-distance dependencies through a self-attention mechanism; the decoder employs a progressive upsampling and skip connection mechanism to fuse deep semantic features with shallow spatial features.

[0036] It's important to understand that the Transformer is used in conjunction with the CNN in the encoder, while skip connections are connection mechanisms that pass information from the encoder to the decoder, assisting the decoder in restoring classification edges during the upsampling process.

[0037] In some embodiments of this application, FocalLoss and DiceLoss are used to optimize the loss function of the deep learning detection model. Focal is used to address the problem of imbalanced sample classes and enhance the model's ability to identify a small number of difficult-to-identify disaster pixels. Dice is used to improve the measurement effect of the overlap of the detection result region.

[0038] In some embodiments of this application, such as Figure 2 As shown, a satellite remote sensing system for detecting lodging of rice based on deep learning is also disclosed, including: The data acquisition module is used to cluster and partition the target research area, select training samples in each area, refine the preprocessing of the classified training samples, and correct the classification by combining spatial regional monitoring features to obtain the rice lodging dataset. The dynamic processing module is used to dynamically process the rice lodging dataset using preset data augmentation strategies to obtain diverse training samples. The model detection module is used to extract features and restore spatial details of training samples based on a preset deep learning detection model to obtain the segmentation results of rice lodging areas. The model optimization module uses FocalLoss and DiceLoss to optimize the loss function of the deep learning detection model.

[0039] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based satellite remote sensing rice lodging detection method, characterized in that, The method comprises the following steps: S1, clustering and partitioning the target research area, selecting training samples in each area, refining the classified training samples, and combining spatial area monitoring features to correct the classification, to obtain a rice lodging data set; S2, using a preset data enhancement strategy to dynamically process the rice lodging data set to obtain diversified training samples; S3, based on a preset deep learning detection model, sequentially performing feature extraction and spatial detail recovery on the training samples to obtain a rice lodging area segmentation result; S4, using FocalLoss and DiceLoss to optimize the loss function of the deep learning detection model.

2. The method according to claim 1, wherein, In S1, the target research area is clustered and partitioned, and training samples are selected in each area, comprising: S11, using spectral features, texture features, and vegetation indices to cluster and partition the target research area, and dividing plots with similar planting structures and management methods into the same area; S12, selecting training samples for different ground object types, and classifying rice lodging disaster conditions; The ground object types at least include normal crops, disaster crops, forests, houses, and roads.

3. The method according to claim 2, wherein, In S1, the classified training samples are refined and preprocessed, and the classification is corrected by combining spatial area monitoring features to obtain a rice lodging data set, comprising: S13, K-means clustering optimization is performed on the classified training samples to further refine the crop boundaries of the training samples; S14, combining field investigation data, cultivated land spatial distribution information, and artificial visual interpretation results to further correct and verify the crop boundaries of the optimized training samples to obtain a rice lodging data set.

4. The method according to claim 3, wherein, In S14, combining field investigation data, cultivated land spatial distribution information, and artificial visual interpretation results to further correct and verify the crop boundaries of the optimized training samples, comprising: For places with unclear classification results, obtain field investigation, cultivated land spatial distribution information, and online maps of other periods, based on human eye comparison and analysis, determine the difference between the classification in the training samples and the actual investigation, and modify the classification in the training samples.

5. The method according to claim 4, wherein, In S2, the preset data enhancement strategy is used to dynamically process the rice lodging data set to obtain diversified training samples, comprising: Using an online data enhancement strategy to dynamically process the original images and corresponding label images in the rice lodging data set; Before each round of training, the system randomly performs rotation, translation, scaling, brightness change, and light disturbance operations to generate diversified training samples.

6. The method according to claim 5, wherein, In S3, the preset deep learning detection model uses an encoder-decoder structure, wherein the encoder is used to extract features of different ground objects in the input training samples, and the decoder is used to recover spatial details.

7. The method according to claim 6, wherein, In S3, the encoder uses a convolutional neural network (CNN) to extract features of normal rice and lodging rice in different ground objects in the input remote sensing image to obtain multi-scale feature representations containing local texture, edge information, and high-level semantics; The decoder recovers spatial details for the deep feature map output by the encoder, gradually recovers spatial resolution and detailed structure, and obtains a segmentation result of the lodging area that is consistent with the spatial scale of the original image, has clear boundaries, and has complete texture.

8. The method according to claim 7, wherein, In the S3, the deep learning detection model further includes a Transformer module that captures long-distance dependencies through a self-attention mechanism; and the decoder adopts a step-by-step upsampling and skip connection mechanism to fuse deep semantic features and shallow spatial features.

9. A deep learning-based satellite remote sensing rice lodging detection system, characterized in that, The method comprises the following steps: The data acquisition module is configured to cluster and partition a target research area, select training samples in each region, perform fine preprocessing on the classified training samples, correct the classification by combining spatial region monitoring features, and obtain a rice lodging dataset; The dynamic processing module is configured to dynamically process the rice lodging dataset by using a preset data enhancement strategy to obtain diversified training samples; The model detection module is configured to sequentially perform feature extraction and spatial detail recovery on the training samples based on a preset deep learning detection model to obtain a rice lodging region segmentation result; The model optimization module is configured to optimize the loss function of the deep learning detection model by using FocalLoss and DiceLoss.

Images