Mobile Networks and Applications

https://doi.org/10.1007/s11036-023-02256-x

Artificial Intelligent IoT‑Based Cognitive Hardware for Agricultural

Precision Analysis
An‑Chao Tsai1 · Anirut Saengsoi1

Accepted: 2 October 2023

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
Traditionally, farmers have used human resources for productivity, which could be more efficient and manageable. Farmers
have been trying to improve agricultural efficiency and optimize productivity with limited cultivation resources. This study
presents an intelligent circulating agricultural farming system that implements monitoring, alerting, automation, and envi-
ronmental prediction functions. With various sensors, the system constantly collects data on climate conditions, including
1) temperature, 2) humidity, and 3) soil content. Furthermore, we integrated machine learning to forecast the requirements
for temperature, humidity, and fertilizer, the most significant growth factors for planting. As a result, the proposed system
successfully controlled the cultivation more precisely. An extensive experimentation was conducted on specific crops and
environmental conditions to evaluate the proposed model's efficacy. The findings of this research contribute to a deeper
understanding of the potential benefits of the proposed integrated system. The results demonstrate how the AI, IoT, and
cognitive hardware framework can significantly enhance agricultural precision, ultimately leading to more sustainable and
efficient crop production practices.

Keywords IoT · Precision Circulating Agriculture · Bean Sprouts · Frass of Hermetia Illucens Larvae

1 Introduction efficiency [6]. These technologies rely on the IoT and plat-
form development. The platform aims to increase produc-
Human labor in agriculture is one of the most important tivity efficiency and quality and minimize climate change
and rapidly transforming the quantity and quality of agri- impact. Specifically, sensors collect data and monitor soil
cultural products [1, 2]. Therefore, traditional methods of quality, weather conditions, crop growth, etc. Through the
agriculture can be replaced by intelligent farming concepts communication between sensors, controllers, and actuators,
to help farmers increase the quantity and quality of their an IoT and automatic irrigation, fertilization, and climate
productivity with limited cultivation resources. Innovative control system helps provide a better growth rate for efficient
Farming tools could collect much data, such as soil quality, crop productivity distribution. Big Data has been applied to
weather, and plant growth. Moreover, these smart sensors smart farming, providing users real-time predictive insights
can operate and control the amount of water, fertilizer, and into farming operations to make informed decisions [7].
other factors through operation between the sensor and con- An intelligent greenhouse can help users manage
troller [3–5]. Technologies such as the Internet of Things resources reliably and efficiently, with high productivity,
(IoT), Big Data, Artificial Intelligence, Cloud Computing, especially when using a remote-controlled farming sys-
Remote Sensing, etc., can significantly improve agriculture tem. This significantly increases the capacity of agriculture
using technology, which will help farmers develop more
productive soils [8]. The farms with low productivity may
* An‑Chao Tsai
actsai@nptu.edu.tw confront many problems, such as unstable water volume,
lack of water, and low efficiency of irrigation systems [9].
Anirut Saengsoi
saengsoi.a@gmail.com Based on a Smart Greenhouse System, the present research
implements an innovative greenhouse's monitoring, auto-
1
International Master Program of Information Technology matic system, alerting, and data analysis features [10]. The
and Application, National Pingtung University, proposed system consists of three subsystems. First, it is
Pingtung 900, Taiwan, Republic of China

13
Vol.:(0123456789)
 Mobile Networks and Applications

an automated monitoring system. It presents data such as collect environmental attributes in the field area. The data
temperature, humidity, and fertilizer. Second, it automati- were collected through an Arduino, which was connected
cally controls irrigation and climate based on sensor data to several sensors and then transmitted to the cloud-based
and alerts the user of the conditions. server. In [11, 15], increased demand for organic vegetables.
Finally, irrigation and automation are implemented As a result of their research, they developed a method to
using machine learning techniques. In conventional farm- maintain an environment conducive to longer, more efficient
ing, farmers use human labour to handle their productivity, growth seasons and increased productivity. They demon-
which leads to inefficient resource use. Imaginative farming strated a comprehensive IoT based on the intelligent green-
concepts help farmers increase the quantity and quality of house system that implements monitoring, alerting, cloud
productivity with limited cultivation resources. The study storage, and automated disease prediction systems. Environ-
aims to develop monitoring and automatic irrigation sys- mental conditions, such as temperature, humidity, and soil
tems using IoT to control climate and irrigation in intelligent moisture, are monitored to increase productivity and quickly
greenhouses [11, 12]. Additionally, the growth relationship redress under poor conditions. In [16], it used IoT-based
between plants and the Frass of Hermetia Illucens Larvae, remote-sensing devices to receive information about the
which affect the growth of mug bean sprouts, have been environment. The system contains several remote-sensing
studied to improve production efficiency through precision devices that transmit data, including temperature, humidity,
farming using technologies to increase crop productivity. and soil moisture. The actual data were obtained from the
The paper introduces the importance of precision agriculture field area. The LM35 and DHT-22 sensors were used for
and the role of technology in optimising agricultural pro- temperature and humidity, respectively.
cesses. Discuss the challenges faced in achieving precision
analysis in agriculture. In the following, we also discuss the 2.2 Forecasting data techniques
application of the amalgamation of IoT (Internet of Things)
and AI (Artificial Intelligence) in agriculture toward smart Various forecasting techniques are discussed in the litera-
farming with data predictive analysis, which focuses on all ture; however, this section mainly discusses forecasting date
aspects of the productivity of crops. techniques used in agriculture. In [17], proposed combining
the aspects of crop growth and support vector regression
(SVR) models with the comprehensive groundwater varia-
2 Literature review tion in several irrigation thresholds throughout the growing
season. Available water was used to simulate the calibrated
2.1 IoT in agriculture daily irrigation requirements. The study found that higher
groundwater volumes significantly affected the groundwa-
IoT in agriculture includes remote sensing, dashboards, ter available for irrigation. Timely suggestions and automa-
alerting, and controlling to monitor agricultural activity tion in agriculture can connect farmers, knowledge, experts,
precisely. In [10], IoT devices for intelligent soil farming at sensors, actuators, and other factors by evolving several
affordable prices. The author demonstrated how to configure technologies [18]. Integrated rainfall forecasting, an essen-
the connections between the sensors and actuators for the tial parameter of interest in agriculture in many countries,
desired farming and effectively maintain devices in preci- assumes that rainfall data are provided to forecast rainfall
sion farming. This enables precision farming by allowing using several forecasting techniques, such as Holt-Winters
farmers to automatically and remotely manage their systems, and Auto Regression Integrated Moving Average (ARIMA).
including irrigation, pesticide control, and fertiliser. In [13, These algorithms provide an idea of the volume of water that
14], discussed the importance of intelligent greenhouse com- may be required from rainfall forecasting information. As
ponents monitoring climate conditions or the environment a result, a forecasting algorithm using historical data from
inside farms. Real-time monitoring can help farmers pro- 40 days of rainfall data was used to forecast rainfall over
tect against adverse conditions and avoid disease infection. the subsequent ten days, and the ARIMA model exhibited a
In [14], deployed the monitoring system with a sensor on greater explanatory power and wider forecasting boundary
an embedded system, which integrates sensor parameters, with 95% confidence intervals.
including soil moisture, temperature, and humidity, with a In [7], model of autocorrelation (ACF) and partial auto-
Raspberry Pi and Arduino board to monitor the soil mois- correlation (PACF) functions that were proposed to test the
ture, temperature, and humidity on a cloud interface and pro- model order, which included p, d, and q. The lowest akaike
vide ideal conditions for the plants inside the greenhouses. information criterion (AIC) score determines the best model
In [8] and [5], deployed IoT devices with several sen- order. In addition, the augmented Dickey–Fuller (ADF) test
sors, such as soil moisture (EC-1258), soil temperature was used to check the stationarity of historical data. In [16],
(DS18B20), and temperature and humidity (DHT11), to long short-term memory based on a neural network was used

13
Mobile Networks and Applications

to test an intelligent irrigation system that contains several so it is one of the areas where technological intervention can
sensors. A neural sensor was used in place of a physical improve the lives of farmers. Greenhouse conditions, such
sensor. The experiment showed highly accurate predictions, as local climate and environmental parameters, can be meas-
especially for parameters within an acceptable range. In [19], ured using sensors, with many nodes capable of performing
considered rainfall as an essential factor and considered rain- remote sensing, actuator control, and communication tasks.
fall and temperature records from 1953 to 2070, as well as In [22], developed intelligent irrigation systems with four
seasonal variations. The analysed data showed a substantial main components. In the first component, a sensor node was
upward trend with a slight increase in rainfall between 2021 placed in the field to measure parameters such as tempera-
and 2050 and 2050–2070. This represents rainfall patterns ture, humidity, soil moisture, and soil temperature. In addi-
with few substantial variations in various climates and crop tion, actuator nodes were connected to a sprinkler to control
seasonality. In contrast, the temperature pattern displayed the amount of water. Weather stations or rain gauge sensors
considerable variations across most seasons. were used to obtain weather and rain-level forecast reports.
The second component was used for data forwarding and
2.3 Automatic system in agriculture receiving from the sensors to the gateway. The third compo-
nent was soil moisture prediction, which was used to predict
Automatic systems in agriculture include automated irri- the soil moisture level for the subsequent day based on the
gation, temperature and humidity controls, and fertiliser climate and previous soil data. Furthermore, a classifica-
amounts to reduce human labour costs and increase pro- tion model and crop data were used to estimate the irriga-
ductivity. In [20], developed a prediction of the irrigation tion system schedules. The last component was the farmer's
system in the field area by using a sensor to collect ground agreement with or alteration of the information.
parameters, such as soil moisture, soil temperature, and cli- In [23–25], proposed a system that provides ideal crop
mate conditions according to weather forecasting from the growth and management efficiency under controlled con-
Internet. Sensor nodes were also involved in the field area. ditions for crop cultivation; however, the system requires
The system has been completely developed and deployed in accurate prediction. This study proposes a long-term pre-
the field area, where the sensor node is wirelessly collected dictor equipped with a collection data system, data analy-
over the web services and provides a decision support sys- sis, and environmental information processing for multiple
tem and real-time visualisation information insights based environmental factors in an intelligent greenhouse, including
on data analysis and weather forecasting. temperature, humidity, C ­ O2, light intensity, etc. The data are
In [21], highlighted the significant effect climate change then automatically transmitted to the server through commu-
has had on agricultural activities and productivity. However, nication at regular intervals and forwarded into the database
recent technologies have revolutionised and provided intel- system. After data analysis and model training, the predic-
ligent systems, such as real-time monitoring, controlling, tion results and recommendations are returned in a period
and visualisation systems for various farm operations. Arti- that more precisely reflects the situation of the greenhouse
ficial Intelligence and IoT can be employed to develop irriga- management system. The experimental greenhouse and inte-
tion systems, fertilisers, and greenhouse cultivation. Water rior scenes were equipped with various IoT devices.
usage efficiency is the first significant reason; hence, the
introduction of technology to manage efficient water usage
has commenced. The development of the IoT and decision- 2.4 Machine learning techniques employed
making support systems have enabled intelligent automa- for smart agriculture
tion of entire irrigation systems. The primary data source
was forwarded to the system for analysis, and the database 2.4.1 For temperature forecasting
system was used to collect historical data that could be used
to train the model by using the classification technique for a) Time-Series Models: Techniques such as Autoregres-
unknown data classification. The second primary reason for sive Integrated Moving Average (ARIMA) and Seasonal
the development of technologies was the low productivity Decomposition of Time Series (STL) can be applied to
of the field area. Soil testing was conducted before the addi- capture temporal patterns in temperature data. More
tion of fertiliser, and required crop nutrients were revealed. advanced models like Seasonal ARIMA (SARIMA) and
Today, organisations have integrated technology into farm- Seasonal Exponential Smoothing (ETS) can account for
ing. IoT devices can help farmers measure fertiliser amounts seasonal fluctuations.
using an NPK sensor. The last primary reason is that main- b) Recurrent Neural Networks (RNNs): RNNs, par-
taining a stable environment inside the greenhouse due to ticularly Long Short-Term Memory (LSTM) networks,
the involvement of many parameters is tedious. However, are well-suited for time-series forecasting. LSTMs can
the environment is an essential factor in crop productivity, capture long-range dependencies and handle sequential

13
 Mobile Networks and Applications

data effectively, making them suitable for temperature or Root Mean Squared Error (RMSE). Cross-validation can
prediction. help estimate the model's generalisation performance on
c) Gradient Boosting Regressors: Algorithms like unseen data.
XGBoost and LightGBM can capture complex relation-
ships in temperature data. They can handle nonlinear
patterns and be optimized for specific forecast horizons. 3 Materials and method

2.4.2 For humidity forecasting The methodology employed in this research is multifaceted,

combining expertise from agricultural sciences, engineering,
a) Time-Series Models: Like temperature forecasting, and data analytics. The study's approach can be summarised
time-series models like ARIMA, SARIMA, and ETS as follows:
can capture temporal patterns in humidity data.
b) Recurrent Neural Networks (RNNs): LSTM networks a) IoT-enabled Data Collection: IoT sensors were
can also be applied to humidity forecasting. They excel deployed strategically across the experimental farm-
at modelling sequential dependencies and can handle the land to collect real-time data on critical environmental
dynamics of humidity variations. variables. These sensors captured temperature, humid-
c) Support Vector Machines (SVMs): SVMs can be used ity, soil moisture, and light intensity information, offer-
for humidity prediction by learning patterns in the data ing insights into the dynamic conditions affecting crop
and finding appropriate hyperplanes for regression tasks. growth.
b) AI-driven Data Processing: The collected IoT data
2.4.3 For fertilizer requirement prediction underwent a robust data preprocessing phase to ensure
quality and consistency. Advanced AI algorithms,
a) Regression Models: Linear regression, Ridge regres- including machine learning and deep learning tech-
sion, and Lasso regression can predict fertilizer require- niques, were employed to analyse the data and extract
ments based on input features such as soil character- meaningful patterns, trends, and correlations.
istics, crop type, historical data, and environmental c) Cognitive Hardware Integration: A significant inno-
conditions. vation in this research is the integration of cognitive
b) Random Forest and Gradient Boosting Regressors: hardware components. These components simulate
Ensemble methods like Random Forest and gradient human-like cognitive functions, enabling the system to
boosting can capture nonlinear relationships between perform intricate pattern recognition, anomaly detection,
input variables and fertilizer requirements. They can and adaptive decision-making based on complex data
handle complex interactions and feature importance inputs.
ranking. d) Precision Analysis and Insight Generation: The
c) Neural Networks: Feedforward neural networks can fusion of AI, IoT, and cognitive hardware led to a com-
be designed to predict fertilizer requirements. However, prehensive analysis of the collected data. Insights into
having a sufficiently large and clean dataset to train neu- optimal growing conditions, resource requirements, and
ral networks effectively is essential. environmental impacts were generated in real time, ena-
bling proactive decision-making.
2.4.4 For data preprocessing and feature engineering e) Experimental Validation: Experimental trials were
conducted across diverse crops and environmental con-
Data preprocessing and feature engineering are crucial in ditions to validate the effectiveness of the proposed
enhancing prediction accuracy regardless of the machine framework. The outcomes of the AI-driven approach
learning technique. Feature engineering involves select- were compared with conventional agricultural practices
ing relevant input variables, scaling data, handling miss- to showcase the framework's advantages
ing values, and creating meaningful features. Time-related
features, lagged variables, and moving averages can benefit 3.1 Proposed architecture
time-series data.
The proposed system architecture, IoT for Precision Cir-
2.4.5 Evaluation and model selection culating Agriculture based on Bean Sprouts and Frass of
Hermetia Illucens Larvae, is shown in Fig. 1. The previ-
It is essential to evaluate the performance of different ous section discussed IoT and machine learning research
machine learning models using appropriate metrics such as in intelligent agriculture. In this section, we introduce the
Mean Absolute Error (MAE), Mean Squared Error (MSE), proposed system, which contains the following subsystems:

13
Mobile Networks and Applications

Fig. 1  The proposed system architecture

1) a monitoring and alerting system for climate conditions, The preprocessing steps include noise removal, data smooth-
2) an irrigation and automation system, and 3) a system for ing, and outlier detection.
forecasting parameters. The system architecture for IoT-
based Precision Circulating Agriculture centered around Crop management and environmental control The system
cultivating Bean Sprouts and utilizing Frass from Hermetia makes informed decisions about crop management and envi-
Illucens Larvae encompasses a comprehensive approach to ronmental control based on the analyzed data. Automated
optimizing agricultural processes through real-time data col- actuators are employed to adjust parameters such as irri-
lection, analysis, and decision-making. Figure 1 illustrates gation levels, light intensity, and temperature to optimize
the various components and their interactions within the growing conditions for Bean Sprouts.
architecture.
Alerts and notifications The system incorporates alert mech-
IoT sensors and data collection The system's foundation anisms that trigger notifications to farmers or stakeholders
consists of an array of IoT sensors strategically placed within in case of deviations from desired conditions. Alerts could
the cultivation environment. These sensors capture essen- include notifications about extreme temperature fluctuations,
tial data such as temperature, humidity, light intensity, soil low soil moisture, or irregularities in larvae farming.
moisture, pH levels, and nutrient content. The IoT sensors
continuously collect data and transmit it wirelessly to a cen- User interfaces and dashboards Farmers and stakeholders
tral data repository. access the system through user interfaces and dashboards,
which provide real-time visualizations of data and insights.
Central data repository and cloud integration The collected The interfaces allow users to monitor crop growth, larvae
data is sent to a central data repository located on-site or farming, and overall system performance.
hosted in the cloud. Cloud integration allows for scalable
storage, real-time data access, and remote monitoring by Data‑driven insights and recommendations The system
authorized users. generates data-driven insights and recommendations for
improved crop yield, resource efficiency, and larvae farming
Data processing and analysis The collected data is subjected productivity. Users can access historical trends and forecasts
to data preprocessing to ensure accuracy and consistency. to make informed decisions.

13
 Mobile Networks and Applications

Fig. 2  Schematic of Arduino

wired circuit

Fig. 3  Schematic of Raspberry

Pi wired circuit

Figure 2 depicts the schematic representation of the Table 1  Experimental setting for collecting data
Arduino Wired Circuit used to design the proposed system. Temperature Humidity Amount
Figure 3. shows the schematic representation of the Rasp-
berry Pi Wired Circuit for interfacing the proposed system 25 60 Frass 1 g
with the stakeholders. Researchers designed experiments 25 60 Frass 3 g
to collect data considering different factors, including the 25 90 Frass 1 g
Frass of Hermetia Illucens Larvae, traditional fertilizer, 25 90 Frass 3 g
and no fertilizer to observe a suitable factor for a mug bean. 30 60 Frass 1 g
A list of factors is shown in Table 1. In this study, authors 30 60 Frass 3 g
performed 18 experiments in total; each was performed 30 90 Frass 1 g
twice for each setting, which amounted to 86,401 datasets 30 90 Frass 3 g
for each setting.
Insufficiently treated food waste is considered an environ-
mental threat in many countries, necessitating special atten- 3.2 Irrigation and automation system
tion to waste treatment activities involving Hermetia Illucens
Larvae that adhere to circulating economic concepts. The IoT component includes a network of sensors placed
strategically in the agricultural environment. These sensors

13
Mobile Networks and Applications

can measure various parameters such as soil moisture levels, to forecast the parameters to prepare a controlling actuator
temperature, humidity, light intensity, weather conditions, schedule [26]. The setting of the temperature and humidity
and crop health indicators. The sensors continuously col- sensor (Arduino Wemos D1 Refer to Figure S2); soil con-
lect real-time data from the field and transmit it to a central tent sensor used to collect soil fertiliser (into InfluxDB in
data repository or a cloud-based platform. The data col- Raspberry Pi4, refer to Figure S3) conductivity, and soil pH.
lected by IoT sensors is transmitted wirelessly to a central MQTT was used to publish and subscribe modules to gather
system using Raspberry Pi 4 and Arduino. The integrated and transfer the data, as shown in Fig. 4. Arduino sends the
data is fed into AI algorithms responsible for data analysis data to the Influx DB server, and these parameters follow
and decision-making. These algorithms can be trained for the experiment. They will appear on the dashboard system,
various tasks, such as predictive modelling, anomaly detec- which is set up on Grafana as an alert sent to the farmer's
tion, and yield forecasting. Cognitive hardware enhances smartphone to increase or decrease the factors as desired.
the AI system's ability to reason, learn, and make complex MQTT is used to publish and subscribe to messaging
decisions by emulating aspects of human cognition, such transport for connecting remote devices with data more effi-
as pattern recognition and adaptive learning. Based on the ciently and lightweight, with minimal bandwidth. MQTT
analysis performed by the AI algorithms, decisions are made comprises a broker (server), clients (publisher/subscriber),
regarding irrigation and automation. The decisions made and topics. A publisher publishes a message to a receiver
by the AI algorithms trigger actions in the physical world via a topic, and the subscriber receives a message via a
through actuators. Actuators are devices that control pro- topic. However, the broker mediates between the publisher
cesses based on the decisions made by the AI system. In and subscriber, and the topic is a data tag. Figure 5 shows
irrigation, actuators control valves that regulate water flow hardware diagram for (a)DHT22 Temperature and Humid-
to different field areas, adjusting irrigation based on real- ity, (b) Arduino Wemos D1, which connects to sensors. (c)
time conditions. The AI algorithms consider soil moisture Soil content sensor, which consists of NPK conductivity and
levels, weather forecasts, historical data, and crop health to pH; (d) Raspberry Pi 4, which is the database server and
determine optimal irrigation schedules and other automation dashboard server; and software result for (e) InfluxDB is a
actions. The decisions made by the AI algorithms trigger time-series database and open source; (f) Grafana dashboard
actions in the physical world through actuators. Farmers and is the source.
agricultural experts can monitor the system's operations and
receive real-time updates through user interfaces, mobile 3.4 Forecast parameters for controlling farmwork
apps, or dashboards.
Box and Jenkins invented the ARIMA model in the early
3.3 Climate condition monitoring and alerting 1970s [22]; the ARIMA model contains three orders:
system autoregressive (AR) of p, d, and q orders, where p, d, and
q, respectively represent the order of auto-regression,
This section aims to monitor the climate parameters and degree of trend difference, and order of the moving aver-
alert the farmer when the parameters deviate from their typi- age. Figure 6 shows ARIMA Framework that comprises
cal values. Continuous monitoring of the parameters was four steps: 1) Stationary Sequence, 2) Identification Model
performed in a database. After the device is connected to Order, 3) Diagnosis and Estimation, and 4) Forecasting
the Internet, the dashboard system allows devices to show and Evaluation. The first step is to sequence the historical
the parameter values on the dashboard, and alerts are sent to data as stationary. If historic data represent a non-station-
the farmer’s smartphone to observe the factors used for the ary time series, it should be modified using the differen-
experimental setting. This section uses the ARIMA Model tiation method. The original sequence diagram suggests

Fig. 4  Messaging transport via MQTT broker

13
 Mobile Networks and Applications

Fig. 5  a DHT22 Temperature and Humidity, b Arduino Wemos D1, server and dashboard server; e InfluxDB is a time-series database and
which connects to sensors. c Soil content sensor, which consists of open source; f Grafana dashboard is the source
NPK conductivity and pH; d Raspberry Pi 4, which is the database

Fig. 6  ARIMA framework

that stationarity should be checked. Statistical software between a time series and its lagged values. It helps us
can process differences and log transformations to obtain identify the potential order of the ARIMA model's Mov-
a stationary time series. The second step is the order of ing Average (MA) component. The ACF plot displays
the identification model. The Auto-correlation Function the auto-correlation values for various time intervals. We
(ACF) plot is valuable in understanding the correlation can identify the lag values where auto-correlation drops

13
Mobile Networks and Applications

significantly by analysing the ACF plot. This aids in deter- The autoregressive (p) assumes Yt is a linear function of the
mining the appropriate order of the MA component in the initial values and is given by Eq. 4.
ARIMA model. The Partial Auto-correlation Function
(PACF) plot further refines our understanding of the cor-
Yt = 𝛼1 Yt−1 + 𝜀t (4)
relation between a time series and its lagged values. It Each observation contained a random component (ran-
helps us identify the potential order of the ARIMA model's dom shock, 𝜀 ) and a linear combination of the preceding
Auto-Regressive (AR) component. observations. 𝛼1 in this equation is the self-regression coef-
AFTER ACCOUNTING FOR THE EFFECTS OF ficient. The integrated (d) may have been affected by some
SHORTER LAGS, the PACF plot shows the partial auto- cumulative effects. While short-term values may fluctuate
correlation values for different lags. A significant positive with many contingencies around the average value, the level
partial auto-correlation at a specific lag indicates a direct of the series over the long term will not change.
relationship between current and past values at that lag. ACF These sequence stationarity differences for an inte-
and PACF plots were used to estimate the ARIMA model grated process are essential during the statistical analysis of
order. Subsequently, the ARIMA model parameters could sequence data. An integrated approach is the archetype of a
be defined by both skills and experience by observing the non-stationary series. A differentiation of order 1 assumes
ACF and PACF plots. After ACF and PACF processing, an that the difference between the two continuous values of Y
ADF test was conducted for hypothesis testing to confirm is constant. The integrated process is defined in Eq. 5.
stationarity [27]. The unit root in the characteristic equa-
tion represents the successive check of data stationarity. The Yt = Yt−1 + 𝜀t , (5)
model for testing the ADF is represented in Eq. 1.
where the random perturbation 𝜀t is white noise. The mov-
𝜕Yr = μ + βt + ρYt − 1 + 𝜕1Yt − 1⋯ + 𝜕1pYt − p + et., ing average (q) linearly combines the latest disturbances and
(1) previous perturbations. The moving average order represents
where µ, p, and β denote the steady value, auto-regression the number of the most recent periods embedded in the latest
order, and trend, respectively. Moreover, it represents a value [16] and can be determined using Eq. 6.
sequence of zero mean and unit variance independent of Yt = 𝜀t − 𝜃1 𝜀t−1 (6)
standard random variables. After the ADF test has been
conducted, the p-value is used to decide whether to reject After training the models, ARIMA was applied to the
or accept the null hypothesis. The null hypothesis is true Raspberry Pi, as shown in Fig. 7, to collect the forecast
at a 95% confidence level, p < 0.05, the value is significant parameters into the database. These forecast parameters
enough to reject the null hypothesis, and the time series is were set as planning to control actuators, such as the time to
stationary [28]. The ARIMA model was used for residual turn on a fan, water heater, or fertiliser.
checking to evaluate the model order using the Ljung (Q)
test. The residual error must be random (significance level,
p > 0.05). The tentative model is considered insufficient if 4 Results and discussion
the Q-statistic is lower than 0.8. In effect, the model should
be refitted. Furthermore, the lowest Bayesian information A monitoring and notification system that uses Raspberry
criterion of Schwarz (BIC) and Akaike Information Crite- Pi 4 as a server to export the dashboard system and uses
rion (AIC) was used to determine a suitable model order Arduino to connect to sensors was developed in this paper.
[12], and its residual was white noise. Table 2 shows the results of the initial setup of the tempera-
AIC: The AIC and BIC formulation is shown in Eqs. 2 ture at 25 °C and humidity of 60%, and fertiliser in 3 g shows
and Eq. 3 the best parameter. Collecting these parameters over 60 days
amounted to 86,401 for each setting. After cleaning the data,
AIC = −2log(maximumlikelihood) + 2k (2)
we split the data into training and test parts.
Figure 8(a) shows that hardware diagram for Arduino
klogn connected to sensors and b) hardware diagram for sensor
BIC = −2log(maximumlikelihood) + , (3)
n was connected to Arduino. Figure 9 shows temperature (yel-
where k is the independently adjusted number of parameters low line) versus humidity (green line).
and n is the total number of data points. The forecasting
ability was evaluated by comparing the forecasted values 4.1 Experimental environment
with actual values. The ARIMA approach requires many
observations to define a suitable model fit for a data series. The proposed system runs in the experimental environment
The ARIMA model is labelled the ARIMA model (p, d, q). (Tables 3, 4 and 5).

13
 Mobile Networks and Applications

Fig. 7  Applying ARIMA model

Table 2  The result of the Temp 25 Temp 25 Temp 30 Temp 30

experimental setting Humid 60 Humid 90 Humid 60 Humid 90
Fresh(g) Dry(g) Fresh(g) Dry(g) Fresh(g) Dry(g) Fresh(g) Dry(g)

No fertiliser 0.25 0.07 0.23 0.07 0.22 0.072 0.22 0.105

BSF 1 g 0.33 0.063 0.25 0.08 0.29 0.082 0.19 0.08
BSF 3 g 0.35 0.085 0.38 0.11 0.30 0.090 0.186 0.086
Normal 1 g 0.28 0.063 0.26 0.07 0.23 0.080 0.19 0.078
Normal 3 g 0.31 0.079 0.29 0.07 0.25 0.073 0.185 0.079

Fig. 8  a Arduino connected to sensors; b That sensor was connected to Arduino

�
The performance of the proposed models was evalu- ∑n 2
(yi − ̂
yi )
ated by calculating mean Absolute Error (MAE) and root- RMSE = i=0 (7)
mean-square error (RMSE). The Root Mean Square Error N
(RMSE) is the standard deviation of the prediction errors
The Mean Absolute Error (MAE) is a measure of errors
in the time-series analysis. The RMSE shows the extent to
between the actual and predicted data (Ahmed et al. 2021),
which the regression point of the actual data is from the
as shown in Eq. 8
predicted data, as shown in Eq. 7.

13
Mobile Networks and Applications

Fig. 9  Dashboard from the system showing temperature (yellow line) and humidity (green line)

Table 3  Environment collecting data platform model are shown in Figure, and the evaluation results are
in Table 6.
Data collection Operating System Raspbian GNU/Linux 11
platform This Fig. 10 presents the experimental results of apply-
CPU Quad-core Cortex-A72
(ARM v8) ing an ARIMA (Auto-regressive Integrated Moving Aver-
RAM 4GB age) model to a temperature test dataset. This Diagnostic
Programming Language Python tests may include examining the residuals (the differences
between predicted and actual values) for auto-correlation
and white noise properties. The absence of significant auto-
Table 4  Arduino’s specification correlation in the residuals is essential. The x-axis represents
the time duration of the test at a span of 2 days, while the
Data collection platform Chipset ESP-8266EX
y-axis indicates the recorded temperature values in degrees
Clock Speed 80MHz/160MHz
Celsius. This Fig. 10 shows that variation of heat tempera-
Flash 4M bytes
ture for each span of 2 days intervals for actual data and
Digital I/O Pins 11
predicated data.
Analog Input Pins 1
In this Fig. 11, experimental results of the humidity test
Programming Language C +  +
dataset is shown for ARIMA model. This figure is a visual
representation of various aspects of the ARIMA modeling
process and its performance in forecasting humidity levels.
The x-axis represents the time duration of the test at a span
Table 5  The experimental environment platform of 2 days experiment, while the y-axis indicates the recorded
Data collection platform Operating System Windows 11 humidity values. This Fig. 11 shows that variation of humid-
CPU i5-10300H ity for each span of 2 days intervals for actual data and predi-
RAM 16GB cated data in ARIMA model.
GPU Nvidia GTX 1650 In this Fig. 12, experimental results of the Soil-Moisture
Programming Language Python test dataset used for predict the soil moisture level by ARIMA
model. The figure could include a comparison between the
ARIMA model's soil moisture forecasts and the actual soil
1 ∑n | moisture values for a specific period beyond the historical
MAE = yi ||,
y −̂ (8) data. This comparison helps assess the model's predictive
N i=0 | i
performance in real-world scenarios. The x-axis represents
where yi denotes the actual value, ̂
yi represents the forecasted the experiment's time duration, while the y-axis denotes
value, and N represents the number of observations. The the recorded soil moisture values.. This Fig. 12 shows that
experimental results of different parameters in the ARIMA

13
 Mobile Networks and Applications

Table 6  RMSE and MAE ARIMA SVR LSTM

parameters comparison for three
models RMSE MAE RMSE MAE RMSE MAE

Temperature 0.208 0.101 0.329 0.197 0.217 0.142

Humidity 0.705 0.246 2.810 2.396 0.895 0.609
Soil Moisture 1.834 1.264 4.514 3.282 5.820 4.195
Conductivity 10.098 3.559 33.343 31.504 11.014 7.027

Fig. 10  Experimental results of

the temperature test dataset

Fig. 11  Experimental results of

humidity test dataset

variation of soil moisture level for every 2 days intervals for practical implications of the results. The x-axis represents
actual data and predicated data in ARIMA model. the duration of the experiment, while the y-axis indicates
This Fig. 13 presents the experimental results obtained the recorded conductivity values. This Fig. 13 shows that
from the conductivity test dataset, explicitly focusing on difference on conductivity level for every 2 days intervals for
evaluating the ARIMA model's performance. Figure may actual data and predicated data in ARIMA model.
conclude with a summary or discussion of the ARIMA mod- The MAE and RMSE of the ARIMA model values were
el's performance. It could highlight the model's strengths, smaller than those of the other models. In addition, the fore-
limitations, areas for potential improvement, and the casted value obtained by ARIMA can better fit the actual

13
Mobile Networks and Applications

Fig. 12  Experimental results of

soil-moisture test dataset

Fig. 13  Experimental results of

the conductivity test dataset

value change trend. Therefore, the ARIMA model is adapted Table 7  Cost device analysis
to a light version and applied to the embedded system to No. Device Quantity Price (NT)
prove that our system can be designed and developed with
low-cost devices, the costs of which are shown in Table 7, 1 Raspberry Pi 4 (RAM four GB) 1 4,968
and that the ARIMA model is a suitable model to forecast 2 Arduino Wemos D1 2 119
the parameters of sensor data. 3 DHT 22 1 250
4 Soil Moisture Sensor 1 140
4.2 Discussion 5 Relay Switch 4 channel 1 170
6 RS485 module 2 20
This section runs the forecasting models on the Raspberry 7 RS485 Soil pH and EC 1 5,142
Pi 4. The scripts running in the background query the his- 8 RS485 Soil NPK 1 3,500
torical data from the database to forecast parameters 30 min Total Cost 15,281
ahead. Then, the written script will estimate the time required
to control the actuators, which means the scripts will estimate
when those values deviate and order the actuators to work. fertilizer and conductivity are used to observe the amount of
The actuators' time estimations were validated from forecasted fertilizer set at a condition lower than 400 µs/cm. The tem-
parameters in the intelligent greenhouse and were scheduled perature and humidity are 25 °C and 90%, respectively. This is
by the model successively, starting with identical initial con- related to the results of the experimental setting for a suitable
ditions. The soil moisture range was set by an experiment to environment for our plants. Table 8 depicts the advantages of
lower than 50 per cent. Irrigation is required, and a range of the proposed system over the traditional one.

13
 Mobile Networks and Applications

5 Conclusions and recommendations

Integrating AI algorithms and cognitive hardware enables precise monitoring

• AI algorithms process vast amounts of data to generate actionable insights.

tion, minimises chemical inputs, and contributes to sustainable agriculture.

Cognitive hardware automates pattern recognition and decision-making pro-
IoT sensors continuously collect real-time data, enabling instant detection of

By optimising resource usage, the proposed system reduces water consump-

• The system's scalability allows easy deployment across various field sizes
The proposed system has completely implemented a com-

• Adaptive decision-making ensures performance in diverse conditions.

prehensive AIoT for Precision Circulating Agriculture based

• Predictive models offer guidance for optimal cultivation strategies.

on Bean Sprouts and Frass of Hermetia Illucens Larvae, sup-
ports a combination of monitoring, alerting, and automation,
and provides a suitable environment for automatic control of
stable parameters for bean sprouts. Furthermore, based on
the ARIMA framework, it uses the most efficient machine-
learning model for temperature, humidity, soil moisture,

cesses, reducing the need for manual labour.

and fertilizer forecasting. The RMSE and MAE evalua-
tion functions were used to verify that the forecast results
deviations from optimal conditions.

obtained using the ARIMA framework were better than

those obtained using the LSTM and SVR. The system can
and adaptive management.

forecast the parameters; thus, it can be deployed in a light

version, which means it can train a model on the Raspberry
Pi, a low-cost device. As a result, the proposed system suc-
cessfully controlled the cultivation precisely. The real-world
and crop types.

implementation of the proposed system provides a reality

check on its promise and potential. The lessons learned
Proposed

from addressing challenges, receiving farmer feedback, and

identifying areas for improvement underscore the dynamic
nature of agricultural technology adoption. By merging
Conventional methods may result in overuse of resources and environmental

Automation and Labor Reduction Manpower-intensive practices demand significant labour resources, contrib-

Scaling up traditional practices might lead to challenges in managing larger

Manually assessing conditions can lead to errors in resource allocation and

Decisions are often based on intuition and experience, needing more data-

theoretical innovation with practical experiences, this study

Periodic manual checks might miss rapid changes in environmental vari-

contributes to the evolution of precision agriculture towards

an intelligent and sustainable future. Integrating scalability
and cost implications into the proposed system's practicality
analysis offers a holistic perspective. Beyond the system's
technical capabilities, the feasibility of its implementation
on a larger scale rests on its economic sustainability. By
addressing scalability and cost, this study provides valuable
ables, delaying responses to critical situations.

insights for stakeholders, farmers, and decision-makers con-

driven precision of the proposed system.

sidering adopting the proposed AI-driven IoT-based cogni-

tive hardware system for agricultural precision analysis.
uting to higher operational costs.
suboptimal decision-making.

Data Availability Data will be provided upon request to the correspond-

fields and diverse crops.
Table 8  Advantages of proposed system with traditional one

ing Author.
degradation.

References
Traditional

1. Ahamd M (2019) State of the art compendium of macro and micro

energies. Adv Sci Technol Res J 13(1):88–109. https://​doi.​org/​
10.​12913/​22998​624/​103425. (Wydawnictwo Naukowe Gabriel
Borowski (WNGB))
2. Ahmad M, Ali M, Turi JA, Manan A, Al-Dala'ien RNS, Rashid K
Environmental Sustainability

Scalability and Adaptability

(2022) Potential use of recycled materials on rooftops to improve

thermal comfort in sustainable building construction projects.
Precision and Accuracy

Data-Driven Decisions
Real-time Monitoring

Front Built Environ 8. Frontiers Media SA. https://​doi.​org/​10.​

3389/​fbuil.​2022.​10144​73
3. Ali A, Fathalla A, Salah A, Bekhit M, Eldesouky E (2021) Marine
Key Features

data prediction: an evaluation of machine learning, deep learning,

and statistical predictive models. Comput Intell Neurosci 2021:1–
13. https://​doi.​org/​10.​1155/​2021/​85511​67. (Hindawi Limited)

13
Mobile Networks and Applications

4. Ahmed A, Karamatullah H, Budhwani A (2020) Analysis of robot- irrigation on daily groundwater levels. Agric Syst 159:248–259.
ics industrial manufacture & promotion of artificial intelligence. https://​doi.​org/​10.​1016/j.​agsy.​2017.​01.​017. (Elsevier BV)
J Adv Technol Eng Res 6(1). TAF Publishing. https://​doi.​org/​10.​ 18. Sunil JS, Charitha R, Kalpana P, Prasad RV (2019) Integrated
20474/​jater-6.​1.1 Framework for Farmer Centric Agriculture Applications Using
5. Ashraf MU, Asfa A, Imran M, Manzoom A (2022) Impact of IoT Architectures. 2019 IEEE International Conference on
climate change on agriculture sector in Pakistan: A case of dis- Advanced Networks and Telecommunications Systems (ANTS).
trict Lodhran, Southern Punjab-Pakistan. Pak J Life Soc Sci IEEE. https://​doi.​org/​10.​1109/​ants4​7819.​2019.​91181​17
20(1):57–62 19 Masum MH, Islam R, Hossen MA, Akhie AA (2022) Time series
6. Atique S, Noureen S, Roy V, Subburaj V, Bayne S, Macfie J (2019) prediction of rainfall and temperature trend using ARIMA model.
Forecasting of total daily solar energy generation using ARIMA: J Sci Res 14(1):215–227. https://d​ oi.o​ rg/1​ 0.3​ 329/j​ sr.v​ 14i1.5​ 4973.
A case study. 2019 IEEE 9th Annual Computing and Communica- (Bangladesh Journals Online (JOL))
tion Workshop and Conference (CCWC). https://​doi.​org/​10.​1109/​ 20. Goap A, Sharma D, Shukla AK, Rama Krishna C (2018) An
ccwc.​2019.​86664​81 IoT based smart irrigation management system using Machine
7. Banerjee A, Chakrabarty M, Bandyopadhyay G, Roy PK, Ray learning and open source technologies. Comput Electron Agric
S (2020) Forecasting environmental factors and zooplankton of 155:41–49. https://​doi.​org/​10.​1016/j.​compag.​2018.​09.​040. (Else-
Bakreswar reservoir in India using time series model. Ecol Inform vier BV)
60:101157. https://​doi.​org/​10.​1016/j.​ecoinf.​2020.​101157. (Else- 21 Subeesh A, Mehta CR (2021) Automation and digitisation of agri-
vier BV) culture using artificial intelligence and internet of things. Artif
8. Bhoi A et al (2021) IoT-IIRS: Internet of Things based intelli- Intell Agric 5:278–291. https://​doi.​org/​10.​1016/j.​aiia.​2021.​11.​
gent-irrigation recommendation system using machine learning 004. (Elsevier BV)
approach for efficient water usage. PeerJ Comput Sci 7:e578. 22 Kashyap PK, Kumar S, Jaiswal A, Prasad M, Gandomi AH (2021)
https://​doi.​org/​10.​7717/​peerj-​cs.​578. (PeerJ) Towards precision agriculture: IoT-enabled intelligent irrigation
9. Bual CLC, Cunanan RD, Bedruz RAR, Kerwin R, Billones C, systems using deep learning neural network. IEEE Sensors J
Vicerra RRP (2019) Microcontroller-based Irrigation System for 21(16):17479–17491. https://​doi.​org/​10.​1109/​jsen.​2021.​30692​
Experimental Design of Mung Bean Sprouts Cultivation. 2019 66. (Institute of Electrical and Electronics Engineers (IEEE))
IEEE 11th International Conference on Humanoid, Nanotechnol- 23. Jin X-B et al (2021) Deep-learning temporal predictor via bidirectional
ogy, Information Technology, Communication and Control, Envi- self-attentive encoder–decoder framework for IOT-based environmen-
ronment, and Management (HNICEM). IEEE. https://​doi.​org/​10.​ tal sensing in intelligent greenhouse. Agriculture 11(8):802. https://​
1109/​hnice​m48295.​2019.​90728​01 doi.​org/​10.​3390/​agric​ultur​e1108​0802. (MDPI AG)
10 Chen W-L et al (2019) AgriTalk: IoT for precision soil farming 24 Karanisa T, Achour Y, Ouammi A, Sayadi S (2022) Smart green-
of turmeric cultivation. IEEE Internet Things J 6(3):5209–5223. houses as the path towards precision agriculture in the food-
https://​doi.​org/​10.​1109/​jiot.​2019.​28991​28. (Institute of Electrical energy and water nexus: case study of Qatar. Environ Syst Decis
and Electronics Engineers (IEEE)) 42(4):521–546. https://​doi.​org/​10.​1007/​s10669-​022-​09862-2.
11. Fatima N, Siddiqui SA, Ahmad A (2021) IoT-based smart green- (Springer Science and Business Media LLC)
house with disease prediction using deep learning. Int J Adv Com- 25 Wolfert S, Ge L, Verdouw C, Bogaardt M-J (2017) Big data in
put Sci Appl 12(7). www.​ijacsa.​thesai.​org smart farming – a review. Agric Syst 153:69–80. https://​doi.​org/​
12 Fattah J, Ezzine L, Aman Z, El Moussami H, Lachhab A (2018) 10.​1016/j.​agsy.​2017.​01.​023. (Elsevier BV)
Forecasting of demand using ARIMA model. Int J Eng Bus Man- 26 Specht K et al (2013) Urban agriculture of the future: an overview
age 10:184797901880867. https://​doi.​org/​10.​1177/​18479​79018​ of sustainability aspects of food production in and on buildings.
808673. (SAGE Publications) Agric Hum Values 31(1):33–51. https://​doi.​org/​10.​1007/​s10460-​
13. Husaein FAR, Ahmad M, Rashid AK (2018) Development of 013-​9448-4. (Springer Science and Business Media LLC)
automated powered precise watering system for large framing in 27 Lopes IG, Yong JW, Lalander C (2022) Frass derived from black
desert. 7(75). http://​www.​ijsei.​com/​papers/​ijsei-​77518-​05.​pdf. soldier fly larvae treatment of biodegradable wastes. A critical
Accessed Jan 2023 review and future perspectives. Waste Manage 142:65–76. https://​
14. Rayhana R, Xiao G, Liu Z (2020) Internet of things empowered doi.​org/​10.​1016/j.​wasman.​2022.​02.​007. (Elsevier BV)
smart greenhouse farming. IEEE J Radio Freq Identif 4(3):195– 28 Wang Y, Shen Z, Jiang Y (2018) Comparison of ARIMA and
211. https://​doi.​org/​10.​1109/​jrfid.​2020.​29843​91. (Institute of GM(1,1) models for prediction of hepatitis B in China. PLOS
Electrical and Electronics Engineers (IEEE)) ONE 13(9):e0201987. https://d​ oi.o​ rg/1​ 0.1​ 371/j​ ourna​ l.p​ one.0​ 2019​
15. Ferrández-Pastor F, García-Chamizo J, Nieto-Hidalgo M, Mora- 87. (Public Library of Science (PLoS))
Pascual J, Mora-Martínez J (2016) Developing ubiquitous sensor
network platform using internet of things: application in precision Publisher's Note Springer Nature remains neutral with regard to
agriculture. Sensors 16(7):1141. https://​doi.​org/​10.​3390/​s1607​ jurisdictional claims in published maps and institutional affiliations.
1141. (MDPI AG)
16 Sami M et al (2022) A deep learning-based sensor modeling for Springer Nature or its licensor (e.g. a society or other partner) holds
smart irrigation system. Agronomy 12(1):212. https://​doi.​org/​10.​ exclusive rights to this article under a publishing agreement with the
3390/​agron​omy12​010212. (MDPI AG) author(s) or other rightsholder(s); author self-archiving of the accepted
17. Guzmán SM, Paz JO, Tagert MLM, Mercer AE, Pote JW (2018) manuscript version of this article is solely governed by the terms of
An integrated SVR and crop model to estimate the impacts of such publishing agreement and applicable law.

13