Aedes habitat identification and extraction method based on remote sensing index and deep learning

Abstract

The invention discloses a remote sensing index and deep learning-based aedes habitat identification and extraction method, which is characterized in that a multi-source information deep fusion method is adopted to realize automatic identification and refined extraction of typical aedes habitat types; the method specifically comprises the steps of acquiring a habitat multispectral image, constructing a high-resolution multiband fused building label data set, training and vectorizing a building identification model, extracting a water body based on a water body index and vectorizing, extracting vegetation based on a vegetation index and vectorizing and the like. Compared with the prior art, the method has good universality and expandability, provides a solid data basis for aedes monitoring, risk assessment and prevention and control decision, reduces the workload of manual patrol and image interpretation, effectively improves the efficiency of aedes habitat extraction, is suitable for wide-range aedes habitat monitoring and dynamic updating requirements, and has a wide application prospect. And the method can be popularized and applied to any region, and provides efficient and accurate technical support for aedes prevention and control in different regions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mosquito monitoring, in particular to an Aedes habitat recognition and extraction method based on remote sensing index and deep learning. BACKGROUND

[0002] Vector organisms refer to organisms that play a role in disease transmission and can transmit pathogens to humans. Common vector organisms include mosquitoes, flies, cockroaches, midges, mites, fleas, ticks, horseflies and mice. Among them, mosquitoes play a key role in the transmission of many diseases, so they need to be monitored. Mosquitoes not only sting and annoy people, but also carry a variety of disease vectors and transmit them to humans, causing global public health problems. The World Health Organization declared mosquitoes as the "number one enemy" in 1996. In recent years, mosquito-borne diseases, especially diseases transmitted by Aedes mosquitoes, such as malaria, dengue fever, yellow fever and Zika virus, have become major public health challenges. With global warming, the acceleration of urbanization, the booming of tourism and trade, and the continuous change of ecological environment, the survival rate, population size and activity range of mosquitoes have increased, increasing the transmission mode of mosquito-borne diseases, and the frequent outbreaks of mosquito-borne diseases have made it increasingly important to control their numbers.

[0003] Currently, the identification of potential mosquito breeding sites relies on manual field investigation and traditional statistical analysis methods, which have certain feasibility in small areas, but have the problems of high work intensity, low efficiency and difficulty in covering large areas. With the development of remote sensing technology, more and more studies have begun to try to extract potential mosquito breeding habitats using high-resolution remote sensing images, especially using water bodies, vegetation and buildings as environmental indicators. However, the existing methods still have many shortcomings: first, although the remote sensing index method has certain advantages in the identification of natural features such as water bodies and vegetation, it is difficult to meet the requirements of high-precision extraction due to the factors of fuzzy boundaries and spectral mixing of urban complex surfaces; second, the recognition ability of buildings, a type of urban artificial environment, is weak, and this type of area is an important breeding place for mosquitoes such as Aedes albopictus; third, current methods mostly fail to effectively fuse multi-band information and lack deep mining of image spatial structure and semantic features, making it difficult to achieve fine description of mosquito habitats in high-heterogeneity urban environments.

[0004] In summary, the existing Aedes habitat recognition technology has the problems of limited Aedes habitat recognition accuracy, insufficient complex ground object extraction capability, and lack of multi-source information deep fusion. Therefore, how to develop a method that can effectively recognize natural environmental factors such as vegetation coverage and water body distribution, and also extract complex man-made objects such as buildings, improve the spatial recognition capability of Aedes habitat, and realize the fine and automatic recognition of Aedes habitat, is a technical problem that needs to be solved in this technical field. SUMMARY

[0005] The purpose of the present application is to provide an Aedes habitat recognition and extraction method based on remote sensing index and deep learning, which uses multi-source information deep fusion method to identify and extract water body and vegetation by remote sensing index, and realizes the automatic recognition and fine extraction of typical Aedes habitat types by combining deep learning and remote sensing index. This method can construct an automatic processing flow covering a large area, significantly reducing the workload of manual patrol and image interpretation, and is suitable for large-scale Aedes habitat monitoring and dynamic updating requirements. It can be applied to any region to provide efficient and accurate technical support for Aedes prevention and control in different regions, and has good universality and scalability. It effectively solves the problems of limited Aedes habitat recognition accuracy, insufficient complex ground object extraction capability, and lack of multi-source information deep fusion, and has good application prospect.

[0006] The purpose of the present application is achieved by an Aedes habitat recognition and extraction method based on remote sensing index and deep learning, which is characterized in that the method specifically includes: Step 1: Obtain habitat multispectral image Construct high-resolution multi-band fused building label dataset and building recognition model training and vectorization, which is water body and vectorization based on water body index extraction and vegetation and vectorization based on vegetation index extraction. The habitat multispectral image is obtained by using a UAV equipped with an orthographic camera module to shoot high-resolution habitat images, which are radiometrically calibrated and two-dimensionally modeled, then numerically converted and band-synthesized to obtain six-band multispectral habitat images.

[0007] Step 2: Construct high-resolution multi-band fused building label dataset The six-band multispectral habitat image is dimensionally reduced by principal component analysis to obtain a three-band multispectral habitat image of the first three principal components. Based on the three-band multispectral habitat image and field investigation, building vector labels are made and a step size is set. The three-band multispectral habitat image and building vector labels are subjected to data augmentation operations such as cropping and rotating to obtain a high-resolution multi-band fused building label dataset.

[0008] Step 3: Building recognition model training and vectorization Using the building label data set of high-resolution multi-band fusion, the building recognition model is trained and the building is semantically segmented. The semantic segmentation result is spliced and vectorized to obtain the final building vector result.

[0009] Step 4: Water body extraction and vectorization based on water body index The water body is extracted based on the water body index, that is, the NDWI index is calculated, the optimal segmentation threshold is determined according to the frequency distribution histogram by the empirical threshold method, the three-band multispectral habitat image is binarized and vectorized, and the final water body vector result is obtained.

[0010] Step 5: Extraction and vectorization of vegetation based on vegetation index The vegetation is extracted based on the vegetation index, that is, the NDVI index is calculated, the optimal segmentation threshold is determined by the Otsu method, the three-band multispectral habitat image is binarized and vectorized, and the final vegetation vector result is obtained.

[0011] The habitat multispectral image is obtained using the Puma D2000 unmanned aerial vehicle system, and the D-CAM2000 orthographic camera module is carried to plan the habitat flight route, determine the heading overlap rate of 80%, the lateral overlap rate of 70%, the unmanned aerial vehicle flight height of 120m, and shoot the habitat six-band high-resolution remote sensing image; the radiation calibration is carried out using YusenseRef software, the two-dimensional modeling is carried out on the image shot by the unmanned aerial vehicle using Pix4DV4.5.6, six high-resolution single-band raster images are obtained, which correspond to 6 spectral bands (blue 450nm, green 555nm, red 660nm, red edge 720nm, red edge 750nm, near-infrared 840nm); the numerical conversion is carried out on the single-band raster image using the Raster Calculator tool in ArcGIS Pro software, first converting the integer type of the image to floating point type, then dividing it by 65536 to convert to reasonable numerical value, that is, using the Composite Bands tool in ArcGIS Pro software to band synthesize the six images after numerical conversion, finally obtain the six-band habitat multispectral image, the habitat multispectral image obtains 8296 six-band multispectral images with a pixel size of 1280 × 960, and the ground sampling interval is 8.9 cm / px.

[0012] The building label dataset constructed by high-resolution multi-band fusion utilizes principal component analysis to perform dimensionality reduction processing on six-band multispectral images, extracts the first three principal components, and obtains three-band high-resolution remote sensing images; utilizes the Training Samples Manager in the ArcGIS Pro software to perform screen tracking vectorization on the buildings in the remote sensing images, and experts make building vector labels; utilizes the Export Training Data For Deep Learning tool in the ArcGIS Pro software to set the slice length and width, and to expand the dataset by taking one-eighth of the slice length as a step and a rotation angle of 90°, to obtain building label single-band binary TIFF images and corresponding slice three-band TIFF images; the building label dataset constructed by high-resolution multi-band fusion generates a total of 4851 building label TIFF images and corresponding slice TIFF images, with a total data amount of about 10 GB.

[0013] The building recognition model training and vectorization utilizes the building label dataset to be divided into a test set, a validation set and a training set: 20% of the building label dataset is taken as the test set, 80% of the remaining 80% is taken as the training set, and the remaining 20% is taken as the validation set; a U 2 -Net deep learning model is trained, and the optimal model parameters are used to predict the test set, and the final test set IOU average is 0.965, the Precision average is 0.986, and the Recall average is 0.975; the trained model is used to perform semantic segmentation and vectorization on the buildings; the Mosaic to New Raster tool in the ArcGIS Pro software is utilized to mosaic the semantic segmentation result into a raster image; the Select By Attribute tool in the ArcGIS Pro software is utilized to extract the records with the gridcode field value of 1, to obtain the final building vector result.

[0014] The water body extraction and vectorization based on the water body index utilizes the Band Arithmetic tool in the ArcGIS Pro software to calculate the NDWI index; the optimal segmentation threshold is determined to be 0.14 according to the frequency distribution histogram by the experience threshold method; the Raster Calculator tool in the ArcGIS Pro software is utilized to perform binaryzation processing on the NDWI image; the Raster to Polygon tool in the ArcGIS Pro software is utilized to convert the raster to shapefile surface vector data; the Select By Attribute tool in the ArcGIS Pro software is utilized to extract the records with the gridcode field value of 1, to obtain the final water body vector result.

[0015] The vegetation and vectorization based on the vegetation index extracts the NDVI index by using the Band Arithmetic tool in the ArcGIS Pro software; the optimal segmentation threshold is determined as 0.38 by the Otsu method; the NDVI image is binarized by using the Raster Calculator tool in the ArcGIS Pro software; the raster is converted into shapefile surface vector data by using the Raster to Polygon tool in the ArcGIS Pro software; and the final vegetation vector result is obtained by extracting the records with the gridcode field value of 1 by using the Select ByAttribute tool in the ArcGIS Pro software.

[0016] Compared with the prior art, the present application has the following beneficial technical effects and significant technical progress: 1) Strong robustness and high recognition accuracy: effectively solves the problems of spatial resolution, recognition accuracy and processing efficiency, realizes high-precision identification and fine extraction of Aedes habitat, and effectively improves the efficiency of Aedes habitat extraction. 2) High-precision target recognition: using high-spatial-resolution unmanned aerial vehicle multispectral images overcomes the problem of insufficient resolution of traditional satellite remote sensing, can accurately identify buildings, water bodies (rivers and lakes, etc.) and vegetation cover areas, significantly reduces the omission rate of each ground object pixel in the Aedes habitat, and improves the accuracy of identification, providing a solid data foundation for Aedes monitoring, risk assessment and prevention and control decision-making. 3) Automatic and efficient processing: by combining deep learning and remote sensing index, automatic identification and fine extraction of typical Aedes habitat types are realized, which has good universality and scalability, can construct an automatic processing flow covering a large area, significantly reduces the workload of manual patrol and image interpretation, is suitable for large-scale Aedes habitat monitoring and dynamic updating requirements, and can be applied to any region, providing efficient and accurate technical support for Aedes prevention and control in different regions. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. EMBODIMENT

[0019] Referring to FIG. 1, the identification and extraction of the Aedes habitat are carried out in the following steps: Step 1: Obtain habitat multispectral image 1-1: Use the Feima D2000 unmanned aerial vehicle system, and carry the D-CAM2000 orthographic camera module to plan the route of the habitat, determine the heading overlap rate as 80%, the lateral overlap rate as 70%, the unmanned aerial vehicle flight height as 120 m, and shoot the habitat six-band high-resolution remote sensing image; 1-2: Use the YusenseRef software to perform radiation calibration; 1-3: Use Pix4D V4.5.6 to perform two-dimensional modeling on the pictures shot by the unmanned aerial vehicle, obtain six high-resolution raster images, which correspond to six spectral bands (blue 450 nm, green 555 nm, red 660 nm, red edge 720 nm, red edge 750 nm, near-infrared 840 nm) respectively, and have all completed geographic registration and unified projection to the WGS 1984 UTM Zone 51N coordinate system; 1-4: Use the Raster Calculator tool in the ArcGIS Pro software to perform numerical conversion on the single-band image, that is, first convert the integer type of the image to the floating-point type, and then divide it by 65536 to convert it to a reasonable numerical value; 1-5: Use the CompositeBands tool in the ArcGIS Pro software to perform band synthesis on the six single-band raster images after numerical conversion, and finally obtain 8296 six-band high-resolution multispectral habitat images with a pixel size of 1280 × 960, with a ground sampling distance of 8.9 cm / px.

[0020] Step 2: Construct high-resolution multi-band fused building label dataset 2-1: Use principal component analysis to process the six-band multispectral image for dimension reduction, extract the first three principal components, and obtain three-band high-resolution remote sensing image; 2-2: Use the Training Samples Manager in the ArcGIS Pro software to perform screen tracking vectorization on the buildings in the remote sensing image, that is, use the polygon to outline the boundaries of the buildings, and the generated shapefile polygon vector file saves the labels of the buildings, the Classname field stores the category name, and the Classvalue field stores the value corresponding to the category; 2-3: Use the Export Training Data For Deep Learning tool in ArcGIS Pro software, set the slice length and width to 512, build the building dataset, and expand the dataset with an eighth of the slice length (64) as the step size and a rotation angle of 90°. Generate 4,851 building label single-band binary TIFF images (1 indicates that the pixel is a building, and 0 indicates that the pixel is the background) and their corresponding slice three-band TIFF images. The total data amount is about 10 GB, and the high-resolution multi-band fused building label dataset is obtained.

[0021] Step 3: Building recognition model training and vectorization 3-1: Divide the high-resolution multi-band fused building label dataset into three subsets, i.e. 20% (970 images) as the test set, 80% of the remaining 80% (3,105 images) as the training set, and 20% (776 images) as the validation set. 3-2: Train using the U 2 -Net model, train the model for a total of 150 rounds, use loss and intersection over union as precision evaluation indicators during model training, perform backward propagation on the loss obtained from the training set to train model parameters, and use the validation set to evaluate the generality and robustness of the model to determine whether there is overfitting. 3-3: Optimize the model using the Adam optimizer, with an initial learning rate of 0.001 and dynamic adjustment of the learning rate through the CosineAnnealingLR method. The batch size is set to 8, the training set Loss is reduced from 0.32 to 0.008, and the IOU is improved from 0.54 to 0.98. The change trend of the validation set Loss and IOU is consistent with that of the training set. 3-4: Use the optimal model parameters to predict the test set, with an average IOU of 0.965, an average Precision of 0.986, and an average Recall of 0.975. 3-5: Merge the predicted raster images into one image using the Mosaic to New Raster tool in ArcGIS Pro software. 3-6: Convert the raster to shapefile surface vector data using the Raster to Polygon tool in ArcGIS Pro software. 3-7: Extract the records with gridcode field value of 1 using the Select By Attribute tool in ArcGIS Pro software to obtain the final building vector result.

[0022] Step 4: Water body extraction and vectorization based on water body index 4-1: Using the Band Arithmetic tool in ArcGIS Pro software, selecting Method as NDWI, inputting band index, assigning variables to the near-infrared and green bands of the image, and calculating the NDWI image; 4-2: The optimal threshold value for distinguishing water bodies from non-water bodies is determined by the frequency histogram using the empirical threshold method, and the optimal threshold value is 0.14; 4-3: Using the Raster Calculator tool in ArcGIS Pro software to perform binaryzation processing on the NDWI image; 4-4: Using the Raster to Polygon tool in ArcGIS Pro software to convert the raster to shapefile surface vector data; 4-5: Using the Select By Attribute tool in ArcGIS Pro software to extract records with gridcode field value of 1, and obtaining the final water body vector result.

[0023] Step 5: Extracting vegetation and vectorization based on vegetation index 5-1: Using the Band Arithmetic tool in ArcGIS Pro software, selecting Method as NDVI, inputting band index, assigning variables to the near-infrared and red bands of the image, and calculating the NDVI image; 5-2: The optimal threshold value for distinguishing vegetation from non-vegetation is determined by the Otsu adaptive threshold segmentation method, and the optimal threshold value is 0.38; 5-3: Using the Raster Calculator tool in ArcGIS Pro software to perform binaryzation processing on the NDVI image; 5-4: Using the Raster to Polygon tool in ArcGIS Pro software to convert the raster to shapefile surface vector data; 5-5: Using the Select By Attribute tool in ArcGIS Pro software to extract records with gridcode field value of 1, and obtaining the final vegetation vector result.

[0024] The above is only a further description of the present application, and is not intended to limit the patent. Any equivalent implementation of the present application shall be included within the scope of the claims of the present patent.

Claims

1. A method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning, characterized in that: The method specifically includes: Step 1: Obtain multispectral imagery of the habitat Use the orthophoto camera module carried by the drone to capture high-resolution habitat images, and then perform numerical conversion and band synthesis after radiometric calibration and two-dimensional modeling to obtain six-band multispectral habitat images; Step 2: Build a high-resolution multi-band fused building label dataset The principal component analysis method was used to reduce the dimensionality of the six-band multispectral habitat image to the first three principal components, obtaining a three-band multispectral habitat image. Based on the three-band multispectral habitat image and field surveys, building vector labels were generated. The slice size and step size were set, and the three-band multispectral habitat image and building vector labels were cropped and rotated for data expansion to obtain a high-resolution multi-band fused building label dataset. Step 3: Building recognition model training and vectorization A building recognition model is trained using a high-resolution multi-band fused building label dataset. Buildings are semantically segmented and the segmentation results are concatenated and vectorized to obtain building vector results. Step 4: Extract water bodies and vectorize based on water body index The NDWI index was calculated, and the optimal segmentation threshold was determined based on the frequency distribution histogram using the empirical threshold method. The three-band multispectral habitat image was binarized and vectorized to obtain the water body vector result. Step 5: Extract vegetation and vectorize based on vegetation index The NDVI index was calculated, the optimal segmentation threshold was determined using the Otsu method, and the three-band multispectral habitat image was binarized and vectorized to obtain the vegetation vector result.

2. The method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning according to claim 1, characterized in that: The obtaining of habitat multispectral images specifically includes: 1-1: Use the Pegasus D2000 UAV system equipped with a D-CAM2000 orthophoto camera module to plan a route over the habitat of Aedes albopictus, determining a heading overlap rate of 80%, a sideways overlap rate of 70%, and a flight altitude of 120m. 1-2: Use YusenseRef software to perform radiometric calibration on the six-band high-resolution remote sensing images of the habitat; 1-3: Use Pix4D V4.5.6 to perform 2D modeling on the drone-captured images, generating six high-resolution single-band raster images corresponding to six spectral bands: blue (450 nm), green (555 nm), red (660 nm), red-edge (720 nm), red-edge (750 nm), and near-infrared (840 nm). 1-4: Use the Raster Calculator tool in ArcGIS Pro software to convert the single-band raster image to a numerical value. First, convert the image's integer type to a floating point type, and then divide it by 65536 to convert the numerical value. 1-5: Use the Composite Bands tool in ArcGIS Pro software to synthesize the six images after numerical conversion to obtain a six-band multispectral habitat image.

3. The method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning according to claim 1, characterized in that: The construction of a high-resolution multi-band fusion building label dataset specifically includes: 2-1: Use principal component analysis to reduce the dimensionality of the six-band multispectral image, extract the first three principal components, and obtain a three-band high-resolution remote sensing image; 2-2: Using the Training Samples Manager in ArcGIS Pro software, screen-track and vectorize buildings in remote sensing images, and experts create vector labels for buildings; 2-3: Use the Export Training Data For Deep Learning tool in ArcGIS Pro to expand the dataset with a step size of one-eighth of the slice length and a rotation angle of 90° to obtain a single-band binary TIFF image of the building label and its corresponding three-band TIFF image of the slice, thus obtaining a high-resolution multi-band fused building label dataset.

4. The method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning according to claim 1, characterized in that: The building recognition model training and vectorization specifically include: 3-1: Divide the high-resolution multi-band fusion building label dataset into a test set, a validation set, and a training set: 20% of the dataset is used as a test set, 80% of the remaining 80% is used as a training set, and 20% is used as a validation set; 3-2: Training U 2 -Net deep learning model, and use the optimal model parameters to predict the test set. The test set IOU mean is 0.965, Precision mean is 0.986, and Recall mean is 0.

975. 3-3: Use the trained model to perform semantic segmentation and vectorization on buildings; 3-4: Use the Mosaic to New Raster tool in ArcGIS Pro software to stitch the semantic segmentation results into a raster image; 3-5: Use the Select By Attribute tool in ArcGIS Pro to extract records with a gridcode field value of 1 to obtain building vector results.

5. The method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning according to claim 1, characterized in that: The water body extraction and vectorization based on the water body index specifically include: 4-1: Calculate the NDWI index using the Band Arithmetic tool in ArcGIS Pro software; 4-2: The optimal segmentation threshold is determined to be 0.14 based on the frequency distribution histogram using the empirical threshold method; 4-3: Use the Raster Calculator tool in ArcGIS Pro software to binarize the NDWI image; 4-4: Use the Raster to Polygon tool in ArcGIS Pro software to convert the raster into shapefile vector data; 4-5: Use the Select By Attribute tool in ArcGIS Pro software to extract records with a gridcode field value of 1 to obtain water body vector results.

6. The method for identifying and extracting Aedes mosquito habitats based on remote sensing index and deep learning according to claim 1, characterized in that: The vegetation extraction and vectorization based on vegetation index specifically include: 5-1: Calculate the NDVI index using the Band Arithmetic tool in ArcGIS Pro software; 5-2: The optimal segmentation threshold is determined to be 0.38 by Otsu's method; 5-3: Use the Raster Calculator tool in ArcGIS Pro software to binarize the NDVI image; 5-4: Use the Raster to Polygon tool in ArcGIS Pro software to convert the raster into shapefile vector data; 5-5: Use the Select By Attribute tool in ArcGIS Pro software to extract records with a gridcode field value of 1 to obtain vegetation vector results.

Images