Highway weed landscape optimization and corn yield increase remote sensing machine learning method
Patent Information
- Application Number
- CN202611191595.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies struggle to quantify the barrier effect of grassland spatial structure on weed spread from a landscape pattern perspective, leading to weed invasion of cornfields, reduced corn yields, high weed control costs, and significant pollution risks from chemical herbicides. Furthermore, there is a lack of precise control methods that couple multiple models.
We used GEE cloud platform to identify wasteland from multi-source remote sensing images, combined with random forest regression and XGBoost models to construct a nonlinear response model, optimize the landscape structure parameters of wasteland, accurately quantify the weed inhibition index through remote sensing machine learning methods, output optimized configuration schemes to block weed spread and increase maize yield.
It achieves low-cost, low-pollution weed suppression, significantly increases maize yield, optimizes the landscape structure of wasteland, reduces the risk of soil pollution from chemical weeding, adapts to diverse environmental conditions, and provides reliable prediction capabilities.
Smart Images
Figure CN123033002A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of landscape optimization and synergistic regulation of maize yield increase, and particularly relates to a remote sensing machine learning method for landscape optimization and weed control in highway wasteland and maize yield increase. Background Technology
[0002] Along the highway, farmland and wasteland are interspersed. Wasteland is an important source of weeds. With the help of wind and surface runoff, weeds continue to spread and invade neighboring cornfields, competing for water, fertilizer and sunlight resources, reducing the photosynthetic efficiency of corn and significantly reducing corn yield. At the same time, the existing wasteland along the highway is mostly scattered and disorderly. The size, density and morphology of the patches lack scientific control. Relying solely on manual weeding and chemical herbicides to control weeds is not only costly and prone to causing soil and non-point source pollution, but also fails to give full play to the ecological function of wasteland as a natural buffer and isolation, making it difficult to ensure stable farmland yields.
[0003] Current research on weed control in farmland mainly focuses on field weeding techniques and the construction of single-vegetation buffer zones. It lacks a technical system that quantifies the role of grassland spatial structure in blocking weed spread from the perspective of landscape pattern. There are few complete methods that combine multi-source remote sensing and machine learning models to identify the optimal landscape configuration threshold for grassland. It is impossible to accurately match grassland control schemes in different environments by region, and it is difficult to simultaneously achieve multiple goals such as blocking weed sources and sinks and increasing corn yield along highways. The industry urgently needs a multi-model coupled, quantifiable, and practical method for coordinated yield increase control of grassland landscape. Summary of the Invention
[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield. The method includes: Includes the following steps: Step 1: Identify understory grasslands along highways based on multi-source remote sensing images from the GEE cloud platform, calculate corn yield and construct weed suppression index, and use GIS to calculate various landscape structure parameters and regional environmental factors of the grasslands. Step 2: Eliminate highly collinear factors, construct a random forest regression model with the weed suppression index as the response variable, and screen core landscape parameters that make significant contributions through permutation importance analysis; Step 3: Use a nonlinear response model to fit the relationship between key landscape parameters and weed suppression index, identify the threshold inflection point of landscape parameters by resampling and differentiation, and complete the significance test; Step 4: Train the XGBoost regression model to represent the quantitative relationship between landscape parameters and weed suppression index, optimize landscape parameters through grid search and simulate and predict the weed suppression effect after optimization, quantify the yield increase regulation effect, output the wasteland spatial optimization configuration scheme, rely on the optimized wasteland to block the spread of weeds to corn fields, and simultaneously achieve the landscape quality improvement of wasteland along the highway and the stable yield increase of corn.
[0005] Preferably, step one specifically includes: using Sentinel-2 multispectral imagery and Sentinel-1 radar imagery as data sources, conducting cloud detection using MSK_CLDPRB cloud probability map and SCL scene classification band, and removing areas with a cloud probability > 30%. Construct a multi-dimensional feature dataset including spectral, radar, texture, spatial, and topographic features, and use the random forest algorithm to identify understory grassland; Simultaneously, based on the 30m land use and the 100m buffer zone of the road network, grassland within 100m of the cultivated land was extracted as the target wasteland; Maize yields were extracted from the 0-60m and 60-180m layers of grassland, and the difference between the two was defined as the weed inhibition index. The larger the positive value of the index, the weaker the effect of grassland on yield reduction. A confusion matrix was used to verify the classification accuracy of forest grassland, ensuring an overall accuracy greater than 90%. Under a custom projection coordinate system, a 5000m×5000m fishing net zone was divided. The mean values of environmental factors and weed suppression index of each grid were statistically analyzed. The zone ratio, patch density, separation degree, landscape spatial separation degree, two reference shape indices, fractal dimension and all other landscape structural parameters were calculated.
[0006] Preferably, the calculation formulas for each landscape structure parameter in step one are as follows: Bambi ; In the formula, This refers to the area of wasteland; The area of cultivated land within 60 meters of the grassland; Patch density ; In the formula, N is the number of grassland patches; A represents the area of cultivated land (hectares) within 60 meters of the grassland, in units of: units / hectare; Resolution ; In the formula, A represents the total area of grassland patches within the region; Let i represent the area of the grassland patch; Landscape spatial separation ; In the formula, A represents the total area of grassland patches within the region; Let i represent the area of the grassland patch; Shape index based on a circle ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch; Shape index based on square ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch; fractal dimension ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch.
[0007] Preferably, the environmental factors mentioned in step one include precipitation, temperature, normalized vegetation index, wind speed, soil organic matter, soil crust factor, clay content, and silt content.
[0008] Preferably, the second step is as follows: First, conduct correlation analysis and variance inflation factor test on all landscape parameters and environmental factors to remove highly collinear variables; divide the dataset into training and test sets in an 8:2 ratio, and configure the random forest regression model with n_estimators=1000, min_samples_split=5, and min_samples_leaf=2; set 100 repeated calculations of permutation importance, sort in descending order of importance mean, and select landscape parameters with z_like≥1.96 and p<0.05 as key variables.
[0009] Preferably, in step three, the nonlinear response model is constructed using the mgcv package in the R language: g(E[Y])=β0+f(x); Y represents the weed control index; f(x) is a smoothing function; The AICc criterion is used to optimize the smoothing term k value, and the REML method is used to estimate the parameters; The first derivative of the model is calculated by Bootstrap resampling. The point where the derivative crosses zero is taken as the threshold inflection point and the 95% confidence interval is calculated. The Mann-Whitney U test is used to compare the significance of the slope difference before and after the threshold.
[0010] Preferably, step four specifically includes: training an XGBoost regression model with the selected landscape parameters and environmental factors, and evaluating the model accuracy through root mean square error, coefficient of determination, and mean absolute error. The optimal parameter combination is obtained by grid search of key landscape parameters, and the prediction feature matrix is constructed by splicing invariant environmental factors. The matrix is then input into the trained XGBoost model to predict the optimized weed suppression index. The change in weed suppression index is calculated as follows: Optimized predicted value - Original predicted value; Percentage change = Change ÷ Absolute value of original value × 100%. The average change, average percentage change, number of optimized grids and their proportions for each zone are statistically analyzed, and the final spatial layout and landscape structure optimization scheme for the wasteland is determined based on the statistical results.
[0011] Based on the above, relying on the GEE cloud platform's remote sensing interpretation, random forest, nonlinear response model, and XGBoost multi-model coupled analysis system, it is possible to accurately quantify the nonlinear effects of various landscape parameters of wasteland on the weed suppression index. By identifying the threshold inflection point of landscape structure through Bootstrap resampling, the optimal range of key indicators such as band ratio and patch density can be determined. This can significantly improve the ability of wasteland along highways to block the spread of weeds to cornfields, effectively weakening the stress of weeds on corn yield reduction. After zonal simulation calculation, the optimized weed suppression index of the entire area has been significantly improved, directly achieving stable and increased corn yield. At the same time, by scientifically regulating the patch layout of wasteland, the current situation of disordered vegetation along roads can be improved, and the regional landscape quality can be improved simultaneously.
[0012] This method is based entirely on remote sensing spatial zoning and quantitative calculations, eliminating the need for large-scale field sampling, thus reducing labor and monitoring costs and avoiding the risk of non-point source pollution to soil and water bodies caused by chemical weeding. Through grid search, it automatically outputs regionally differentiated wasteland spatial optimization configuration schemes, adapting to heterogeneous environmental conditions such as precipitation, soil, and topography. The model has been verified by confusion matrix and multiple accuracy indicators, demonstrating stable and reliable predictive capabilities. It can be applied on a large scale to the ecological management of farmland along dryland highways, taking into account the multiple values of agricultural production benefits, road ecological landscape, and arable land soil protection. Attached Figure Description
[0013] Figure 1 This is a flowchart of a remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing corn yield according to the present invention. Figure 2 This is a schematic diagram of the importance analysis results of a random forest permutation according to the present invention, showing the ranking of the contributions of various landscape parameters and environmental factors to the weed suppression index; A schematic diagram of the fitting curve and threshold identification of the nonlinear response model, where the solid line is the fitting curve, the shaded area is the 95% confidence interval, and the dashed line is the threshold inflection point. Figure 3 This is a comparison image of the landscape feature parameters of the wasteland along the highway before and after optimization in Implementation Method 1. Figure 4This is a comparison image of the landscape feature parameters of the wasteland along the highway before and after optimization in the second specific implementation method. Figure 5 This is a comparison chart of the landscape feature parameters of the wasteland along the highway before and after optimization in the third specific implementation method. Figure 6 This is a comparison chart of the weed suppression index before and after optimization in Implementation Method 1; Figure 7 This is a comparison chart of the weed suppression index before and after optimization in the second specific implementation method; Figure 8 This is a comparison chart of the weed suppression index before and after optimization in Implementation Method 3. Detailed Implementation
[0014] The present invention will now be described in detail with reference to the accompanying drawings.
[0015] This method constructs a spatial response model between the landscape structure of grassland along highways (including key parameters such as band ratio, patch density, and separation) and the weed suppression index of maize yield, accurately quantifying the impact mechanism of grassland configuration on maize yield. Based on this, the spatial layout and structural parameters of the grassland are further optimized to significantly enhance its barrier capacity against weed spread, thereby effectively reducing the negative impact of weeds on adjacent maize yields. Ultimately, this achieves a scientific, efficient, and sustainable increase in maize yield in dryland areas, providing key technical support for enhancing the yield-increasing function of dryland agricultural ecosystems.
[0016] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield, including the following steps: Step 1, data acquisition and preprocessing. Based on Sentinel-2 multispectral imagery and Sentinel-1 radar imagery from the GEE cloud platform, cloud detection is performed using the cloud probability map MSK_CLDPRB and scene classification SCL band, and areas with a cloud probability > 30% are removed. A comprehensive dataset containing spectral, radar, texture, spatial, and topographic features is constructed, and the spatial distribution of understory grassland in the study area is identified using the random forest algorithm. At the same time, grassland within 100 meters of cultivated land is extracted based on 30-meter land use data, and grassland and understory grassland data within 100 meters of cultivated land around roads are extracted as wasteland based on a 100-meter buffer constructed from road network data. Based on the extracted wasteland, maize yield data are extracted within the ranges of 0-60 meters and 60-180 meters, respectively. The difference between the two is used as the weed suppression index for maize yield. The larger the positive value of the index, the weaker the negative edge effect of wasteland on yield. The accuracy of the classification results was verified using a confusion matrix: the overall accuracy of the extracted data was greater than 90%, indicating that the identification accuracy of understory grassland was high.
[0017] Using ArcGIS Pro software, fishing nets were created within a custom NER BlackSoil Albers projection coordinate system, covering the study area, with a net size of 5000m × 5000m. Environmental factors and weed suppression indices were extracted from each net. The method involved first associating the environmental factor and weed suppression index layers, then extracting raster values by partitioning the data, and finally extracting the mean environmental factors and the mean weed suppression index for each net. Spatial statistical tools were used to extract the grassland landscape structure parameters for each net, including grassland area, cultivated land patch area, and patch perimeter. These parameters were then used to calculate the grassland structure parameters, including banding ratio. ; In the formula, This refers to the area of wasteland. The area of cultivated land within 60 meters of the grassland; Patch density ; In the formula, N is the number of grassland patches, and A is the area of cultivated land within 60 meters of the grassland (hectares), unit: patches / hectare; Resolution ; In the formula, A represents the total area of the grassland patches within the region. Let i represent the area of the grassland patch; Landscape spatial separation ; In the formula, A represents the total area of the grassland patches within the region. Let i represent the area of the grassland patch; Shape index based on a circle ; In the formula, P is the perimeter of the grassland patch, and A is the area of the grassland patch; Shape index based on square ; In the formula, P is the perimeter of the grassland patch, and A is the area of the grassland patch; fractal dimension ; In the formula, P is the perimeter of the grassland patch, and A is the area of the grassland patch.
[0018] Step 2: Random Forest Permutation Importance Analysis. Before model construction, correlation and variance inflation factor tests were performed on the factors to remove variables with high collinearity. The weed suppression index of maize yield was used as the response variable, and grassland landscape parameters (SCR, PD, SPLIT, SIS) and environmental factors (precipitation, temperature, normalized difference vegetation index, wind speed, soil organic matter, soil crust factor, clay content, silt content) were used as explanatory variables. The dataset was divided into training and test sets in an 8:2 ratio. After training, permutation importance was calculated with 100 repetitions. Variables were sorted in descending order of importance mean, and the z-like values of each variable were calculated for empirical significance assessment (z-like ≥ 1.96 corresponds to p < 0.05). Landscape parameters with higher importance were selected as key variables for subsequent analysis.
[0019] The following is a key example code snippet: rf = RandomForestRegressor( n_estimators=1000, max_depth=None, min_samples_split=5, min_samples_leaf=2, random_state=42, n_jobs=1 ) rf.fit(X_train, y_train) perm_imp = permutation_importance( rf, X_test, y_test, n_repeats=100, random_state=42, n_jobs=-1 ) perm_imp = permutation_importance( rf, X_test, y_test, n_repeats=100, random_state=42, n_jobs=-1 ) imp_df = pd.DataFrame({ 'feature': features, 'importance_mean': perm_imp.importances_mean, 'importance_std': perm_imp.importances_std }).sort_values('importance_mean', ascending=False) Step 3: Nonlinear Response Model Construction and Threshold Identification. Using the mgcv package in R, a nonlinear response model is constructed with the weed suppression index as the response variable Y and key landscape parameters (such as PD) as explanatory variables: g(E[Y])=β0+f(x), where f(x) is the smoothing function. The optimal k value is selected using the AICc information criterion, and the parameters are estimated using the restricted maximum likelihood REML method. The p-value of the smoothing term is extracted to evaluate the model significance. The first derivative is calculated using Bootstrap resampling, and the zero-crossing point of the derivative is identified as the threshold inflection point. The 95% confidence interval of the threshold is calculated. The Mann-Whitney U test is used to compare the significance of the difference in the average slope within the window before and after the threshold.
[0020] Step 4: XGBoost Model Construction and Parameter Optimization: Train an XGBoost regression model based on the selected original landscape parameters and environmental factor data to establish the relationship between input features and the weed suppression index. Evaluate the model performance using the root mean square error, coefficient of determination, and mean absolute error of the training and test sets. Then, read the dataset containing key landscape parameters and perform a grid search on the landscape parameters. Combine the optimized landscape parameters with the unchanged environmental factors to construct a prediction feature matrix. Use the trained XGBoost model to estimate the optimized weed suppression index. Calculate the change as the difference between the optimized predicted value and the original predicted value plus the original value. For the weed suppression index, the change is calculated as the optimized predicted value minus the original predicted value plus the original value. Calculate the percentage change by dividing the change by the absolute value of the original value and then multiplying by 100%. Statistically calculate the average change, average percentage change, number of optimized fishing nets and their proportion to evaluate the effectiveness of the optimization scheme.
[0021] This method constructs a spatial response model of the landscape structure parameters of grassland along highways and the weed suppression index of maize yield, accurately quantifying the impact mechanism of different grassland configurations on maize yield. Based on this, the spatial layout and structural parameters of grassland are further optimized, significantly enhancing the grassland's ability to block weed spread, thereby effectively reducing the negative impact of weeds on adjacent maize yields. Ultimately, this achieves a scientific, efficient, and sustainable increase in dryland maize yield, providing key technical support for enhancing the yield function of dryland agricultural ecosystems.
[0022] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.
[0023] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield, characterized by: Includes the following steps: Step 1: Identify understory grasslands along highways based on multi-source remote sensing images from the GEE cloud platform, calculate corn yield and construct weed suppression index, and use GIS to calculate various landscape structure parameters and regional environmental factors of the grasslands. Step 2: Eliminate highly collinear factors, construct a random forest regression model with the weed suppression index as the response variable, and screen core landscape parameters that make significant contributions through permutation importance analysis; Step 3: Construct a nonlinear response model to fit the relationship between key landscape parameters and weed suppression index, identify the threshold inflection point of landscape parameters by resampling and differentiation, and complete the significance test; Step 4: Train the XGBoost regression model to represent the quantitative relationship between landscape parameters and weed suppression index, optimize landscape parameters through grid search and simulate and predict the weed suppression effect after optimization, quantify the yield increase regulation effect, output the wasteland spatial optimization configuration scheme, rely on the optimized wasteland to block the spread of weeds to corn fields, and simultaneously achieve the landscape quality improvement of wasteland along the highway and the stable yield increase of corn.
2. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 1, is characterized in that: Step one specifically includes: using Sentinel-2 multispectral imagery and Sentinel-1 radar imagery as data sources, conducting cloud detection using MSK_CLDPRB cloud probability map and SCL scene classification band, and removing areas with a cloud probability > 30%. Construct a multi-dimensional feature dataset including spectral, radar, texture, spatial, and topographic features, and use the random forest algorithm to identify understory grassland; At the same time, combining 30m land use data and a 100m buffer zone of road network, grassland within 100m of cultivated land is extracted as target wasteland; Maize yields were extracted from the 0-60m and 60-180m layers of grassland, and the difference between the two was defined as the weed inhibition index. The larger the positive value of the index, the weaker the effect of grassland on maize yield reduction. A confusion matrix was used to verify the classification accuracy of forest grassland, ensuring an overall accuracy greater than 90%. Under a custom projection coordinate system, a 5000m×5000m fishing net zone was divided. The mean values of environmental factors and weed suppression index of each grid were statistically analyzed. The zone ratio, patch density, separation degree, landscape spatial separation degree, two reference shape indices, fractal dimension and all other landscape structural parameters were calculated.
3. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 2, is characterized in that: The calculation formulas for each landscape structure parameter in step one are as follows: Bambi ; In the formula, This refers to the area of wasteland; The area of cultivated land within 60 meters of the grassland; Patch density ; In the formula, N is the number of grassland patches; A represents the area of cultivated land (hectares) within 60 meters of the grassland, in units of: units / hectare; Resolution ; In the formula, A represents the total area of grassland patches within the region; Let i represent the area of the grassland patch; Landscape spatial separation ; In the formula, A represents the total area of grassland patches within the region; Let i represent the area of the grassland patch; Shape index based on a circle ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch; Shape index based on square ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch; fractal dimension ; In the formula, P is the perimeter of the grassland patch; A represents the area of the grassland patch.
4. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 2, is characterized in that: The environmental factors mentioned in step one include precipitation, temperature, normalized vegetation index, wind speed, soil organic matter, soil crust factor, clay content, and silt content.
5. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 1, is characterized in that: Step two's operation process is as follows: First, conduct correlation analysis and variance inflation factor test on all landscape parameters and environmental factors to remove highly collinear variables; divide the dataset into training and test sets in an 8:2 ratio, configure the random forest regression model with n_estimators=1000, min_samples_split=5, and min_samples_leaf=2; set 100 repeated calculations of permutation importance, sort in descending order of importance mean, and select landscape parameters with z_like≥1.96 and p<0.05 as key variables.
6. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 1, is characterized in that: In step three, the nonlinear response model is constructed using the mgcv package in the R language: ； Y represents the weed control index; It is a smoothing function; The AICc criterion is used to optimize the smoothing term k value, and the REML method is used to estimate the parameters; The first derivative of the model is calculated by Bootstrap resampling. The point where the derivative crosses zero is taken as the threshold inflection point and the 95% confidence interval is calculated. The Mann-Whitney U test is used to compare the significance of the slope difference before and after the threshold.
7. The remote sensing machine learning method for optimizing the landscape of highway wasteland, suppressing weeds, and increasing maize yield according to claim 1, is characterized in that: Step four specifically includes: training an XGBoost regression model using the selected landscape parameters and environmental factors, and evaluating the model accuracy using root mean square error, coefficient of determination, and mean absolute error. The optimal parameter combination is obtained by grid search of key landscape parameters, and the prediction feature matrix is constructed by splicing invariant environmental factors. The matrix is then input into the trained XGBoost model to predict the optimized weed suppression index. The change in weed suppression index is calculated as follows: Optimized predicted value - Original predicted value; Percentage change = Change ÷ Absolute value of original value × 100%. The average change, average percentage change, number of optimized grids and their proportions for each zone are statistically analyzed, and the final spatial layout and landscape structure optimization scheme for the wasteland is determined based on the statistical results.