Soil health is vital for getting a good crop yield. Analysis of available soil nutrients done at the right time not only helps in conservation of soil fertility but can also help in getting a good crop yield by limiting the usage of external inputs to the soil such as fertilizers, water etc. The use of AI in agriculture is being explored in recent times to optimize the crop yield. Machine Learning techniques are used in developing smart soil sensing systems to provide accurate soil nutrients distribution. In this study, a sample of 40 spectral data in the frequency range of 500MHz to 1000MHz was passed to the ParLeS software. The PLSR cross validation in ParLeS gave us an RMSE of 2.87. However, when Ridge regression based on machine learning was applied, we obtained a RMSE of 1.02 with parameter alpha set to 0.005. Thus, we can say conclusively that, machine-learning methods yield better results than traditional methods. In addition, implementation of ParLeS needs LabVIEW type of environment and needs external graphics support, whereas, Ridge regression can be implemented using simple Python environment, which is now a day most often used programming language. The implementation does not require compulsory graphics support.
ESTIMATION USING MACHINE LEARNING TOOLS⋆
Agriculture is the backbone of any thriving economy. Advancements in technology has seen lot
of influence on the way agriculture is practiced. Smart farming has paved a way for sustainable
agriculture and increasing the crop productivity. Soil fertility plays an important role in crop
production [
Artificial Intelligence (AI) is the most rapidly growing technology embedded into all aspects of
human life. In agriculture AI technologies can be used in precision farming for soil and irrigation
management, weather forecasting, plant growth, disease prediction, and animal management [
Digital soil mapping (DSM) is used to generate digital maps of the type and quality of soil by
combining soil sensing data with environmental factors [
AI models and DSM have been utilized in soil fertility prediction, offering a decision-making tool
capable of forecasting the best crop based on soil pH, soil nutrients, soil moisture, environmental
variables, and other components [
In another review study on using machine learning methods for predicting soil properties,
agricultural yield, and soil fertility, it was observed that for soil prediction, Random Forest (RF) and
deep learning techniques surpass traditional ML algorithms. Depending on the model's inputs, the
RF and deep learning techniques can reliably forecast soil conditions and crop to be grown. It was
also found from the study that inaccurate data has the ability to reduce forecasting precision.
Variations in geographical elements, meteorological circumstances, and farming techniques can
hamper the process of generalizing models. Furthermore, selecting relevant characteristics from
numerous influencing factors necessitates subject expertise and testing [
To obtain the RF spectra of various samples a cell is designed based on the principle of dielectricity.
The design details of the cell are discussed in [
The RF spectrum analyzer connected at the receiver end of the cell captures signal proportional to the radiation loss due to the sample solution. The cell has a capacity of holding 15ml of liquid. Soil samples were prepared in the laboratory by mixing 5 different components urea, potash, sodium chloride, calcium carbonate and phosphate in distilled water. Molar solutions for each of the component was prepared and for 15ml of water the amount of each component required to be added was calculated. It was found that amount of urea required was 225mg/15ml. Similarly, for the remaining components the amount required to be added for 15ml of water was calculated. The amount of each component to be added is shown in table 1.
Urea 112.5 225 337.5 450 838.2 1890 3780 5670 7560 A b s o r p t i o n A U
Frequency(MHz
The soil urea estimation using this spectral data of 40 samples was done using two methods. The
first method was using the ParLeS software based on Partial Least Squares Regression(PLSR)
model. The second method was using Machine Learning Algorithm i.e. Ridge Regression.
ParLeS is a chemometrics software for multivariate modelling and prediction. It provides users with
various algorithm options to transform, preprocess and pretreat spectra. It may be used to implement
principal components analysis (PCA); partial least squares regression (PLSR) with leave-n-out cross
validation; and bootstrap aggregation-PLSR (bagging-PLSR). ParLeS facilitates the implementation
of a large number of preprocessing techniques as well as bagging -PLSR, which can improve the
robustness and accuracy of PLSR models. Other unique features of ParLeS include the provision of a
number of assessment statistics and graphical output as well as a user-friendly interface and
functionality [
In this study, a sample of 40 spectral data in the frequency range of 500MHz to 1000MHz was passed into the ParLeS software. The PLSR cross validation technique was used for soil urea estimation where the n=5 was chosen for the cross validation. Using this an RMSE of 2.87 was obtained. A screenshot of the ParLeS software is as shown in Figure 3. applied to linear regression models. Regularization is a statistical method used to prevent errors due to the overfitting of training data. Ridge regression is specifically tailored to address multicollinearity in regression analysis, which is crucial when developing machine learning models with many parameters, particularly when these parameters are significantly weighted.
A standard, multiple-variable linear regression equation is: = 0 + 1 1 + 2 2 + ⋯ + ……..(1) In the above equation, Y represents the expected value or the dependent variable, X is the predictor or independent variable, B denotes the regression coefficient linked to that independent variable, and X0 is the value of the dependent variable when the independent variable is zero, also referred to as the y-intercept. It's important to observe how the coefficients illustrate the relationship between the dependent variable and a specific independent variable. The best-fitting line for a given dataset is obtained by calculating coefficients for each independent variable that result in the smallest residual sum of squares (also called the sum of squared errors).
The Residual sum of squares (RSS) represents how well a linear regression model matches the training data and is represented by the formula: =
∑ =1( − ⏞ )2……..(2) This formula is used to calculate the accuracy of model predictions against the expected values in the training data. If the Residual Sum of Squares (RSS) is zero, it indicates that the model perfectly predicts the dependent variables. If two or more variables have a strong linear correlation, highvalue coefficients are generated, causing the model's output to be sensitive to minor changes in the input data. This indicates that the model overfitted on a single training dataset and is unable to correctly generalise to new test datasets. This causes the model to be unstable.
Multicollinearity exists when two or more predictors have a near-linear relationship or are highly correlated, which results into unreliable and unstable estimates of regression coefficients. Ridge regression is a procedure for eliminating the bias of coefficients and reducing the mean square error by shrinking the coefficients of a model towards zero in order to solve problems of overfitting or multicollinearity that are normally associated with ordinary least squares regression.
Ridge regression corrects for high-value coefficients by introducing a regularization term (often called the penalty term) into the RSS function. This penalty term denoted as L2, is the sum of the squares of the model’s coefficients. It is represented in the formulation: 2 =
∑ =1( − ⏞ ) 2
+ ∑ =1 2……..(3) The L2 penalty term reduces all coefficients to balance the high ones. This process is utilized in ridge regression, to calculate new coefficients that minimize the residual sum of squares (RSS) for a model, thereby reducing overfitting.
Ridge regression doesn't reduce all coefficients equally and is proportional to their original
magnitude. As the lambda (λ) parameter increases, coefficients with higher values diminish more
rapidly than those with lower values, resulting in a greater penalty for the former [
In machine learning, ridge regression is used to reduce overfitting that results from model
complexity. Model complexity can be due to a model possessing too many features and features
possessing too much weight. Feature weight refers to a given predictor’s effect on the model output.
In machine learning terms, ridge regression amounts to adding bias into a model for the sake of
decreasing that model’s variance. Bias measures the average difference between predicted values and
true values and variance measures the difference between predictions across various realizations of
a given model. As bias increases, a model predicts less accurately on a training dataset. As variance
increases, a model predicts less accurately on other datasets. Bias and variance thus measure model
accuracy on training and test sets respectively. To reduce the model bias and variance, ridge
regression technique can be used [
Using Ridge regression technique allows control over the bias-variance trade-off. Increasing the value of λ increases the bias but reduces the variance, while decreasing λ does the opposite. The goal is to find an optimal λ that balances bias and variance, leading to a model that generalizes well to new data. Selection of an appropriate value for the ridge parameter k is crucial in ridge regression, as it directly influences the bias-variance tradeoff and the overall performance of the model. There are several methods for the selection of ridge parameter:
Cross-validation is one of the most popular method used in the selection of the ridge parameter. In this method, the dataset is divided into multiple subsets, and the model is trained on some subsets while being validated on the remaining ones. The process is repeated over multiple iterations, and the average performance across all iterations is used to determine the optimal value of λ.
• •
K-Fold Cross-Validation: The dataset is divided into K subsets (folds). The model is trained on K- folds and validated on the remaining fold. This process is repeated K times, with each fold being used as the validation set once. The average performance across all folds is used to select λ.
Leave-One-Out Cross-Validation (LOOCV): A special case of K-fold cross-validation where K equals the number of observations. Each observation is used as a validation set once, and the model is trained on the remaining observations. This method is computationally intensive but provides an unbiased estimate of the model’s performance. 2. Grid Search: This method defines a grid of possible values for λ and the ridge regression model is trained for each value of λ. The performance of the model is evaluated for each value of λ from the grid and the one with the best performance is then selected as the ridge parameter.
Bayesian optimization is used to efficiently explore the space of possible λ values and find the optimal value. This method can be more efficient than grid search for large search spaces.
Use information criteria like Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) to select the optimal value of λ. These criteria balance model fit and complexity.
• Incorporate domain knowledge about the problem to guide the choice of λ. For example, if you know that overfitting is a significant concern, you might choose a larger value of λ [17]. Ridge regression was implements using the sklearn python package in the python programming language. The sklearn package includes the Application Programming Interface (API) interface to implement the same. The linear_model.Ridge() API is used to implement the ridge regression. The only parameters supplied to the API is alpha with a value of 0.005 . Here it may be noted that the alpha is equivalent to λ specified above. The others parameters have default values. The parameter with the default values are copy_X=True, fit_intercept=True, tol=0.0001, max_iter=None, positive=False, solver='auto', and random_state=None.
The dataset consisting of 40 samples was used for training and testing the ML model using Ridge regression. With the parameter alpha set to 0.005, the RMSE obtained using this technique was found to be 1.02.
It may be seen that the analysis using Ridge technique (Which is a machine learning based tool for regression analysis) gives excellent performance with error as low as 1.02. Whereas, the error in traditional technique of ParLeS regression is 2.87, which is nearly three times more than that of Ridge technique. As mentioned earlier that, in addition to the advantage of less error, the implementation of the algorithm can be done in a simpler computational platform, not necessarily requiring complicated LabVIEW back end. This is reflected in the table 2. The regression graph shown in figure 5 show good agreement between the actual and predicted values using the ridge regression technique.
In this article we studied the application of Ridge Regression Technique for analysis of urea in the soil for better productivity of the crops. In past we had done such analysis using ParLeS (which is propriety and not an open source software). The results obtained were encouraging with errors as low as 1.02mg/15ml. Therefore, we conclude here that Ridge Technique is far superior to the traditional technique of regression analysis. 5. References