A farmland condition parameter inversion method fusing point-plot-region scale data

Abstract

The application discloses a kind of quantitative analysis method of agricultural condition parameter of fusion point-block-region scale data, improve the utilization rate of remote sensing image, realize the fine agricultural condition parameter quantitative analysis of less sample.The present application collects remote sensing image and agricultural condition parameter measured point data of different space-time resolution, carries out data preprocessing;Using gray level co-occurrence matrix extracts the texture feature of high-resolution image and carries out superposition;According to certain strategy, divide image with 256x256 row and column number, and scale up low-resolution image, make into data set;Using superimposed texture feature high-resolution image and scaled low-resolution image data set training SRGAN model, on the basis of this training weight, train high-resolution image without superimposed texture feature and scaled low-resolution image;Finally, the generated high-resolution image and agricultural condition information point data are fused to carry out random forest regression, obtain high spatial resolution surface scale agricultural condition information, and verify.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fine monitoring of crop growth parameters, and particularly relates to a crop growth parameter inversion method fusing point-plot-region scale data. BACKGROUND

[0002] Obtaining long-time series crop growth parameters from remote sensing data is an important data source for crop growth monitoring and yield estimation. Crop growth information parameters are very important structural parameters for monitoring crop growth and reflecting crop growth. Taking leaf area index (LAI) as an example, the change of its area affects the light received by crops and the microclimate of farmland, and it is one of the important indicators for evaluating the photosynthetic capacity of plant communities. In addition, normalized difference vegetation index (NDVI) is also a good indicator of crop growth.

[0003] There are mainly two methods for measuring crop growth parameters, direct measurement by field sampling and indirect measurement by non-destructive methods using measuring instruments and remote sensing technology. In remote sensing data, multi-spectral data and hyperspectral data are often used to extract crop growth parameters and study their spatial and temporal distribution. Remote sensing images can be used for large-area synchronous observation, saving cost and excluding human interference, and the data has comprehensiveness and comparability, but is easily affected by environmental factors, instrument and equipment errors, and changes in the photochemical process of vegetation. How to improve the monitoring accuracy of crop growth parameters has become a research hotspot and difficulty.

[0004] Different scales of remote sensing images have different resolutions and contain different amounts of information, and the accuracy of the obtained data also differs. Low-resolution images often have a serious mixed pixel problem, and high-resolution images are difficult to obtain and often have low temporal resolution. In order to improve the problems and effects caused by low-resolution images on information extraction, data fusion technology is often used to obtain remote sensing images with higher spatial and temporal resolution to obtain more information data. Traditional data fusion methods generally directly enlarge or reduce the whole image, and the actual effect is not very obvious. With the development of machine learning and other artificial intelligence algorithms, image fusion methods are more abundant and can enhance more detailed information.

[0005] The reconstruction of image data can effectively improve the accuracy of remote sensing image data acquisition. However, due to the diversity of remote sensing satellites and the different imaging principles and technical limitations of various sensors carried by the satellites, any single image source cannot fully reflect the characteristics of the target object. Taking a small field as an example, in order to realize the monitoring and estimation of crop growth at field scale, more detailed data are usually required. However, single remote sensing image often cannot meet the detailed requirements of actual situation. At present, there are few satellites with high temporal and spatial resolution, and the cost of data acquisition is high. In order to make the application more convenient, the spatio-temporal data fusion technology provides a solution for continuous time series crop growth monitoring at field scale. That is, a mathematical model is used to integrate multiple remote sensing images of the same area from different sensors into one image that meets the specific application requirements, combining the advantages of different remote sensing sensors to improve the temporal and spatial resolution of the image and achieve more accurate and reliable estimation and judgment of the target.

[0006] Patent document CN115170916A, published on October 11, 2022, discloses a multi-scale feature fusion image reconstruction method and system in the field of image reconstruction. The initial reconstruction image is constructed using the measurement vector generated by the original image and the residual modules are used for feature extraction to obtain a residual feature set. The extracted features are input into the dense module of multiple scale convolution kernels to extract dense features, and the global residual feature fusion is obtained by repeated iteration. The residual of the global fusion feature is calculated and added to the initial reconstruction image to obtain the final reconstruction image. The invention improves the reconstruction quality and reduces the calculation amount.

[0007] Patent document CN1677085A, published on October 5, 2005, discloses an agricultural application integrated system and method of earth observation technology. The invention uses advanced earth observation technology to accurately manage agriculture. First, high-spectral data are obtained and the obtained high-spectral data are preprocessed. The high-spectral data are selected by waveband. The high-spectral data selected by waveband are extracted for feature extraction to obtain agricultural condition parameters. The obtained agricultural condition parameters are used to output related agricultural information in the high-spectral data.

[0008] However, there are some problems in current research:

[0009] High temporal and spatial resolution remote sensing images are difficult to obtain and have high acquisition cost. The temporal and spatial resolution of easily obtainable images is low, the ground object contour information is fuzzy, the information recognition effect is poor, and the data accuracy is difficult to meet the actual production needs. The method of using easily obtainable remote sensing images to retrieve more detailed images needs to be explored.

[0010] The data fusion algorithm has the disadvantages of unstable model training, large number of parameters, slow model convergence speed, etc. The reconstruction effect is difficult to meet the accuracy requirements, and the process data analysis and calculation amount is large.

[0011] Different scale data resolution, different information quantity, small scale research area is not effectively combined with multi-scale data to realize the application of remote sensing data in the field, improve the utilization rate of remote sensing image. SUMMARY

[0012] The present application provides a kind of point-farmland-region scale data fusion agricultural condition parameter inversion method for the defects in the prior art, to solve the above technical problems, the present application provides the following technical scheme:

[0013] A kind of point-farmland-region scale data fusion agricultural condition parameter inversion method, comprising:

[0014] Collect two different spatio-temporal resolution remote sensing images and agricultural condition information point data, and carry out data preprocessing;

[0015] Extract the texture features of high spatial resolution remote sensing image, and superimpose it;

[0016] Strategy image segmentation is carried out to each image set image, and training set, verification set and test set for image fusion are prepared;

[0017] The fusion model of different spatial resolution remote sensing images is constructed, the step-by-step training and fine-tuning of the model are carried out using the transfer learning strategy, the final weight of the model is obtained, and the trained image fusion model is established;

[0018] The point-face fusion model is constructed, the model is trained using the generated high-resolution image and agricultural condition information point data, and the trained point-face fusion model is established;

[0019] The inverted agricultural condition information is evaluated in multiple scales.

[0020] Optionally, the collection of two different spatio-temporal resolution remote sensing images and agricultural condition information point data, and the data preprocessing, comprises:

[0021] Collect remote sensing images of different resolutions corresponding to spatial location information;

[0022] Preprocessing of multi-scale remote sensing data set: geometric correction, atmospheric correction, radiation calibration, orthorectification and spatial registration;

[0023] Resample low spatial resolution image to be consistent with high spatial resolution image.

[0024] Optionally, the extraction of texture features of high spatial resolution remote sensing image, and superimposition thereof, comprises:

[0025] The texture features of high spatial resolution image are extracted using gray level co-occurrence matrix method and superimposed into the image to deepen the contour information of the image;

[0026] By calculating the gray image to obtain its co-occurrence matrix, then calculate the co-occurrence matrix to obtain the partial eigenvalue of the matrix, to represent some texture features of the image respectively, the commonly used texture feature statistical attributes are: mean, standard deviation, homogeneity / inverse difference moment, contrast, dissimilarity, entropy, angular second moment / energy, maximum probability, compare the extraction effect of each statistical attribute, select the best statistical attribute to extract the texture features of high spatial resolution image.

[0027] Optionally, the strategy image segmentation of each image set image, the training set, the verification set, the test set for image fusion are prepared, including:

[0028] The resampled low resolution image set, the high resolution image set superimposed with texture features and the original high resolution image set are respectively cropped to 256*256 size, and a certain repetition rate needs to be set at the four edges of the image during cropping, wherein the repetition rate of the upper edge and the left edge is 10%, and the repetition rate of the lower edge and the right edge is 5%, after cropping, the low spatial resolution image is up-scaled by 4*4, that is, the mean value of every 16 pixels is taken, to convert the row and column number from 256*256 to 64*64, and finally the training set, the verification set and the test set are divided according to 7:3:1.

[0029] Optionally, the fusion model of remote sensing images with different spatial resolutions is constructed, the model is trained and fine-tuned in steps using the transfer learning strategy, the final weight of the model is obtained, and the trained image fusion model is established, including:

[0030] The image super-resolution reconstruction model SRGAN based on the generative adversarial network, the model is composed of a generator and a discriminator, the generator structure in the model is composed of residual blocks, skip layers and convolution layers, and the loss function is optimized as the target;

[0031] The loss function of the generator:

[0032]

[0033] The loss function of the discriminator:

[0034]

[0035] The loss function method of the model is composed of the weighted sum of the content loss and the adversarial loss. Its formula definition is:

[0036]

[0037] Among them, the pixel-based MSE loss in the content loss is defined as:

[0038]

[0039] The formula definition of the adversarial loss is:

[0040]

[0041] First, the SRGAN model is trained using high-resolution images with texture features and low-resolution image data sets after upsampling. Then, based on the pre-trained weights, the high-resolution images without texture features and the low-resolution images after upsampling are further migrated and trained. Finally, the precision of the model is verified using the original high-resolution image and the low-resolution image after upsampling. The SRGAN model with optimal training weights is input into the test set of the low-resolution image after upsampling to generate high-resolution images with spatial resolution consistent with the high-resolution images.

[0042] Optionally, the point-face fusion model is constructed using the generated high-resolution images and agricultural information point data to train the model and establish the trained point-face fusion model, including:

[0043] The random forest regression model is used to fuse the generated high-resolution remote sensing images with agricultural information point data to obtain high spatial resolution field-scale agricultural information, and realize multi-scale data fusion of agricultural information inversion.

[0044] Optionally, the multi-scale effect of the inverted agricultural information is evaluated, including:

[0045] The R 2 The RMSE is used to verify the point-scale agricultural parameter inversion effect, and the linear trend method and the percentage change method are used to analyze the change trend and stability of the agricultural parameter inversion results using high-resolution images and generated high-resolution images in the face scale.

[0046] The formula of the percentage change method is:

[0047]

[0048] Wherein, The percentage change value of a certain agricultural parameter to be monitored at a certain spatial location, Index is the name of the agricultural parameter, high_Value Index The agricultural parameter value at a certain spatial location inverted using the original high-resolution image, generate_Value Index The agricultural parameter value at a corresponding spatial location inverted using the generated high-resolution image from the low-resolution image. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on the provided drawings.

[0050] Figure 1 A flow chart of the agricultural condition parameter inversion method fusing point-plot-area scale data;

[0051] Figure 2 A technical framework diagram of the agricultural condition parameter inversion method fusing point-plot-area scale data;

[0052] Figure 3 Image texture features extracted using the gray level co-occurrence matrix method;

[0053] Figure 4 A schematic diagram of strategic cutting and splicing of remote sensing images;

[0054] Figure 5 A high-resolution image after reconstruction of a low-resolution image;

[0055] Figure 6 A leaf area index after fine inversion of the reconstructed high-resolution image;

[0056] Figure 7 A percentage change spatial distribution diagram of the inversion results of the reconstructed image and the original high-resolution image. DETAILED DESCRIPTION

[0057] 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 only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort fall within the scope of the present application.

[0058] The terms "first", "second", and the like in the specification and claims of the present application and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the terms used in this way can be interchanged, and this is only a distinguishing way used in the description of the embodiments of the present application to describe the objects with the same attributes. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, so that the processes, methods, systems, products or equipment containing a series of units do not have to be limited to those units, but can include other units not clearly listed or inherent to these processes, methods, products or equipment.

[0059] The application provides a farmland parameter inversion method for fusing point-plot-region scale data. Figures 1-7 The application provides a farmland parameter inversion method for fusing point-plot-region scale data.

[0060] Figure 1 The application provides a farmland parameter inversion method for fusing point-plot-region scale data, Figure 2 The application provides a farmland parameter inversion method for fusing point-plot-region scale data, and the technical framework of the application is as follows, Figure 1 The application provides a farmland parameter inversion method for fusing point-plot-region scale data, and the technical framework of the application is as follows, Figure 2 The application provides a farmland parameter inversion method for fusing point-plot-region scale data, and the technical framework of the application is as follows,

[0061] Step 1: Collecting two kinds of remote sensing images with different spatio-temporal resolutions and farmland information point data, and performing data preprocessing.

[0062] In order to obtain satellite images with different spatio-temporal resolutions in the target region, Worldview-2 sub-meter satellite images or other high-resolution images and Sentinel-2 satellite images provided by the European Space Agency can be used. In order to ensure that the remote sensing images of different sensors have a small time difference, the remote sensing images with different spatio-temporal resolutions corresponding to the target research area and having a time difference of no more than 5 days and spatial positions are selected for downloading.

[0063] For farmland information point data, it can include but is not limited to leaf area index, leaf chlorophyll content, leaf nitrogen, etc. The acquisition method can be field collection or field experiment, or it can be obtained by a field farmland sensor. However, it should be noted that the data collection time should be consistent with the high-resolution remote sensing image.

[0064] After obtaining remote sensing images with different spatio-temporal resolutions and farmland point data, the point and surface data can be preprocessed. For remote sensing data, geometric correction, atmospheric correction, radiation calibration, orthorectification and spatial registration can be performed. The point data needs to be processed for missing values, and the sensor data also needs to be processed for thinning.

[0065] Step 2: Extracting texture features of high spatial resolution remote sensing images and superimposing them.

[0066] The texture features of Worldview-2 satellite images are extracted by using the gray level co-occurrence matrix method and superimposed on the images to deepen the contour information of the images. The co-occurrence matrix of the gray image is obtained by calculating the gray image, and then the partial eigenvalues of the matrix are calculated to represent some texture features of the image. Commonly used texture feature statistical attributes include: mean, standard deviation, homogeneity / inverse difference moment, contrast, dissimilarity, entropy, angular second moment / energy, maximum probability, etc. Figure 3The extraction effects of various statistical attributes are compared, and the best statistical attribute is selected for texture feature extraction of high spatial resolution images.

[0067] The texture features of the Worldview-2 image are superimposed on the high-resolution image to obtain a high-resolution image superimposed with texture features.

[0068] Step 3: Strategically segment each image set image to prepare a training set, a validation set, and a test set for image fusion.

[0069] The resampled Sentinel-2 image set, the Worldview-2 image set superimposed with texture features, and the original Worldview-2 image set are respectively cropped to 256*256 in size. When cropping, a certain repetition rate needs to be set at the four edges of the image, such as Figure 4 As shown, the repetition rate of the upper and left edges is 10%, and the repetition rate of the lower and right edges is 5%;

[0070] After cropping, the low spatial resolution image is up-scaled by 4*4, i.e., taking the mean value of every 16 pixels, to convert the row and column number from 256*256 to 64*64.

[0071] The Sentinel-2 image set, the Worldview-2 image set superimposed with texture features, and the original Worldview-2 image set are divided into a training set, a validation set, and a test set in a 7:3:1 ratio according to their spatial positions.

[0072] Step 4, build a fusion model of remote sensing images with different spatial resolutions, use the transfer learning strategy to train and fine-tune the model in steps, obtain the best weight of the model, and establish the trained image fusion model.

[0073] The SRGAN model is constructed to include a generator and a discriminator. The generator structure is composed of residual blocks, skip layers, and convolutional layers, with the optimization loss function as the target.

[0074] The loss function of the generator:

[0075]

[0076] The loss function of the discriminator:

[0077]

[0078] The loss function method of the model is composed of the weighted sum of the content loss and the adversarial loss. Its formula definition is:

[0079]

[0080] Among the content loss, the pixel-based MSE loss is defined as:

[0081]

[0082] The formula definition of the adversarial loss is:

[0083]

[0084] The SRGAN model is pre-trained using the training set of Worldview-2 images superimposed with texture features and the up-scaled Sentinel-2 images, and the training weights are saved.

[0085] The SRGAN model is further trained using the original Worldview-2 images and the up-scaled Sentinel-2 images as the training set based on the pre-trained weights.

[0086] The model is verified for accuracy using the original Worldview-2 images and the up-scaled Sentinel-2 images as the validation set.

[0087] The SRGAN model with the optimal training weights is inputted with the up-scaled Sentinel-2 test set to generate high-resolution images consistent with the spatial resolution of Worldview-2, as shown in Figure 5

[0088] High-resolution images after reconstruction of the Sentinel-2 test set images.

[0089] Step 5: Build a point-to-area fusion model using the generated high-resolution images and agricultural information point data to train the model and establish the trained point-to-area fusion model.

[0090] A random forest regression model is built for the fusion of remote sensing images and agricultural information point data.

[0091] The random forest regression model is trained using the original Worldview-2 image training set and the corresponding regional agricultural point data (leaf area index) training set.

[0092] Step 6: Perform multi-scale effect evaluation on the inverted agricultural information.

[0093] The model is verified for accuracy using the original Worldview-2 image validation set and the corresponding regional agricultural point data (leaf area index) validation set, and the R 2 and RMSE are used to verify the point-scale agricultural parameter inversion effect.

[0094] ​The trained random forest regression model is inputted with the test set of the original Worldview-2 image to generate a fine leaf area index spatial distribution map corresponding to the region of the image.

[0095] The trained random forest regression model is inputted with the test set of the generated high-resolution image to generate a fine leaf area index spatial distribution map corresponding to the region of the generated high-resolution image, and the result is as shown in Figure 6

[0096] The linear trend method and the percentage change method are used to analyze the variation trend and stability of the area scale of the agricultural condition parameter inversion results obtained by using the original Worldview-2 and the generated high-resolution image respectively, and the result is as shown in Figure 7

[0097] The formula of the percentage change method is as follows:

[0098]

[0099] Among them, is the percentage change value of a certain agricultural condition parameter to be monitored at a certain spatial position, Index is the name of the agricultural condition parameter, high_Value Index is the agricultural condition parameter value at a certain spatial position inverted by using the original Worldview-2, generate_Value Index is the agricultural condition parameter value at a corresponding spatial position inverted by using the high-resolution image generated by the Sentinel-2 image.

[0100] In summary, the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit it. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the above embodiments, or make equivalent replacement to part of the technical features. The modification or replacement does not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solutions of the embodiments of the present application.​​

Claims

1. A method for fusing point-plot-region scale data to retrieve agro-ecological parameters, characterized in that, The application relates to a method for multi-scale remote sensing data fusion of agricultural information, comprising the following steps: Collecting remote sensing images and agricultural information point data of two different space-time resolutions and performing data preprocessing; Extracting texture features of high spatial resolution remote sensing images and superimposing the texture features on the images; Strategically segmenting the images of each image set to prepare a training set, a verification set and a test set for image fusion; Constructing a remote sensing image fusion model of different spatial resolutions, using a transfer learning strategy to perform step-by-step training and fine-tuning of the model, obtaining the optimal weight of the model, and establishing a trained image fusion model; the fusion model is a super-resolution reconstruction model SRGAN based on a generative adversarial network, the model is composed of a generator and a discriminator, and the generator structure in the model is composed of a residual block, a jump layer and a convolution layer; first, the SRGAN model is trained by using the high-resolution image superimposed with the texture features and the low-resolution image data set after upsampling, then the high-resolution image without superimposed texture features and the low-resolution image after upsampling are further transferred and trained on the basis of the pre-training weight, finally, the model is verified in precision by using the original high-resolution image and the low-resolution image after upsampling as the verification set, and the SRGAN model with the optimal training weight is input into the low-resolution image after upsampling as the test set to generate a high-resolution image with the same spatial resolution as the high-resolution image; Constructing a point-surface fusion model, training the model by using the generated high-resolution image and the agricultural information point data, and establishing a trained point-surface fusion model; the point-surface fusion model is a random forest regression model, which is used for point-surface data fusion of the high-resolution remote sensing image generated by fusion and the agricultural information point data to obtain high spatial resolution field-scale agricultural information, realizes multi-scale data fusion of the agricultural information, and reverses the agricultural information; Multi-scale effect evaluation is performed on the reversed agricultural information.

2. The method of claim 1, wherein, The collection of remote sensing images and agricultural information point data of two different space-time resolutions and the data preprocessing comprise the following steps: Collecting remote sensing images of different resolutions corresponding to spatial position information; Preprocessing of the multi-scale remote sensing data set: geometric correction, atmospheric correction, radiation calibration, orthographic correction and spatial registration; Resampling the low spatial resolution image to be consistent with the high spatial resolution image.

3. The method of claim 1, wherein, The extraction of texture features of the high spatial resolution remote sensing image and the superimposition thereof comprise the following steps: Using a gray level co-occurrence matrix method to extract the texture features of the high spatial resolution image and superimpose the texture features on the image to deepen the contour information of the image; The co-occurrence matrix of the gray image is obtained by calculation, then the partial eigenvalues of the co-occurrence matrix are calculated to represent certain texture features of the image, and the extraction effects of commonly used texture feature statistical attributes are compared, and the best statistical attribute is selected for the extraction of the texture features of the high spatial resolution image.

4. The method of claim 1, wherein, The strategic image segmentation of the images of each image set to prepare a training set, a verification set and a test set for image fusion comprises the following steps: The resampled low-resolution image set, the high-resolution image set superimposed with the texture features and the original high-resolution image set are respectively cropped to a size of 256*256, and a certain repetition rate needs to be set at the four edges of the image during the cropping, wherein the repetition rate of the upper edge and the left edge is 10%, and the repetition rate of the lower edge and the right edge is 5%. After the cropping, the low spatial resolution image is up-scaled by 4*4, that is, the mean value is taken every 16 pixels, to convert the row and column number from 256*256 to 64*64, and finally the training set, the verification set and the test set are divided according to 7:3:

1.

5. The method of claim 1, wherein, The fusion model of remote sensing images with different spatial resolutions is constructed, the model is trained and fine-tuned in steps using the transfer learning strategy, the most powerful weight of the model is obtained, and the trained image fusion model is established, including: The SRGAN model takes the optimization loss function as the target; The loss function of the generator: The loss function of the discriminator: The loss function method of the model is composed of the weighted sum of the content loss and the adversarial loss, and the formula definition is: In the content loss, the pixel-based MSE loss is defined as: The formula definition of the adversarial loss is: 。 6. The method of claim 1, wherein, The multi-scale effect evaluation of the inversed agricultural information is performed, including: R 2 The point scale validation of the agricultural parameter inversion effect with RMSE, the linear trend method and the percentage change method are used to analyze the change trend and stability of the area scale of the agricultural parameter inversion results using high resolution images and generated high resolution images respectively. The percentage change method formula is: wherein, is the percentage change value of a certain agricultural parameter at a certain spatial location, Index is the name of the agricultural parameter, high_Value Index is the value of the agricultural parameter at a certain spatial location, which is derived from the original high-resolution image, generate_Value Index is the value of the agricultural parameter at a corresponding spatial location, which is derived from the high-resolution image generated using the low-resolution image.

Images