A soil fertility prediction method based on a simplified neural network

Abstract

The application discloses a soil fertility prediction method based on a simplified neural network, and comprises the following prediction process: a simplified neural network model based on organic matter content is constructed; the model is embedded into a UAV; the UAV flies close to the ground, and images of a position to be predicted are shot; an organic matter content prediction result of the measured land is obtained; the soil organic matter content prediction result calculated by the model is combined with an improved Pearson coefficient to determine soil fertility information; and through the differentiation of the fertility grades, the final soil fertility grade of the position to be predicted is obtained. The application can obtain accurate soil fertility prediction information, and further helps to accurately determine crops suitable for planting on the land, can dynamically and accurately determine the soil fertility before sowing by agricultural workers, can make suggestions and recommendations for the planted crops, and can realize the maximum utilization of the land.

Description

Technical Field

[0001] This invention relates to a method for predicting soil fertility, and more particularly to a method for predicting soil fertility based on a simplified neural network. Background Technology

[0002] Currently, most soil fertility prediction studies involve on-site sampling and indoor acquisition of soil reflectance using spectrometers. This method eliminates the influence of vegetation and surface sediments, resulting in high accuracy. However, the obtained spectral data is point-like, making it difficult to dynamically monitor the organic matter content within a region. Furthermore, existing soil fertility studies rely on various satellite images, but satellite images suffer from limitations such as long revisit periods, pixel mixing, and weather conditions, making them unsuitable for accurate analysis of soil nutrient information. Gradually, low-altitude UAVs equipped with hyperspectral cameras have been developed for image acquisition. While this method is more expensive, images acquired using hyperspectral cameras can easily obtain high spatial resolution (centimeter-level) and multi-band (from visible to near-infrared) remote sensing data, thus achieving a balance between cost and usability. Neural network models can quickly and accurately obtain predictive information; combining them with soil fertility prediction will provide more reliable support for dynamic monitoring and accurate analysis of soil fertility. Summary of the Invention

[0003] To address the shortcomings of the aforementioned technologies, this invention provides a soil fertility prediction method based on a simplified neural network.

[0004] To solve the above technical problems, the technical solution adopted by this invention is: a soil fertility prediction method based on a simplified neural network, comprising the following prediction process:

[0005] Step 1: Construct a simplified neural network model based on organic matter content;

[0006] Step 2: Embed the simplified neural network model into the drone;

[0007] Step 3: A drone embedded with a simplified neural network flies close to the ground and takes images of the location to be predicted;

[0008] Step 4: Obtain the predicted organic matter content of the tested land by using a simplified neural network;

[0009] Step 5: Combine the predicted soil organic matter content calculated by the model with the improved Pearson coefficient to determine the soil fertility information.

[0010] Step 6: By differentiating the fertility levels, the final soil fertility level of the location to be predicted is obtained.

[0011] Furthermore, in step one, the process of constructing the simplified neural network model is as follows: data collection, splitting the training set and the test set, defining the neural network model, model training, model evaluation, and lightweight model tuning to obtain a simplified neural network suitable for soil fertility prediction.

[0012] Furthermore, during the model construction process, the defined neural network model is a combination of a CNN model and a support vector machine model.

[0013] Furthermore, during the model building process, the CNN model uses a residual network to optimize the convolution kernel, a support vector machine to find the maximum margin hyperplane, and then optimizes the SMO algorithm of the model to solve the dual problem of convex quadratic programming.

[0014] Furthermore, during the model construction process, Pearson correlation coefficient analysis is used to compare the accuracy of the models and evaluate them; the Pearson correlation coefficient analysis is shown in the following formula:

[0015]

[0016] In the formula, X i This represents the organic matter content of the i-th sample; Y represents the average organic matter content of all samples; i This represents the spectral value of the i-th sample. This represents the average spectral values ​​of all samples.

[0017] Furthermore, in step three, when using the drone to photograph the location to be predicted, the speed of the drone is controlled so that it can be correctly predicted by the simplified neural network model; the ground-hugging flight speed of the drone is determined by calculating the approximate time predicted by the model.

[0018] Furthermore, in step five, the accuracy of the prediction is determined based on the improved Pearson coefficient, as shown in the following formula:

[0019]

[0020] In the formula, ρ L，Q The value represents the accuracy of the prediction; L represents the actual organic matter content, which was obtained in advance during the model building phase; Q represents the organic matter content predicted by the model; and N represents the number of samples.

[0021] Furthermore, in step six, the organic matter content in the cultivated soil is graded according to its condition, and divided into three levels: excellent, good, and medium. When the organic matter content is <6g / kg, it is medium; when the organic matter content is between 6-10g / kg, it is good; and when the organic matter content is between 10-15g / kg, it is excellent.

[0022] This invention discloses a soil fertility prediction method based on a simplified neural network, which can obtain accurate soil fertility prediction information, thereby helping to accurately determine the crops suitable for planting on the land. It can dynamically and accurately determine soil fertility before agricultural workers sow seeds, and make suggestions and recommendations for the crops to be planted, thereby maximizing the utilization rate of the land. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the soil fertility prediction process of the present invention.

[0024] Figure 2 This is a schematic view of the simplified neural network structure of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] This invention discloses a soil fertility prediction method based on a simplified neural network, such as... Figure 1 As shown, the main prediction steps include the following:

[0027] Step 1: Construct a simplified neural network model based on organic matter content;

[0028] Step 2: Embed the simplified neural network model into the drone;

[0029] Step 3: A drone embedded with a simplified neural network flies close to the ground and takes images of the location to be predicted;

[0030] Step 4: Obtain the predicted organic matter content of the tested land by using a simplified neural network;

[0031] Step 5: Combine the predicted soil organic matter content calculated by the model with the improved Pearson coefficient to determine the soil fertility information.

[0032] Step 6: By differentiating the fertility levels, the final soil fertility level of the location to be predicted is obtained.

[0033] The process of constructing a simplified neural network model can be summarized as follows: data collection, splitting the training set and the test set, defining the neural network model, model training, model evaluation, and lightweight model tuning to obtain a simplified neural network suitable for soil fertility prediction.

[0034] The data required for model construction are mainly provided by existing measured soil organic matter content data and photos taken by drones. After cross-validation, the data is divided into n parts, and one part is used as the test set and the other n-1 parts are used as the training set.

[0035] A neural network is a widely parallel interconnected network composed of adaptive, simple units. Its organization can simulate the interactive responses of biological nervous systems to real-world objects. This invention uses a CNN model combined with a support vector machine to define the neural network.

[0036] Support Vector Machines (SVMs) are a machine learning method based on statistical theory, possessing a strong mathematical foundation and theoretical support. Leveraging the advantages of VC theory and structural risk minimization theory in small-sample, nonlinear, and high-dimensional pattern recognition, SVMs improve the model's generalization ability by minimizing structural risk. During operation, the problem is ultimately transformed into a quadratic optimization, thereby achieving a global optimum and obtaining better statistical results with fewer known samples.

[0037] The key to Support Vector Machines (SVMs) lies in the choice of kernel function. The kernel function computes inner products in a low-dimensional space, rather than directly mapping low-dimensional features to a high-dimensional space. This addresses the linear inseparability problem in low-dimensional spaces while avoiding the curse of dimensionality and the complexity of subsequent processing. According to functional theory, if a kernel function satisfies Mercer's condition, it will correspond to an inner product in a certain transformation space. Kernel functions satisfying Mercer's condition include polynomial functions, linear functions, radial basis functions (RBFs), and sigmoid functions, etc. Different types of SVMs can be constructed based on different kernel functions.

[0038] This invention mainly uses CNN and support vector machine models for prediction. The CNN model uses a residual network to optimize the convolution kernel, while the support vector machine mainly finds the hyperplane with the greatest distance from each type of sample point, that is, finds the maximum margin hyperplane. Then, the SMO algorithm of the model is optimized to solve the dual problem of convex quadratic programming.

[0039] Then, the accuracy of the model is compared and evaluated through Pearson correlation coefficient analysis. Specifically, Pearson correlation coefficient analysis is performed between the measured soil organic matter content and the spectral reflectance of the spectral set at the sampling points. The Pearson correlation coefficient is a statistical method used to reflect the degree of correlation between two data variables. Its r value is between -1 and 1; the larger the absolute value, the higher the correlation between the two variables, and the smaller the absolute value, the weaker the correlation. As shown in the following formula:

[0040]

[0041] In the formula, X i This represents the organic matter content of the i-th sample; Y represents the average organic matter content of all samples; iThis represents the spectral value of the i-th sample. This represents the average spectral values ​​of all samples.

[0042] The final optimal simplified neural network model is mainly obtained by pruning the model size and optimizing the heuristic algorithm, thus obtaining a simplified neural network model, such as... Figure 2 As shown in the figure, the input image format is 32×32, which is successively cropped into images of 28×28, 14×14, 10×10, and 5×5, and then output after being stacked by 120 layers and 84 layers.

[0043] After obtaining the simplified neural network, it is embedded into the drone. The neural network runs in the form of code, and the operating environment of the neural network is built on Raspberry Pi to realize the embedding of the neural network in the drone. The corresponding running page mainly displays a bounding box for image acquisition. When the command is issued, the Raspberry Pi will control the hyperspectral camera to take pictures of the image within the bounding box. The simplified neural network is used to predict the soil organic matter content of the corresponding area of ​​the captured image. Finally, the page will also display the organic matter content predicted by the model.

[0044] A drone embedded with a simplified neural network flies close to the ground to capture images of the location to be predicted. The geolocation system on the drone is used to determine the specific location information of the predicted location, thereby obtaining specific land information (mainly including the area to be measured) and the land distribution to be predicted (mainly including determining whether the measurement area is all land and marking buildings and water bodies).

[0045] When using a drone to photograph a predicted location, the speed of the drone is controlled so that it can be correctly predicted by a simplified neural network model; the drone's ground-hugging speed is determined by calculating the approximate time predicted by the model.

[0046] Furthermore, the captured images of the predicted locations are input into a simplified neural network to obtain the predicted organic matter content of the tested land. A support vector machine regression model constructed with first-order differential transformation is used to build an estimation model for soil organic matter content. Soil fertility is studied and analyzed by detecting the organic matter content in the soil, and the soil organic matter content is calculated based on the spectral information of the images.

[0047] The predicted soil organic matter content calculated by the model is combined with the improved Pearson coefficient to determine soil fertility information. The accuracy of the prediction is judged based on the Pearson coefficient, as shown in the following formula:

[0048]

[0049] In the formula, ρ L，QThe value represents the accuracy of the prediction; L represents the actual organic matter content, which was obtained in advance during the model building phase; Q represents the organic matter content predicted by the model; and N represents the number of samples.

[0050] Finally, by differentiating fertility levels, the soil fertility level of the location to be predicted is determined.

[0051] Whether the soil used for arable land is suitable for crop growth is one of the key factors determining the potential productivity of farmland. In the process of improving farmland soil, rationally arranging planting types and management measures, remote sensing technology can effectively monitor soil fertility and obtain soil fertility maps, providing a basis for farmland management. Soil fertility monitoring includes soil particle size, texture, and clay content, which affect soil structural parameters' ability to retain soil moisture and the migration of nutrients. These parameters can be used to assess fertilizer utilization and soil irrigation and drainage capacity.

[0052] This invention classifies soil organic matter content according to the organic matter content in cultivated land. The soil organic matter content is divided into three levels: excellent, good, and medium, based on threshold values. Specifically, the threshold values ​​are divided into three levels: <6g / kg, 6-10g / kg, and 10-15g / kg. By defining these three levels, soil fertility can be effectively classified and categorized into excellent, good, and medium fertility grades.

[0053] Soil organic matter content was determined using the potassium dichromate titration method with oil bath heating. Under heating conditions, excess potassium dichromate-sulfuric acid solution was used to oxidize soil organic carbon. The excess potassium dichromate was titrated with ferrous sulfate solution. The amount of organic carbon was calculated from the amount of potassium dichromate consumed using an oxidation correction factor. This result was then multiplied by a conversion factor of 1.724, obtained from the relevant chemical formulas, to determine the soil organic matter content.

[0054]

[0055] Where M is the molar concentration of the standard ferric salt solution, Vo is the volume (ml) of the ferric solution used in the blank titration, V is the volume (ml) of the ferric solution used in the sample titration, W is the dry soil type = sample weight * moisture coefficient, and 1.724 is the number of grams of organic matter equivalent to one gram of carbon.

[0056] Therefore, the soil fertility prediction method based on a simplified neural network disclosed in this invention uses a simplified neural network model combined with UAVs to predict soil fertility. The improved Pearson coefficient can specifically analyze the accuracy of soil fertility prediction, and can obtain accurate soil fertility prediction information, which helps to accurately determine the crops suitable for planting on the land. Through the soil fertility prediction method of this invention, agricultural workers can make dynamic and accurate judgments on soil fertility before sowing, so as to make suggestions and recommendations for the crops to be planted, thereby maximizing land utilization. At the same time, the rapid and accurate acquisition of soil fertility information can ensure the early prediction of soil changes.

[0057] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A method for predicting soil fertility based on a simplified neural network, characterized in that: The prediction process includes the following: Step 1: Construct a simplified neural network model based on organic matter content; The construction strategy includes: defining the neural network by combining a CNN model with a support vector machine; optimizing the convolution kernel using a residual network in the CNN model; finding the hyperplane with the greatest distance from each type of sample point to the hyperplane using the support vector machine, i.e., finding the maximum margin hyperplane, and then optimizing it using the SMO algorithm; comparing the accuracy of the neural network model through Pearson correlation coefficient analysis to achieve model evaluation; pruning the size of the neural network model and optimizing the heuristic algorithm to obtain a simplified neural network model; Step 2: Embed the simplified neural network model into the drone; Step 3: A drone embedded with a simplified neural network flies close to the ground and takes images of the location to be predicted; Step 4: Obtain the predicted organic matter content of the tested land by using a simplified neural network; Step 5: Combine the predicted soil organic matter content calculated by the model with the improved Pearson coefficient to determine the soil fertility information. Specifically, this involves inputting images of the predicted location taken by a drone flying close to the ground into a simplified neural network to obtain the predicted organic matter content of the tested land. A simplified neural network was used to construct a support vector machine regression model based on first-order differential transformation to build an estimation model for soil organic matter content. Soil fertility was studied and analyzed by detecting the organic matter content in the soil, and the soil organic matter content was calculated based on the spectral information of the image. The predicted soil organic matter content calculated by the model is compared with the spectral reflectance of the spectral set using Pearson correlation coefficient analysis. This is combined with an improved Pearson coefficient to determine soil fertility information. The accuracy of the prediction is then judged based on the improved Pearson coefficient, as shown in the following formula: In the formula, For the accuracy of the prediction; L The actual organic matter content was obtained in advance during the model building phase; Q The organic matter content predicted by the model. N The number of samples; Step 6: By differentiating the fertility levels, the final soil fertility level of the location to be predicted is obtained.

2. The soil fertility prediction method based on a simplified neural network according to claim 1, characterized in that: In step one, the process of constructing a simplified neural network model is as follows: data collection, splitting the training set and the test set, defining the neural network model, model training, model evaluation, and lightweight model tuning to obtain a simplified neural network suitable for soil fertility prediction.

3. The soil fertility prediction method based on a simplified neural network according to claim 2, characterized in that: During the model construction process, the Pearson correlation coefficient analysis is shown in the following formula: In the formula, X i Indicates the first i Organic matter content of each sample; This represents the average organic matter content across all samples. Y i Representing the i Spectral values ​​of each sample This represents the average spectral values ​​of all samples.

4. The soil fertility prediction method based on a simplified neural network according to claim 3, characterized in that: In step three, when using the drone to photograph the predicted location, the speed of the drone is controlled so that it can be correctly predicted by the simplified neural network model; the ground-hugging flight speed of the drone is determined by calculating the approximate time predicted by the model.

5. The soil fertility prediction method based on a simplified neural network according to claim 4, characterized in that: In step six, the organic matter content in the cultivated soil is graded according to its condition, and divided into three levels: excellent, good, and medium. When the organic matter content is <6g / kg, it is medium; when the organic matter content is between 6-10g / kg, it is good; and when the organic matter content is between 10-15g / kg, it is excellent.

Images