Unmanned aerial vehicle multispectral image processing and corn biomass estimation method

Abstract

The invention provides an unmanned aerial vehicle multi-spectral image processing and corn biomass estimation method, which comprises the following steps of: splicing collected images to obtain images of various reflectivity wavebands, cutting according to the boundaries of the plots to obtain the image of each plot, and segmenting vegetation and soil by adopting a green normalized vegetation index threshold segmentation method, the method comprises the following steps of: extracting mean values of various parameters of a vegetation region, including wave band reflectivity, vegetation indexes, texture features and the like, respectively taking the mean values as input values, establishing an AGB estimation model, and meanwhile, extracting the mean values of various parameters of the whole plot as input parameters of the AGB model for comparison; according to the method, accurate estimation of the corn AGB under soil background interference can be realized.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing technology, and in particular relates to a method for processing multispectral images from unmanned aerial vehicles (UAVs) and estimating maize biomass. Background Technology

[0002] Aboveground biomass (AGB) plays a crucial role in light energy utilization and is a vital foundation for crop yield formation. Traditional methods for measuring AGB involve manual sampling and weighing, which is labor-intensive, inefficient, and destructive. Current research often estimates AGB using spectral and textural features from UAV multispectral images. However, the extracted parameters typically represent features from the entire image, including non-vegetated areas. This limits the accuracy of AGB estimation, especially for sparsely vegetated areas with significant bare soil. Summary of the Invention

[0003] Technical Solution: To solve the above-mentioned technical problems, this invention first segments the vegetation and soil in the image to obtain the vegetation region before extracting relevant features, and then extracts relevant parameters of the vegetation region to further estimate the aboveground biomass (AGB). Specifically, it proposes a method for estimating maize biomass based on UAV multispectral image processing and the following steps:

[0004] Step 1: Data Acquisition and Preprocessing

[0005] Collect reflectivity bands corresponding to drone flight images and real field data;

[0006] Step 2: Calculate the vegetation index

[0007] The corresponding vegetation index is obtained by performing relevant band calculations based on the initial reflectance bands in step one.

[0008] Step 3: Extract texture features

[0009] In remote sensing image processing platforms, such as ENVI, the gray-level co-occurrence matrix method is used to extract texture features of five reflectance bands, and the corresponding texture features are obtained for each band.

[0010] Step 4: Process the image

[0011] Images of each plot were obtained by cropping the plot boundaries. All experimental plot images were processed, and vegetation-soil segmentation was performed using the green normalized vegetation index (NRVI). A threshold ε was set. v Pixel value ≥ ε v The pixel is a vegetation pixel, value < ε v The pixels are soil pixels;

[0012] Step 5: Extract each parameter

[0013] For each plot, the pixels of the vegetation mask region Mask=1 in all parameter images are extracted, the average value of each parameter in the mask region is calculated, and the average value of each parameter of all pixels in the unsegmented whole plot is directly calculated as a control. According to the variable type, they are used as input parameters for estimating aboveground biomass AGB.

[0014] Step Six: Selection of Feature Parameters

[0015] Based on the Pearson correlation coefficient and the criteria of high correlation between fresh biomass and dry biomass, multiple characteristic parameters under different variable types were selected. The Pearson correlation coefficients (r) between all parameters and biomass were calculated and arranged in descending order as characteristic parameters. A threshold for the number of characteristic parameters was set.

[0016]

[0017] Where, x i This represents a parameter value for the i-th plot. This represents the average value of this parameter across all plots, y. i This represents the actual biomass value of the i-th plot. This represents the average of the actual biomass values ​​across all plots.

[0018] Step 7: Construction of the biomass estimation model

[0019] The feature parameters extracted from the entire plot without segmentation were used as independent variables, and the fresh biomass and dry biomass were used as dependent variables. The training set and validation set were divided using the five-fold cross-validation method. An estimation model was established using two machine learning algorithms: Random Forest (RF) and Partial Least Squares Regression (PLSR). The results of estimating the fresh biomass and dry biomass after different image processing parameters were output.

[0020] As an improvement, the specific steps in step one are as follows:

[0021] (1.1) UAV data: First, the UAV is planned to fly, then the UAV images are collected, and finally the images collected by the UAV are stitched together to obtain five reflectivity bands, including RED, GREEN, BLUE, REDEDGE and NIR images.

[0022] (1.2) Field data: Three corn plants were selected from each plot. The destructive sampling method was used to collect the corn plants, put them into marked sealed bags, and immediately transported them to the laboratory to weigh their fresh weight. After drying until the mass was constant, their dry weight was weighed. The fresh biomass and dry biomass were calculated based on the planting density.

[0023] As an improvement, step two includes 14 vegetation indices, specifically: green chlorophyll index, red-edged chlorophyll index, chlorophyll vegetation index, differential vegetation index, enhanced vegetation index, overgreen index, vegetation cover index, green normalized vegetation index, green ratio vegetation index, optimized soil-regulating vegetation index, normalized vegetation index, normalized differential red-edged index, renormalized vegetation index, and ratio vegetation index.

[0024] As an improvement, in step three, each band obtains eight corresponding texture features.

[0025] As an improvement, in step four, the normalized green vegetation index (GNDVI) is used to segment vegetation and soil, and the GNDVI is calculated using the NIR and GREEN bands.

[0026] GREEN=(NIR-GREEN) / (NIR+GREEN)

[0027] First, the GNDVI image is processed using a reclassification operation, selecting "Natural Discontinuity Grading" as the classification method. Then, the number of categories is set, and finally, the threshold ε is obtained. v Pixels with values ​​greater than a threshold are classified as vegetation pixels, and pixels with values ​​less than a threshold are classified as soil pixels. The resulting binary mask image is then classified.

[0028]

[0029] Where x represents the width of the image, y represents the length of the image, value(i,j) represents the pixel value in the i-th row and j-th column of the GNDVI image, and Mask(i,j) represents the pixel value in the i-th row and j-th column of the vegetation and soil mask.

[0030] As an improvement, the variable types in step five include initial reflectance band, vegetation index, texture features, and initial reflectance + vegetation index + texture features; in step six, since the number of initial reflectance variables is relatively small, feature selection is not performed. However, since the number of vegetation index, texture features, and initial band reflectance + vegetation index + texture features is relatively large, the top n variables that are highly correlated with fresh biomass and dry biomass are selected as feature parameters, where n is a positive integer, preferably 10.

[0031] As an improvement, methods for evaluating the accuracy of the estimation model are also included, specifically: when constructing crop phenotypic parameter estimation models, the commonly used coefficient of determination R is calculated. 2 The root mean square error (RMSE) is used to evaluate model performance. 2 The closer the value is to 1 and the smaller the RMSE, the better the model performs.

[0032]

[0033] Where n is the number of samples, y i It is the true value of the i-th sample. It is the estimated value of the i-th sample. It is the average of the actual observed values.

[0034] Beneficial Effects: The UAV multispectral image processing and maize biomass estimation method proposed in this invention first stitches together the acquired images to obtain images of various reflectance bands. Then, it crops the images according to the plot boundaries to obtain images of each plot. The vegetation-soil segmentation is performed using the green normalized vegetation index. The mean values ​​of various parameters (band reflectance, vegetation index, and texture features) of the vegetation area are extracted and used as input values ​​to establish an AGB estimation model. The AGB estimation model is established using the extracted parameters of the entire plot as input parameters as a control, and finally, the aboveground biomass is accurately estimated under soil background interference.

[0035] This invention predicts biomass by using parameters of vegetated areas after image segmentation. This can reduce the impact of non-vegetated areas (such as bare soil), improve the accuracy of biomass estimation, and achieve more accurate results through simple operations. Furthermore, this method is applicable to both dense and sparse vegetation. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention.

[0037] Figure 2 This is the result of vegetation and soil segmentation in Embodiment 1 of the present invention.

[0038] Figure 3 This is a visualization of Pearson correlation in Embodiment 1 of the present invention.

[0039] Figure 4 This is a diagram showing the predicted AGB effect in Embodiment 1 of the present invention. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] Step 1: Data Acquisition and Preprocessing

[0043] Plan the flight in DJIGSPRO software, set the flight altitude to 80m, use an S-shaped scanning mode, and point the camera vertically downwards to capture drone images. Stitch the captured drone images in DJI Terra mapping software to obtain reflectivity images for five bands: RED, GREEN, BLUE, REDEDGE, and NIR.

[0044] Three corn plants were selected from each plot. Using a destructive sampling method, the collected corn was placed in a marked sealed bag and immediately transported to the laboratory to be weighed for fresh weight. After drying until the mass was constant, the dry weight was measured. The fresh biomass and dry biomass were calculated based on the planting density.

[0045] Step 2: Calculation of Vegetation Index

[0046] By performing band calculations on five reflectance bands (RED, GREEN, BLUE, REDEDGE, NIR), 14 vegetation indices were obtained: CIG, CIRE, CVI, DVI, EVI, ExG, LCI, GNDVI, GRVI, OSAVI, NDVI, NDRE, RDVI, and RVI, as shown in Table 3.

[0047] Table 3: Calculation of Vegetation Index

[0048]

[0049] Step 3: Extraction of texture features

[0050] The gray-level co-occurrence matrix (GLCM) method was used in ENVI software to calculate the texture features of reflectance images. Specifically, the sliding window size of GLCM was set to 3×3 pixels, the step size to 1, and the gray-level quantization levels to 64. Eight texture features were obtained for each reflectance image, including mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation, resulting in a total of 40 texture features.

[0051] Step 4: Vegetation and Soil Segmentation

[0052] Vegetation and soil are separated by setting a threshold using the Normalized Green Vegetation Index (GNDVI). Optionally, a reclassification operation is performed on the GNDVI image in ArcGIS Pro, selecting "Natural Discontinuity Classification" as the classification method, and setting the number of categories. In this example, it is set to 2, resulting in a threshold ε. vPixels with values ​​greater than a threshold are classified as vegetation pixels, and pixels with values ​​less than a threshold are classified as soil pixels. The resulting binary mask image is shown below. Figure 2 As shown.

[0053]

[0054] Where x represents the width of the image, y represents the length of the image, value(i,j) represents the pixel value in the i-th row and j-th column of the GNDVI image, and Mask(i,j) represents the pixel value in the i-th row and j-th column of the vegetation and soil mask.

[0055] Step 5: Extract parameters

[0056] For each plot, the average value of each parameter within the vegetation mask area of ​​all parameter images is extracted. In addition, the average value of all parameters within the entire plot area (all pixels) is directly extracted as a control.

[0057] Step Six: Pearson Correlation Calculation

[0058] The Pearson correlation coefficient r for all parameters and biomass was calculated using Pearson correlation calculation. The parameters were then arranged from largest to smallest as characteristic parameters. A threshold for the number of characteristic parameters was set. The formula for calculating the Pearson correlation coefficient r is as follows:

[0059]

[0060] Where, x i This represents a parameter value for the i-th plot. This represents the average value of this parameter across all plots, y. i This represents the actual biomass value of the i-th plot. This represents the average of the actual biomass values ​​across all plots.

[0061] Calculate the correlation between all variables and fresh biomass and dry biomass in Python, and visualize the results as follows: Figure 3 As shown, preferably, in this invention, the top ten variables with the highest Pearson correlation are selected as feature variables, and the filtered feature variables are shown in Table 4.

[0062] Table 4: Feature Variables

[0063]

[0064]

[0065] Step 7: Construction of the AGB estimation model

[0066] The selected feature variables were used as input parameters, and fresh biomass and dry biomass were used as output parameters. A five-fold cross-validation method was employed to partition the data. Two machine learning algorithms, Random Forest (RF) and Partial Least Squares Regression (PLSR), were used to build AGB estimation models. The effectiveness of different image processing methods in estimating fresh and dry biomass was compared for each parameter extracted. Figure 4 As shown.

[0067] See Figure 4 As shown, image segmentation significantly improves estimation accuracy: compared to directly extracting all feature parameters from the entire plot, performing vegetation and soil segmentation (based on GNDVI) before extracting feature parameters effectively reduces soil background interference and significantly improves AGB estimation accuracy. The PLSR model shows a particularly significant improvement in accuracy (R...). 2 The highest increase was 0.543.

[0068] Example 2

[0069] The results of estimating biomass using different image processing methods and parameter combinations are compared, as shown in Tables 5 and 6.

[0070] Table 5: Comparison of Estimated Fresh Biomass Results

[0071]

[0072]

[0073] See Table 5: RBs, VIs, TFs, and RBs+VIs+TFs represent different parameter combinations, no_Segement means no segmentation, and Segement means the image has been segmented.

[0074] Table 6: Comparison of estimated dry biomass results

[0075]

[0076] Comparison revealed that, regardless of whether it is fresh or dry biomass, regardless of the combination of variables used, and regardless of whether the machine learning algorithm is RF or PLSR, extracting parameters from the vegetation area to estimate maize biomass yields better results in general, and image segmentation can improve the accuracy of biomass estimation.

[0077] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for multispectral image processing and maize biomass estimation using unmanned aerial vehicles (UAVs), characterized in that: The steps of the method are as follows: Step 1: Data Acquisition and Preprocessing Collect drone flight images and process them to obtain the corresponding reflectivity bands and collect real field data; Step 2: Calculate the vegetation index The corresponding vegetation index is obtained by performing relevant band calculations based on the initial reflectance bands in step one. Step 3: Extract texture features The gray-level co-occurrence matrix method was used in the remote sensing image processing platform to extract the texture features of five reflectance bands, and the corresponding texture features were obtained for each band. Step 4: Process the image Images of each plot were obtained by cropping the plot boundaries. All experimental plot images were processed, and vegetation-soil segmentation was performed using the green normalized vegetation index (NRVI). A threshold ε was set. v Pixel value ≥ ε v The pixels are vegetation pixels, value < ε v The pixels are soil pixels; Step 5: Extract each parameter For each plot, the pixels of the vegetation mask region Mask=1 in all parameter images are extracted, the average value of each parameter in the mask region is calculated, and the average value of each parameter of all pixels in the unsegmented whole plot is directly calculated as a control. According to the variable type, they are used as input parameters for estimating aboveground biomass AGB. Step Six: Selection of Feature Parameters Based on the Pearson correlation coefficient, multiple characteristic parameters were selected for different variable types with high correlation to fresh biomass and dry biomass. The Pearson correlation coefficients (r) between all parameters and biomass were calculated (see Formula 1). These parameters were arranged from largest to smallest as characteristic parameters. A threshold for the number of characteristic parameters was set. Where, x i This represents a parameter value for the i-th plot. This represents the average value of this parameter across all plots, y. i This represents the actual biomass value of the i-th plot. This represents the average of the actual biomass values ​​across all plots. Step 7: Construction of the biomass estimation model The feature parameters of the vegetation area were extracted after the entire plot was divided, and the feature parameters extracted without division were used as independent variables, while fresh biomass and dry biomass were used as dependent variables. The training set and validation set were divided using the five-fold cross-validation method. The estimation model was established by two machine learning algorithms, Random Forest (RF) and Partial Least Squares Regression (PLSR), and the results of estimating fresh biomass and dry biomass after different image processing were output.

2. The method for UAV multispectral image processing and maize biomass estimation according to claim 1, characterized in that: The specific steps in step one are as follows: (1.1) UAV data: First, the UAV is planned to fly, then the UAV images are collected, and finally the images collected by the UAV are stitched together to obtain five reflectivity bands, including RED, GREEN, BLUE, REDEDGE and NIR images. (1.2) Field data: Three corn plants were selected from each plot. The destructive sampling method was used to collect the corn plants, put them into marked sealed bags, and immediately transported them to the laboratory to weigh their fresh weight. After drying until the mass was constant, their dry weight was weighed. The fresh biomass and dry biomass were calculated based on the planting density.

3. The method for UAV multispectral image processing and maize biomass estimation according to claim 1, characterized in that: Step two includes 14 vegetation indices, specifically: green chlorophyll index, red-edged chlorophyll index, chlorophyll vegetation index, differential vegetation index, enhanced vegetation index, overgreen index, vegetation cover index, green normalized vegetation index, green ratio vegetation index, soil-optimized vegetation index, normalized vegetation index, normalized differential red-edged index, renormalized vegetation index, and ratio vegetation index.

4. The method for UAV multispectral image processing and maize biomass estimation according to claim 1, characterized in that: In step three, each band yields eight corresponding texture features.

5. The method for UAV multispectral image processing and maize biomass estimation according to claim 1, characterized in that: In step four, the normalized green vegetation index (GNDVI) is used to segment vegetation and soil, and the GNDVI is calculated using the NIR and GREEN bands. GREEN=(NIR-GREEN) / (NIR+GREEN) First, the GNDVI image is processed using a reclassification operation, selecting "natural discontinuity classification" as the classification method. Then, the number of categories is set, and finally, the threshold ε is obtained. v Pixels with values ​​greater than a threshold are classified as vegetation pixels, and pixels with values ​​less than a threshold are classified as soil pixels. The resulting binary mask image is then classified. Where x represents the width of the image, y represents the length of the image, value(i,j) represents the pixel value in the i-th row and j-th column of the GNDVI image, and Mask(i,j) represents the pixel value in the i-th row and j-th column of the vegetation and soil mask.

6. The method for estimating maize biomass based on UAV multispectral image processing according to claim 1, characterized in that: In step five, the variable types include initial reflectance band, vegetation index, texture features, and initial reflectance + vegetation index + texture features. In step six, since the number of initial reflectance variables is relatively small, no feature selection is performed. However, since the number of vegetation index, texture features, and initial band reflectance + vegetation index + texture features is relatively large, the top n variables that are highly correlated with fresh biomass and dry biomass are selected as feature parameters, where n is a positive integer.

7. The method for UAV multispectral image processing and maize biomass estimation according to claim 1, characterized in that: It also includes methods for evaluating the accuracy of the estimation models, specifically: when constructing crop phenotypic parameter estimation models, calculating the commonly used coefficient of determination R. 2 The root mean square error (RMSE) is used to evaluate model performance. 2 The closer the value is to 1 and the smaller the RMSE, the better the model performs. Where n is the number of samples, y i It is the true value of the i-th sample. It is the estimated value of the i-th sample. It is the average of the actual observed values.

Images