An Integrative Approach for Mineral Nutrient Quantification in Dioscorea Leaves: Uniting Image Processing and Machine Learning

11.png

Figure 1. Image of Dioscorea leaves

2.2.2 Feature extraction

The extraction of multiple features to bolster the accuracy of regression models has been well-documented by several researchers [1-3, 9, 14]. The initial images are saved in an 8-bit RGB format, a mixed color space that leaves the color features vulnerable to external environmental influences such as lighting intensity. It has been suggested that using a set of invariant moments to normalize these features can provide positive outcomes [1, 15, 16]. The subsequent normalization formula for the color features can mitigate the light's impact on the image color, as displayed in Eqns. (1)-(3).

$r=R / \sqrt{R^2+G^2+B^2}$          (1)

$g=G / \sqrt{R^2+G^2+B^2}$          (2)

$b=B / \sqrt{R^2+G^2+B^2}$          (3)

Invariant moments of the image were employed, yielding a set of parameters that effectively eliminated the influence of external factors such as the environment and camera type on the image color. Certain color spaces utilize saturation and brightness as encapsulations of color perception. These color spaces, which include the HSI (Hue, Saturation, Intensity) color space, the Lab* color space (operating on three axes: a=green to red; b=blue to yellow; L=brightness), and the XYZ color space, have been employed to reduce the effect of light brightness on color. Eqns. (4)-(7) illustrate the conversion of RGB images to HSI, XYZ, and Lab* color spaces.

$H=\left\{\begin{array}{l}\theta\ (G \geq B) \\ 2 \pi-\theta\ (G<B)\end{array}\right.$          (4)

where, $\theta=\arccos \left\{\frac{\left[\left(R-G^{}\right)+(R-B)\right] / 2}{\left\lceil(R-G)^2+(R-B)(B-G)\right\rceil^{1 / 2}}\right\}$

$S=1-3 /(R+G+B) \cdot \min (R, G, B)$          (5)

$I=(R+G+B) / 3$          (6)

$\left\{\begin{array}{l}R_1=\operatorname{gamma}\left(\frac{R}{255.0}\right) \\ G_1=\operatorname{gamma}\left(\frac{G}{255.0}\right) \\ B_1=\operatorname{gamma}\left(\frac{B}{255.0}\right)\end{array}\right.$          (7)

where, $\operatorname{gamma}(x)= \begin{cases}\left(\frac{x+0.055}{1.055}\right)^{2.4}(x>0.04045) \\ \frac{x}{12.92} (\text { otherwise })\end{cases}$

$\left[\begin{array}{l}X \\ Y \\ Z\end{array}\right]=\left[\begin{array}{lll}0.4124 & 0.3576 & 0.1805 \\ 0.2126 & 0.7152 & 0.0722 \\ 0.0193 & 0.1192 & 0.9505\end{array}\right] *\left[\begin{array}{l}R_1 \\ G_1 \\ B_1\end{array}\right]$          (8)

$L=\left\{\begin{array}{l}116 \times f\left(Y / Y_n\right)-16.0 ; \text { if } \ Y / Y_n>0.008856 \\ 903.3 \times\left(Y / Y_n\right) ; \ \text { if } Y / Y_n \leq 0.008856\end{array}\right.$          (9)

$a^*=500\left[f\left(X / X_n\right)-f\left(Y / Y_n\right)\right]$          (10)

$b^*=200\left[f\left(Y / Y_n\right)-f\left(Z / Z_n\right)\right]$          (11)

where, $f(t)=\left\{\begin{array}{l}1 / 3\left(\frac{6}{29}\right)^3 t+4 / 29, \text { if } \ t \leq 0.008856 \\ t^{\frac{1}{3}}\ \ {i f }\ \ t>0.008856\end{array}\right.$

and $X_n, Y_n, Z_n$  describe a specified white object-color stimulus.

An average of all pixels across three color spaces in the target image was employed as a corresponding color evaluation indicator for the image. In this context, mean values of RGB mode images are denoted as R, G, and B respectively. Similarly, mean values for HSI and Lab* mode images are denoted as H, S, I, L, a*, and b* respectively. The formula to calculate the image mean index, denoted as µ, is presented in Eq. (12):

$\mu=\frac{1}{M N} \sum_{i=1}^M \sum_{j=1}^N P(i, j)$          (12)

where, M and N represent the rows and columns of effective pixels in the target image, and $P(i, j)$  represents the color grayscale value of the effective pixels.

In the realm of image analysis, a rise in color difference amongst effective pixels in an image corresponds to an increase in standard deviation for the same image feature. Thus, the standard deviation of the color of effective pixels has been utilized as an index of significance, as shown in Eq. (13).

$\sigma=\sqrt{1 / M N \sum_{i=1}^M \sum_{j=1}^N[P(i, j)-\mu]^2}$          (13)

Innovative indicators, derived from interactive processing or arithmetic combinations of basic color indicators, exhibit a more stable and closely related relationship with the chlorophyll content in plant leaves [1]. Previous literature [17-21] and multiple experimental studies have facilitated the selection of a combined set of indicators for the detection of mineral nutrient content in wheat, namely: R-G, R+B, R-G-B, R+G-B, NRI, NBI, NGI, and DGCI. The corresponding formulas are shown in Eqns. (14)-(17).

$N R I=\frac{R}{R+G+B}$          (14)

$N G I=\frac{G}{R+G+B}$          (15)

$N B I=\frac{B}{R+G+B}$          (16)

$D G C I=\frac{(H-60) / 60+(1-S)+(1-I)}{3}$          (17)

2.3 Analysis of modeling approaches

2.3.1 Regression via XGBoost

The employment of the XGBoost algorithm, a sophisticated version of the gradient boosting algorithm, has been evidenced as enhancing the realm of plant nutrient diagnosis through image analysis. In contrast to parallel computing algorithms like random forests, XGBoost implements a sequential operation, constructing and evaluating trees one after another. This approach enables a detailed and precise model to be built, augmenting the accuracy of plant nutrient diagnoses. The gradient boosting algorithm has applications in both regression and classification problems, and it should be noted that the internal decision tree utilized by XGBoost is a regression tree.

2.3.2 Regression via K-Nearest Neighbor (KNN)

The KNN algorithm, a relatively uncomplicated machine learning method, is versatile enough to handle both classification and regression tasks. In the context of image-based plant nutrient diagnosis, the algorithm can classify or predict plant nutrient statuses based on color features extracted from images. In classification, the nutrient status of a plant is determined by the most frequently occurring nutrient status amongst its k-nearest neighbors in the feature space. For regression problems, the mean nutrient content of its k-nearest neighbors is used to predict the plant's nutrient content. This approach enhances the accuracy and efficiency of plant nutrient status diagnoses, improving the efficacy of nutrient management strategies.

2.3.3 Regression via Lasso

The Lasso method, also known as the Least Absolute Selection and Shrinkage Operator, applies a shrinkage estimation technique that compacts variable coefficients and sets specific regression coefficients to zero through the construction of a penalty function. This technique results in variable selection and mitigates the overfitting problem commonly associated with standard linear regression. In image-based plant nutrient diagnosis, Lasso regression can be deployed to select the most significant color features from images for the prediction of plant nutrient status. By reducing the coefficients of less relevant features to zero, Lasso regression effectively diminishes the feature space's dimensionality and prevents overfitting, which in turn, improves the model's generalizability and predictive accuracy.

2.3.4 Regression via Multilayer Perceptron (MLP)

MLP, also referred to as an Artificial Neural Network (ANN), is a type of neural network characterized by the potential to incorporate multiple hidden layers, in addition to input and output layers. In an MLP, complete connectivity is maintained between layers, with the lowermost layer serving as the input layer, the middle layers functioning as hidden layers, and the final layer acting as the output layer. Within the context of image-based plant nutrient diagnosis, MLP can be employed to discern intricate patterns in color features extracted from images for predicting plant nutrient status. The input layer receives the color features, the hidden layers learn and represent the intricate relationships between these features and the nutrient status, and the output layer provides the predicted nutrient status. By learning these complex patterns, MLP effectively models the non-linear relationships between color features and nutrient status, enhancing the predictive accuracy of the model.

The utilization of digital image processing technology and machine learning methodologies was explored in this study, with the primary objective of extracting targeted images of Dioscorea leaves. A total of 54 color feature values were computed using RGB, HSI, XYZ, and Lab color spaces. The established correlations between these color features and SPAD values, alongside nitrogen (N), phosphorus (P), and potassium (K) content in Dioscorea leaves were evaluated.

It has been deduced from the findings that:

(1) Chlorophyll content in Dioscorea leaves can be predicted effectively using digital image technology and machine learning methodologies. This offers a robust foundation for non-destructive testing of Dioscorea leaves. In the arena of nutrient content prediction, the MLP model showed superior performance compared to other tested methods, such as KNN, XGBoost, and Lasso.

(2) A significant correlation was observed between color features and mineral nutrients. The process of normalization appears to diminish the influence of light on leaf characteristics, with the normalized RGB color space demonstrating a higher correlation with mineral nutrients than unnormalized color features.

(3) Combining color features produced a higher correlation with mineral nutrient content in Dioscorea leaves than a single variable. This suggests that a non-linear combination of features provides a more comprehensive understanding of the chlorophyll content in Dioscorea leaves. It paves the way for future research into the construction of additional image color features to enhance the evaluation of crop nutrient content.

(4) It should be noted that this study was conducted in a controlled laboratory environment. Therefore, future research should include the collection of leaf images in field prototypes, paving the way towards non-destructive and rapid detection of Dioscorea leaves.