An automated AI-powered IoT algorithm with data processing and noise elimination for plant monitoring and actuating

Design of the measurement and monitoring system

The greenhouse is a special planting structure that can generate suitable conditions for the plants, enabling the plants to grow from sowing until harvesting. The main challenge in the greenhouse is how to isolate it from outside weather by selecting good greenhouse materials. For the weather inside the greenhouse, one can use the sensors to measure the plant’s surrounding conditions and apply the necessarily actions either by people in the manual methods or by actuators in automatic farming. Such actions will be applied based on the type of plants and recommendations by agricultural engineer. Thus, the practical challenges related to the variability in outside environmental conditions, geographic region and plant responses due to outside-weather, have no considerable effects on the planting processes in the greenhouse.

In the concept design of the proposed system as shown in Fig. 1A, three main different subsystems have been used, namely, irrigation, CO₂ detecting and plant infects detection systems, to simultaneously work with microcontroller and IoT systems.

In the irrigation system, a soil moisture sensor is immersed in the soil to detect the moisture level of the soil surrounding the plants and send the signals to the microcontroller. If the soil moisture level is low, the microcontroller will turn on the pump 1 and the water will be pumped into the soil. With CO₂ detection system, the CO₂ level of the greenhouse environment is measured every minute and if the CO₂ level around the plant is high, the microcontroller will turn on the exhaust fan to circulate the air inside the greenhouse environment. Therefore, the CO₂ level in the air will return to its normal level again. In the plant infects detection system, a camera installed on top of the plant, is used to give a live steaming video of the plant. If the plant is found to be infected by pests, pesticide will be pumped to the plant using pump 2.

This concept of design was generated based on three main aspects, namely, sensors, actuators and data processing. Five sensors are used to obtain the measurements for each system, such as soil moisture, humidity and temperature detection, CO₂ detector and camera sensors. The system applies the necessary actions on the plant using actuators such as irrigation water pumps, fans and LED lamps. The data processing of the system includes the signal and image processing of sensors reading and IoT system. In this design, the camera detects the conditions of the plant’s leaves through a live stream video which is processed in MATLAB using image acquisition and image processing toolbox. The soil moisture sensor is used to detect the moisture level of the soil of the chili plant. The sensor signals are then processed in the embedded system and the results are sent to the irrigation pump. The CO₂ detector is used to detect the concentration of CO₂ levels in the air inside the greenhouse. When CO₂ concentration is high, the exhaust fan is used to circulate the air inside the greenhouse and reduce the concentration of CO₂.

The air humidity sensor is able to detect the humidity, as well as the temperature of the air in the greenhouse, so that enable to determine if the condition is suitable to the plant. The relays are used to automatically turn on/off the power supply to the pumps and fans, according to the data coming from microcontroller of the proposed system.

Figure 1 shows the concept design of the proposed system.

The electronic breadboard is used to create a tidy and well managed circuit as shown in Fig. 1B. Firstly, 5V and 3.3V (V_CC) power supply from the Arduino microcontroller is connected to the top and bottom positive terminal respectively. Then, the ground terminal is connected to both negative terminals to complete the circuit. The V_CC terminal of the relay is connected to the 5V circuit while the V_CC terminals of the soil moisture, CO₂ detector and temperature sensors are connected to the 3.3V circuit. Next, the ground terminal from either side of the 5V and 3.3V circuit is connected to a new column of the breadboard. The ground terminals from all sensors and relays are then connected to this column. For data input, A₀ terminal from the soil moisture sensor is connected to the A₀ terminal of Arduino, A₀ terminal of CO₂ detector sensor is connected to the A₁ terminal of Arduino and the OUT terminal of temperature sensor is connected to the D₇ terminal of Arduino. The inputs of relay’s channel 1 and channel 2 are connected to the D₈ and D₉ terminal of Arduino, respectively. For the actuators, both positive terminals of the batteries are connected respectively to the Common (C) terminal of the relay channel 1 and channel 2. Then, the positive terminal of the pump is connected to the Normally Closed (NC) terminal in channel 1 while the negative terminal is connected to the negative terminal of the battery 1 to complete the circuit. Lastly, the positive terminal of the exhaust fan is connected to the Normally Opened (NO) terminal in channel 2 while the negative terminal is connected to the negative terminal of the battery 2.

Development of monitoring system prototype

A small greenhouse prototype with dimensions of 70 × 50 × 50 cm has been developed in the lab to test the proposed system as shown in Fig. 2. A waterproof box with a plastic cover are placed at the sides of the plant to seal the greenhouse and protect the electrical components and circuit from the surrounding environments. A webcam called, 360 Degree with Microphone, is located at one corner of the box for stream live recording of the plant status. Four sensors, namely, moisture sensor module (Model: SN-MOISTURE-MOD), CO₂ detector sensor module (Model: SN-MQ135-MOD), temperature and humidity sensor (Model: DHT22) are put in the plant soil, beside the plant and inside the green house, respectively. The fan with DC5V Brushless Fan 6,000 rpm-Jetson Nano model is used for circulating the air inside greenhouse. Two Micro-Submersible Water Pumps DC 3V–5V are used to supply water and apply the suitable treatment of the pest. In addition, Arduino with a model of MKR WiFi 1000, is used as microcontroller to receive data from sensors, send data to actuators and processing data the data in C-language program. 2 Channel 5V Active Low Relay Module (Model: BB-RELAY-5V-02) is used to switch the pump and fan with the data coming from microcontroller. The total cost of the proposed system is around 91.8USD.

The system is built in the lab with a small scale greenhouse prototype using cheap sensors, actuators and microcontroller. It can be enlarged for big scale environments such the standard greenhouse with dimensions of 2,000 × 900 × 300 cm, by increasing the number of sensors, actuators and micro-controllers in three rows arrangements: two on the sides of greenhouse and another one on the ceiling. The data processing can be done in separate microcontrollers and sent the decisions to farmer through IoT. Such system is cost-effective as the capital cost of the system may be higher at the beginning comparing with the manual system but the operational cost is lower. The most important feature of the proposed system is that it can discover the diseases that harm the plant once it happens which cannot be detected by manual methods with operators.

The proposed system can be used in the greenhouse for any kind of plants by utilizing the sensors to adjust the weather conditions inside the greenhouse based on the plants standards and recommendations of agricultural engineer. We have applied the proposed system on the chili plant as it is effected too much by the weather changing.

In other words, the proposed system can be applied to any other plant with changing of the threshold of the sensors during processing, such as temperature, humidity and CO₂ level, so that can control all the above-mentioned parameters based on the required conditions. In fact, the sensors will keep updating on the status of plant conditions and the decision making will be taken by signal processing and ELSL algorithms which is then converted into electrical signals to turn on/off pumps and fans.

Sensors’ signal and image processing

Two programming languages, namely, Arduino IDE and MATLAB are used in this project for signal and image processing, respectively. The algorithms used for both processes are analyzed in details as follows:

-Sensors signal processing

Three sensors are used to sense the soils moisture, CO₂ level and humidity with temperature inside greenhouse. The signals of those sensors are processed in Arduino microcontroller to extract the suitable features as follows:

- Soil moisture sensor:

SN-MOISTURE-MOD sensor module is used to measure the moisture in the soil. If the moisture sensor output is low, then the pump will switch on, however, when the moisture sensor output is high, the pump will be switched off. The pseudo-code algorithm for signal processing of this sensors is illustrated bellow:

• →

  Assign the input of relay to channel 1.

• →

  Assign soil moisture sensor to analog input pin.

• →

  Establish Setup function () for reading.

  • Setup baud rate into 9600

  • Read serial port of sensor.

  • Switch relay-pin into out mode

  • Switch sensor moisture into input mode

  • End function setup

• →

  Establish loop function () for reading.

  • Create Variable soilmoisturePin to read the data of analog port analogRead() function

  • Use map function () inside loop function.

    • Display the moisture level.

    • If statement (moisture level is low)

      • then relay channel 1 will stay at the NC channel, and the pump is turned on.

      • else the relay channel 1 will switch to NO channel, and the pump is turned off.

      • End if statement

    • End map function

• →

  End loop function

- Gas sensor

The MQ-135 gas sensor is used to measure the level of CO₂ inside the greenhouse. If the gas conductivity is low (less than 50 ppm), then the exhausted fan will switch off, however, when the gas conductivity of gas sensor is low, the fan will be switched off. The pseudo-code algorithm for signal processing of this sensors is illustrated bellow:

• →

  Assign the input of relay to channel 2.

• →

  Assign MQ-135 to analog input pin.

• →

  Establish Setup function () for reading.

  • Setup baud rate into 9600

  • Read serial port of gas sensor.

  • Switch relay-pin into output mode

  • Switch sensor moisture into input mode

  • End function setup

• →

  Establish loop function () for reading.

  • Calculate conductivity in percents

  • Create Variable gasPin to read the data of analog port analogRead() function

  • Use map function () inside loop function.

    • Display the gas level.

    • If statement (conductivity <50) Then

      • relay channel 2 will stay at the NC channel, and the exhausted fan is turned off.

      • else the relay channel 2 will switch to NO channel, and the exhausted fan is turned on.

      • End if

    • End map function

• →

  End loop function

- Temperature and humidity sensor

DHT22 sensor is used to measure the humidity and temperature of greenhouse. It enables to track the temperature inside the greenhouse and check whether the condition is healthy for the plant or not. The pseudo-code the algorithm is described herewith:

• →

  Include DHT.h library.

• →

  define DHTPIN pin to indicate the location of the sensor input pin

• →

  define DHTTYPE DHT22 to indicate the type of DHT sensor in our system

• →

  Initialise DHT sensor for normal 16mhz Arduino

• →

  Define variables humidity and temperature as float variable

• →

  Establish loop function () for reading.

  • Create Variable soilmoisturePin to read the data of analog port analogRead() function

  • Use map function () inside loop function.

    • Display the moisture level.

    • If statement (moisture level is low)

      • then relay channel 1 will stay at the NC channel, and the pump is turned on.

      • else the relay channel 1 will switch to NO channel, and the pump is turned off.

      • End if statement

    • End map function

• →

  End loop function

Video and image processing

The video and image processing are accomplished in MATLAB as it has ready image processing toolboxes. Three main stages are used to detect the pest and defects on the leaves of chili as follows:

Video pre-processing

A live streaming video is used to detect the pests on the leaves which is decomposed into 60 frames per second. A frame per second is used for images analysis and features detection. The image pre-processing stage is used to prepare the frame in a good format for further processing in the processing and post-processing stages. It includes operations that can adjust the brightness, crop certain part of image and converting into grey scale. The results of the algorithm pre-processing processes are shown in Fig. 3. The Pseudo-code of image pre-processing algorithm is described herewith:

• →

  Construct an input video object using vid=videoinput(’winvideo’,1);

• →

  Preview the constructed video object: preview(vid);

• →

  Start decomposing of the video into frames: start(vid)

• →

  Take one frame per second from 60frame/s: input=getdata(vid,1);

• →

  imcrop() function to crop the image into a desired size: GH=imcrop(input,[x1 y1 x2 y2]);

• →

  rgb2gray() function to convert into grayscale : GS=rgb2gray(GH)

Image processing

Several operations will be applied on the frame that has already passed the pre-processing stage. This stage is aimed to find the features of the plant such as pests dark spots, holes and change of color on the leaves, which will be analyzed later to determine whether it is a pest or noise on the leaves. For this purpose, multiple filters are used for blurring image and edge detection using Gaussian filter and Canny filters, respectively as shown in Fig. 4. Canny filter has shown better performance in comparison with Prewitt and Sobel filters as shown in Figs. 4D–4F. The Pseudo-code of image processing algorithm is illustrated herewith:

• →

  Apply blurring to the image using Gaussian filter: PSF=fspecial(’gaussian’,5,8);I=imfilter(GS,PSF,"symmetric","conv");

• →

  Apply edge detection methods using Sobel, Prewitt and Canny: BWS=edge(I,’sobel’); BWP=edge(I,’prewitt’); BWC=edge(I,’canny’);

• →

  Choose suitable edge detection

• →

  Fill up the dark patches or holes in the image with white colour using the imfill() function P = imfill(BWC, ’holes’);

• →

  Invert black and white colour of image P: P= ~P.

The results of image processing algorithm are shown in Fig. 4.

Figure 4 shows that the pest spots on the leaves from the image are normally occurred as big dark holes. However, it is mixed with noise, which have to be eliminated using AI algorithms.

Noise elimination in signals and images

After applying the image and signal processing, there still some noise that must be eliminated to make the right decision and apply the necessary actions. The signals of moisture, CO₂ emission and temperature sensors includes some wrong reading due to the sensor quality and unexpected condition. This include two main noise: (1) the missing data and (2) the data that are too high or too low than normal measurements. The result of image processing includes many noise that are looked similar to the defects on the leave. These noise are not avoidable and coming from many sources such as: (1) light homogeneity (2) background of the image (3) unexpected object (4) shadow that may occur during capturing the live video.

Thus, there is a need for AI algorithm to classify the spots in the leave into defective and non-defective areas.

Signal based noise elimination algorithm

For the signals that acquired by moisture, CO₂ emission and temperature sensors, several types of pass filters are used to remove the unwanted signals which are too far from the range of values. This is done by determining the threshold upper and lower level for each sensor. The low pass filter is used for moisture sensor measurement as in Eq. (1), however high pass filter and band pass filters are used for CO₂ emission as in Eq. (2) and temperature sensors as in Eq. (3), respectively.

(1) $$y_{ns}=\left\{\begin{array}{ll}{x}_1& x_1\le 20\mathrm{\%}\\ \hfill 0& x_1>20\mathrm{\%}\end{array}\right\}$$where $x_1$ is the reading values of the moisture sensor, y_ns is the moisture sensor measurements after applying the low pass filter.

(2) $$y_{co2}=\left\{\begin{array}{ll}{x}_2& x_2\ge 50ppm\\ \hfill 0& x_2<50ppm\end{array}\right\}$$where $x_2$ is the reading values of the CO₂ emission detector sensors, y_co2 is the emission detector of CO₂ after applying high pass filter

(3) $$y_{ts}=\left\{\begin{array}{cc}0& x_3<12C^0\hfill \\ x_3& 12C^0\le x_3\le 15C^0\hfill \\ 0& x_3>15C^0\hfill \end{array}\right\}$$where $x_3$ is the reading values of the temperature sensor, y_ts is the temperature sensor measurement after applying band pass filter

Image-based noise elimination algorithm

Since, there are a lot of noise in the images after post-processing, a new AI algorithm called enhanced laser simulator (ELSL) has been introduced here to accurately find the defected areas of the leaves.

Principle of Enhanced laser Simulator Logic (ELSL):

It is an extension of Laser Simulator Logic (LSL) (Ali et al., 2022) that has been used for noise reduction. Both LSL and ELSL are capable for handling noise through high inference of linguistic variables with a dynamic range of membership functions. Such accommodation of the noise enables to avoid the drawbacks of fuzzy logic membership functions when it is dealing with highly overlapping membership functions of the linguistic variables, as shown in Fig. 5. Equations (4)–(13) present the fuzzification process in ELSL with lowly and highly inferenced linguistic variables. In comparison with LSL, ELSL simplifies the computation of membership functions for high inference linguistic variables, as derived in Eqs. (11)–(14). The main features of ELSL comparing with LSL are listed herewith:

- A step function to delete the effect of non-occurrence of linguistic variable is newly proposed, as given in Eqs. (11) and (12).

- A step function for calculation of any lowly or highly overlapped membership function of the crisp input is introduced in Eq. (13).

- A general equation that simplifies the calculation of membership values for both low and high overlapping situations is formulated in Eq. (10).

LSL approach depends on finding the ratio between the crisp values location and the universe of discourse of linguistic variables, i.e., the membership value of the crisp input is computed as the division of its value position into the whole input linguistic variable range as described in Eq. (4). The membership function of the crisp input/output in LSL and ELSL is modeled as drawn in Fig. 6. The membership values of inputs linguistic variables are then implicated on the corresponding output linguistic variables, according to fuzzy inference rules, to determine the output membership values.

Let us define the membership value of a fuzzy input/output set (Q) in the universe of discourse (X) as: ${\mu }_i:X\to \left[01\right]$ with $Q=\left\{\left(x,{\mu }_i\left(x\right)\right)|x\in X\right\}$. The ELSL membership values ${\mu }_i\left(x\right)$ are simply presented in Fig. 6 and computed using Eq. (4):

(4) $$\mu \left(x\right)=\left\{\begin{array}{ll}0& x\le s\\ {\displaystyle \frac{x-a}{\left(x-a\right)+\left(m-x\right)}}& s<x<m\\ {\displaystyle \frac{b-y}{\left(b-x\right)+\left(x-m\right)}}& m<x<e\\ 0& x\ge e\end{array}\right\}$$where s, e and m are the starting, ending and midst of linguistic variable as illustrated in Fig. 6.

For x = $y_1$ and x = $y_2$ as shown in Fig. 6, the memberships are:

$$\mu \left(y_1\right)={\displaystyle \frac{y_1-s}{\left(y_1-s\right)+\left(m-y_1\right)},\mu \left(y_2\right)={\displaystyle \frac{e-y_2}{\left(e-y_2\right)+\left(y_2-m\right)}}}$$

The accumulative membership values for a specific crisp input across all linguistic variables has to verify Eq. (5):

(5) $$\sum _{i=1}^j{\mu }_i\left(x\right)\le 1|x\in X$$where j is the number of inferenced linguistic variables.

ELSL derivation

Let us consider that there are j inferenced-linguistic variables in the input set with universe of discourse X. The inferenced-linguistic variables have the following starting, midst and ending points as depicted in Fig. 7:

- s₁, s₂, s₃………, s_j are the beginning values of the ranges in the linguistic variables

- m₁, m₂, m₃ ……..…., m_j are the midst points of the ranges in the linguistic variables

- e₁, e₂, e₃……..…, e_j are the utmost values of the ranges in the linguistic variables

It is required to formulate a general equation that can be used to calculate the membership values in both low and high overlapping scenarios. To achieve this, Eqs. (4) and (6) are derived to calculate the membership values and the accumulative membership value in lowly overlapping situations for a specific crisp input, respectively.

(6) $$\sum _{i=1}^j{\mu }_i\left(y\right)=\left\{\begin{array}{ll}0& x<s_1\\ {\displaystyle \frac{x-a_1}{\left(x-a_1\right)+\left(m_1-x\right)}}& s_1<x<m_1\\ {\displaystyle \frac{e_1-y}{\left(e_1-x\right)+\left(e_1-m_1\right)}+{\displaystyle \frac{x-a_2}{\left(x-s_2\right)+\left(m_2-x\right)}+\dots +{\displaystyle \frac{x-s_j}{\left(x-s_j\right)+\left(m_j-x\right)}}}}& m_1<x<e_1\\ {\displaystyle \frac{e_2-x}{\left(e_2-x\right)+\left(e_2-m_2\right)}+{\displaystyle \frac{x-s_3}{\left(x-a_3\right)+\left(m_3-x\right)}+\dots +{\displaystyle \frac{x-s_j}{\left(x-a_j\right)+\left(m_j-x\right)}}}}& e_1<x<m_2\\ ⋮& \\ {\displaystyle \frac{e_{j-1}-x}{\left(e_{j-1}-y\right)+\left(e_{j-1}-m_{j-1}\right)}+{\displaystyle \frac{x-s_j}{\left(y-s_j\right)+\left(m_j-y\right)}}}& m_{j-1}<x<e_{j-1}\\ {\displaystyle \frac{e_j-x}{\left(e_j-x\right)+\left(x-m_j\right)}}& m_j<x<e_j\\ 0& x>e_j\end{array}\right\}$$

The accumulative membership values for highly overlapping case can be written as in Eq. (7):

(7) $$\sum _{i=1}^j{\mu }_i\left(x\right)=\left\{\begin{array}{ll}0& x\le s_1\\ {\displaystyle \frac{x-s_1}{\left(x-s_1\right)+\left(m_1-x\right)}}& s_1<x\le s_2\\ {\displaystyle \frac{x-s_1}{\mathrm{\Delta }D}+{\displaystyle \frac{x-s_2}{\mathrm{\Delta }D}+\dots +{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}}& a_2<x\le m_1\\ {\displaystyle \frac{e_{1-x}}{\mathrm{\Delta }D}+{\displaystyle \frac{x-s_2}{\mathrm{\Delta }D}+\dots +{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}}& m_1<x\le e_1\&x\le m_2\\ {\displaystyle \frac{e_{1-x}}{\mathrm{\Delta }D}+{\displaystyle \frac{e_{2-x}}{\mathrm{\Delta }D}+\dots +{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}}& m_1<x\le e_1\&x\ge m_2\\ {\displaystyle \frac{e_{2-x}}{\mathrm{\Delta }D}+{\displaystyle \frac{x-s_3}{\mathrm{\Delta }D}+\dots +{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}}& e_1<x\le e_2\&x\le m_3\\ {\displaystyle \frac{e_{2-x}}{\mathrm{\Delta }D}+{\displaystyle \frac{e_{3-x}}{\mathrm{\Delta }D}+\dots +{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}}& e_1<x\le e_2\&x\ge m_3\\ & ⋮\\ {\displaystyle \frac{e_{j-1}-x}{\mathrm{\Delta }D}+{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}& e_{j-2}<x\le e_{j-1}\&x\le m_j\\ {\displaystyle \frac{e_{j-1}-x}{\mathrm{\Delta }D}+{\displaystyle \frac{x-s_j}{\mathrm{\Delta }D}}}& e_{j-2}<x\le e_{j-1}\&x\ge m_j\\ {\displaystyle \frac{e_i-x}{\left(e_i-x\right)+\left(x-m_i\right)}}& e_{j-1}<x\le e_j\\ 0& x\ge e_j\end{array}\right\}$$where ΔD in Eq. (7), has the values as in Eq. (8):

(8) $$\mathrm{\Delta }D=\sum _{i=1}^j\mathrm{\Delta }{D}_i=\left\{\begin{array}{ll}0& x\le s_1\\ \left(x-s_1\right)+\left(x-m_1\right)& s_1<x\le s_2\\ \left(x-s_1\right)+\left(x-s_2\right)+\dots +\left(x-s_j\right)& s_2<x\le m_1\\ \left(e_1-x\right)+\left(x-s_2\right)+\dots +\left(x-s_j\right)& m_1<x\le e_1\\ \left(e_2-x\right)+\left(x-s_3\right)+\dots +\left(x-s_j\right)& s_1<x\le e_2\\ ⋮& \\ \left(e_{j-1}-x\right)+\left(x-s_j\right)& e_{j-2}<x\le e_{j-1}\\ \left(e_j-x\right)+\left(x-m_j\right)& e_{j-1}<x\le e_j\\ 0& x\ge e_j\end{array}\right\}.$$

To address well ΔD sub-ranges in the Eqs. (7) and (8), they have been implicitly modified with Eq. (9), to determine the membership value as a minimum sub-range measured from the midst point to the starting and ending points in each respected linguistic variable. This allows to compute the membership values at any location, no matter it is located before or after the midst point of the linguistic variable, as given in Eq. (9):

(9) $$min\left(\left(x-s_i\right),\left(e_i-x\right)\right)=\left\{\begin{array}{c}x-s_i{s}_i<x<m_i\\ e_i-x{m}_i<x<e_i\end{array}\right\}.$$

Thus, ΔD in Eq. (8) is written as Eq. (10) using the sub-ranges defined in Eq. (9):

(10) $$\mathrm{\Delta }D=\sum _{i=1}^j\mathrm{\Delta }{D}_i=\left\{\begin{array}{c}\begin{array}{c}\sum _{i=1}^j\mathrm{m}\mathrm{i}\mathrm{n}(\left(x-s_i\right),\left(e_i-x\right))s_2<x\le e_1\\ \begin{array}{c}\sum _{i=2}^j\mathrm{m}\mathrm{i}\mathrm{n}(\left(x-s_i\right),\left(e_i-x\right))e_1<x\le e_2\\ ⋮\end{array}\end{array}\\ \begin{array}{c}\sum _{i=j-1}^j\mathrm{m}\mathrm{i}\mathrm{n}(\left(x-s_i\right),\left(e_i-x\right))e_{j-2}<x\le e_{j-1}\\ \end{array}\end{array}\right\}.$$

To deal with the cases where the beginning and ending ranges points are not occurred in the order, a step function $f_i$ is introduced to eliminate the effect of the linguistic variables when they are already exceeded or still not reached yet. $f_i$ is given in Eq. (11):

(11) $$f_i=\left\{\begin{array}{l}1min(\left(x-s_i\right),\left(e_i-x\right)\ge 0\\ 0min(\left(x-s_i\right),\left(e_i-x\right)<0\end{array}\right\}.$$

To nullify the effect of the non-occurrence linguistic variables in ΔD, Eq. (12) is used:

(12) $$f_i\times \mathrm{m}\mathrm{i}\mathrm{n}\left(\left(x-s_i\right),\left(e_i-x\right)\right)=\left\{\begin{array}{ll}min\left(\left(x-s_i\right),\left(e_i-x\right)\right)& x\ge s_i\\ 0& x<s_i\end{array}\right\}.$$

To find the membership value of a crisp input, regardless they are lowly or highly overlapped, the expression $\left|\left(m_i-x\right)\right|$ in Eq. (4) has to be involved in Eq. (10) when there exists a low inferencing of linguistic variable. However, it has to be excluded when the high inferencing is occurred. This task can be accomplished using a step function $N_x$ in Eq. (13):

(13) $$N_x=\left\{\begin{array}{ll}1& x\le s_1\vee x\ge e_{j-1}\\ 0& s_1<x<e_{j-1}\end{array}\right\}.$$

The general equation for calculating the membership values of both low and high inference linguistic variables is written in Eq. (14).

(14) $${\mu }_i\left(x\right)=\left\{\begin{array}{ll}0& x<s_i\vee x>e_i\\ \frac{min((x-s_i),(e_i-x)}{[\sum _{i=i-1}^1{f}_i\times min\left(\left(x-s_i\right),\left(e_i-x\right)\right)]+\left|\left(m_i-x\right)\right|\times N_x}& s_i\le x\le e_i\end{array}\right\}.$$

Implication and Defuzzification of ELSL:

In the implication process, the membership values of the input linguistic variables are implicated on the corresponding output linguistic variables, according to fuzzy inference rules, to determine the output membership values. The implication values in ELSL is calculated using Eq. (15).

(15) $$z_i={\mu }_i\left(x\right)\times Z_i$$where Z_i is the output linguistic variable range, z_i is the result of input implication on the corresponding output. The crisp output is computed as an average value of the implicated values of z_i in all rules using Eq. (16):

(16) $$z={\displaystyle \frac{{\sum }_{i=1}^n{z}_i}{n}}$$where n is the number of implicated rules on the corresponding output linguistic variables.

Implementation of the proposed AI

To identify whether the spots in Fig. 4F are corresponding to diseases or noise, we use two input sets:

1^st one is the different of intensity between the spots and the average of leave color intensity. If the different is high, it means that this is possible of defects. This is due to that the color of the disease spot is has intensity which looks totally different than other spots that maybe occur due to the shortage in chlorophyll. If the different is high, it means that this is possible of defects.

Intensity = {low, medium, high}

To accomplish the calculation of intensity, the following steps must be done:

(a) The location of the spots is determined from the result of image processing as in Fig. 4F.

(b) We should find the intensity of these spot areas in the original image as shown in Fig. 3B.

(c) The average of the original image intensity is calculated from image in Fig. 3B.

(d) The difference in intensity between step (b) and (c) is considered as the crisp input for intensity input set.

The 2^nd Input set is the area of spots, where the big areas are possible to be defects:

Area = {small, Medium, big}.

The area is calculated for the spots with black pixels in the result of image processing algorithm as shown in Fig. 4F, which will be supplied as the crisp input for Area input set.

Figure 8 shows the input sets of ELSL.

The output set of ELSL is the classification of the tested spots state into a defective or non-defective spots. In fact, we preferred to select binary decisions for the spots either defect or noise, as if we increase the linguistic variables of the output to include more situations such as possible noise and possible defects, it is required during defuzzification process to grab these new linguistic variables (possible noise and possible defect) into either noise or defect decisions, respectively:

Spot State = {Defect, Noise}.

The ranges of linguistic variables in the input and output sets are given in Table 2. The ranges have been selected based on empirical trials for some spots that we know its state whether defect or noise. The selected ranges cover all possibility cases.

The membership functions in ELSL is shown in Fig. 9.

The rules for implicating the input on the output are listed as follows:

• 1)

  If Intensity is Low and Area is Small THEN Spot_area is Noise

• 2)

  If Intensity is Low and Area is Big THEN Spot_area is Noise

• 3)

  If Intensity is Medium and Area is Small THEN Spot_area is Noise

• 4)

  If Intensity is Medium and area is Medium THEN Spot_area is Defect

• 5)

  If Intensity is high and Area is Medium THEN Spot_area is Defect

• 6)

  If Intensity is High and Area is High THEN Spot_area is Defect

The ELSL steps and procedures are illustrated in Fig. 10.

The decision of ELSL on the spot area is highlighted in red circle if it is defect and no color if there it is noise. Figure 11 shows the resulted image after implementing ELSL.

The final results of video pre-processing, processing and post-processing with ELSL algorithms are shown in Fig. 12. When the red circles are detected on the leaves, the pump will spray the pesticide on the leaves, however, due to the effect of such pesticide to the people in lab, we, instead, spray water to the leaves.