Utilizing AIoT to Achieve Sustainable Agricultural Systems in a Climate-Change-Affected Environment

Abstract

Smart agricultural systems are continually evolving to provide high-quality planting and defend against threats such as climate change, which necessitate improved adaptation and resource allocation. IoT technology offers a cost-effective approach to monitoring and managing system performance. However, this approach faces challenges, including connectivity issues and complex decision-making. While researchers have studied these problems individually, no fully automated solution has addressed them simultaneously. There is still a need for an offline solution that manages multiple processes and reduces human error. This paper introduces an AI-powered edge computing system that serves as an early-warning solution for climate impacts. This system enables autonomous management through an Agentic AI model that observes, predicts, decides, and adapts. It provides a low-cost AIoT platform for data forecasting, classification, and decision-making, converting sensor data into actionable insights. The system integrates forecast evaluation with real-time data comparisons to optimize scheduling, efficiency, sustainability, and yields. Moreover, this solution is totally autonomous and independent of internet connectivity. Demonstrating its superior performance, it reduced errors by 50% and achieved an R-squared value of 0.985.

1. Introduction

Nowadays, the world faces numerous challenges that require swift, informed adaptation to survive [1]. Water resources are depleting and altering life on Earth [2]. Cultivated land is shrinking, resulting in a sharp shortage in food production [3]. Natural nutrients and the normal ecology of the food cycle are being disrupted [4]. Climate change threatens the planting process from multiple directions, including through weather variability, global warming, soil disturbance, and pollution [5].

Various mitigation, adaptation, and remediation strategies have been implemented to sustain planting and meet the ever-growing demand for food. Water depletion has been mitigated through the sanitation and recycling of rainwater, sewage, and seawater [6]. In addition to smart water management within agricultural fields, this helps minimize overall water use. Notably, recent reports indicate that up to 70% of water is used for irrigation [7]. Widely accepted efforts to minimize water consumption in irrigation are encouraged to ensure a more abundant water supply for humanity and a better life on Earth. Land cultivation can be improved through wise fertilization, incorporating nanotechnology to maximize soil nutrition benefits and reduce losses [8]. Furthermore, a recent agricultural approach known as regenerative agriculture aims to mitigate the impacts of climate change by enhancing and expanding cultivated land [9]. One key aspect of soil regeneration solutions is the preservation of soil organic carbon [10]. Organic carbon in the soil is maintained by the diffusion of increasing amounts of CO₂ from the atmosphere into the soil. This build-up of organic matter creates a viable surface layer, enabling more effective soil preparation for planting and cultivation. Climate change is addressed from another perspective through shelters and greenhouses [11]. In this context, conditioning the air and creating a suitable environment for plants are essential to sustain the planting process and achieve higher production rates. However, this solution is limited to enclosed agricultural environments, such as greenhouses.

All these solutions require continuous monitoring and the timely application of appropriate amounts of fertilizers, medicines, and chemicals. As agriculture has evolved, animals, humans, and machines have been adopted to maintain the principles of monitoring and acting [12]. Over the past decade, a new principle has emerged through the well-known Industry 4.0 technology. This technology advances precision agriculture by enabling smart agricultural systems. A smart agriculture system is a vital framework that allows a productive response to the urgent global threat to food security [13]. Its goal is to maximize crop yields while minimizing energy and water consumption. The system continually advances by integrating emerging technologies, including the Internet of Things (IoT), machine learning, and artificial intelligence (AI). Recently, the integration of IoT and AI has given rise to the Artificial Intelligence of Things (AIoT). This integration automates various farming processes, notably irrigation, fertilization, seeding, and harvesting [14]. The system oversees these processes and promotes sustainable operations by monitoring and controlling them. The current system model lacks the decision-making ability to autonomously manage all these processes concurrently.

There is also a connectivity issue when accessing a cloud dashboard or system platform from one side and connecting from within an agricultural field from the other. Providing heterogeneous wireless communication across all fields is a key solution for maintaining sustainable operations. In addition, the diversity and interoperability of sensors require a heterogeneous wireless infrastructure. Maintaining a heterogeneous wireless infrastructure provides a resilient network for sensor data collection and enables continuous data acquisition in smart systems. Achieving resilient connectivity among system components with minimal delays and outages is also essential. However, other key factors, such as system scalability and process management, should also be considered [15]. Process management also requires continuous data collection, which involves maintaining data flow between IoT sensing elements and controllers [16]. Distributing IoT data across multiple heterogeneous wireless technologies may cause delays to vary due to fluctuating data processing times [17], potentially disrupting operations or leading to failures in agricultural management processes. In addition to the diversity of communication technologies, the primary cause of delay variation lies within the data collection, analysis, and management processes. Furthermore, monitoring the environmental impacts on all physical aspects from potential changes due to climate change requires an early-warning system [18]. Soil degradation, which decreases soil organic carbon levels, is among the most critical factors [19]. Those critical factors and hazardous conditions, such as fire, flooding, insects, and the spread of infection, require a faster response.

Old and current smart models that rely on cloud or remote dashboards are not beneficial due to their delayed response times. It is beneficial to strengthen the embedded board’s abilities for better organization and response, rather than merely collecting and transmitting data to a cloud dashboard. This approach not only optimizes the edge solution but also eliminates processing and propagation delays, as data processing is maintained locally within the field perimeter [20]. The recently deployed smart system controller employs edge computing technologies on embedded system boards [21]. Diverse solutions are found to address smart irrigation, fertilization, and plant remediation individually. The decision tree (DT) is found to be one of the most appropriate solutions for irrigation and fertilization. However, there remains a need to determine quantities rather than simply deciding upon them. There remains a need for integrated smart systems that provide a hybrid solution for the simultaneous management of smart irrigation, fertilization, and plant remediation. In fact, this could be achieved through a cloud-based solution. However, several challenges remain, including operating costs, latency, complexity, and limited connectivity. Limited connectivity is the main constraint, especially in remote and underserved areas. For this reason, reliance on cloud solutions is problematic; even under the best connectivity conditions, it results in delays that severely undermine efforts to address climate change and take urgent action.

In this paper, we propose an agentic AI model that integrates smart irrigation, fertilization, and plant medicine management. Managing heterogeneous sensor data and incorporating predictive models for sensed data are key assets of the proposed solution. The forecasts from agricultural system sensors are transformed into inputs for a hybrid model.

This hybrid model uses lightweight techniques to reduce processing delays and simpler hosting hardware to reduce costs. This hybrid management solution is proposed through an intelligent edge framework that maintains sustainability objectives as an agentic workflow. This agentic workflow employs a DT with the Fuzzy Analytic Hierarchy Process (FAHP). This improves classification to autonomously manage irrigation, fertilization, and plant treatment through adaptive, explainable processes. Moreover, it provides early warnings on climate change, autonomous system management, and urgent action. The paper is organized into four main sections: related work, methodology, simulation and results, and conclusions. The main contributions of this paper are as follows:

• Adopting a new general decision-making model by incorporating heterogeneous sensor data. This processing of heterogeneous data effectively supports more accurate decisions.
• Proposing an intelligent edge solution that adopts the Hybrid-DT algorithm for managing remote farm fields that lack internet connectivity.
• Proposing a more intelligent agricultural solution that predicts farming inputs and provides real-time concurrent processing of irrigation, fertilization, and remediation by employing a modified multiple linear regression (MLR). The proposed predictive solution aims to mitigate the impact of climate change on soil by accounting for both linear and non-linear responses to environmental change.

2. Related Work

Smart agriculture systems have some significant connectivity challenges, including communication and processing delays [22]. Communication delays are due to the propagation delay component in reaching the system platform or dashboard. Several factors contribute to this delay, including link failures, congestion, and cloud or server overhead. On the other hand, processing delays are the time required to process data and take appropriate action. This delay component may be related to system capacity, server or platform capabilities, and complexity. These delays significantly affect the performance of smart agricultural systems, particularly in data analysis and the timely execution of necessary actions [23]. Traditional agricultural model solutions use a human-dependent automation approach, so they are not affected by such delays [24]. This conventional agrarian system is limited in its decision accuracy, resulting in delays in decision-making and action. Irrigation and fertilization are typically important for resource conservation. Furthermore, other agrarian processes are time-sensitive and require a prompt response [25]. These issues are often related to environmental factors, including climate change and natural disasters. Recently, modern smart agricultural models have increasingly focused on the autonomous execution of farming processes. However, the aforementioned connectivity and processing delays create constraints. Extensive research has focused on addressing these delays, and various solutions have emerged in the recent literature. Discussions of the use of artificial intelligence (AI) have dominated research in recent years, particularly in prediction and action. For instance, AI has been harnessed to better manage cultivated areas by predicting soil moisture levels and subsequently providing optimal irrigation practices, as highlighted in [26]. Furthermore, AI is used to predict crop-yield timing, enabling timely harvesting and practical fruit preservation, as noted in [27]. The authors of [28] discussed how predicting the impacts of climate change on soil health could potentially reduce reliance on certain fertilizers. Another study, as reported in [29], explores how predicting pollution rates in flood irrigation systems can help maintain yield quality and plant health. Additionally, research reported in reference [30] focused on forecasting potential plant infections to prevent disease outbreaks and subsequent crop failures. Reference [31] examined the broader implications of climate change for agriculture, emphasizing irrigation and fertilization strategies to mitigate adverse effects. Another innovative direction, initiated by [32], confirms that forecasting sensor readings is vital for mitigating the connectivity failures that are commonly observed in wireless communications. These predictions are essential for monitoring environmental parameters along with more precise decision-making. A recent study [33] further investigated this by developing an early-warning system that predicts threats posed by climate change.

On the other side, several authors have posited that AI plays an essential role in preserving plant health and in yield prediction, as reported in [34]. Others provide early symptoms of malnutrition, as in [35]. Another study suggested an effective strategy to identify an immediate need for irrigation [36]. On the other hand, ref. [37] was more interested in studying the predicted data by analyzing the insect footprint. These collective efforts have established a framework for anticipating sensor data and recommending timely management actions. However, further research is still required to enhance the reliability of smart agricultural systems.

From another perspective, automation in the field has undergone several developments. An automation system with a smart scheduling system based on an expert judge was adopted in [38], which was considered semi-autonomous. This semi-automatic solution resulted in substantial losses of resources and in planting quality due to inaccurate decisions. Another wave of adoption of fog computing was reported in [39], which relied on internet connectivity and system management. This dependence on the internet and system management led to resource losses and a degradation in quality. The cloud computing solution described in [40] offers a superior approach to autonomous management via a cloud dashboard. However, it remains internet-dependent, and the difficulty of modifying the cloud system has led some to consider it a waste of resources. The promising era in edge computing warrants greater investment in developing autonomous systems [41]. The main limitation is their resource-constrained capacity to host AI models, whether in fog or cloud environments. Several developments have been made to provide a cloud–edge solution, as in [42]. However, this system remains cloud-dependent and is susceptible to outages or mis-synchronization. A better development is investing in lightweight techniques that run totally on the edge [43]. Developing an offline intelligent edge solution remains challenging, particularly when hosting an agentic AI model. An agentic model provides full autonomous capability to observe, forecast, process, evaluate, and adapt [44].

However, implementing a trustworthy edge solution in the agricultural sector would enable more efficient resource use, improved planting practices, and early warnings of crises and issues. The most affected agricultural component by climate change is soil. This soil impact manifests in various forms; however, the most critical is the soil organic carbon (SOC) component [45]. The SOC is related to the soil’s quality for planting; its degradation eliminates the possibility of planting. This paper presents an early-warning system that forecasts impacts on SOC and supports informed decision-making to mitigate their effects. The proposed solution should maintain the autonomous forecasting of heterogeneous field data on weather, soil, and plants, followed by a decision-making process. The decision-making process aims to enable the autonomous management of multiple processes: irrigation, fertilization, and plant remediation.

However, prioritizing one process over another is a critical task and should be managed carefully. Determining which process to prioritize is a key challenge in implementing an autonomous system. Moreover, the timing of merging or executing multiple processes is a significant factor in achieving autonomous operation. This management requires careful processing of heterogeneous data followed by an orchestrated decision system that answers the following questions:

• Which method should be used to decide on priorities?
• What is the heterogeneous data to be analyzed?
• Which process should be started first?
• When should multiple processes be merged or executed?
• How much of a quantitative amount of a resource, or time, should be allocated by the process?
• How are the data and decision system evaluated?

3. Model Architecture

The proposed model for the agricultural system is shown in Figure 1 and includes a control unit, wireless gateways, and sensing stations. The sensing stations feature a range of sensors for weather, soil, and plant health. The controller’s function is to gather and analyze sensor data while also monitoring and managing relevant agricultural processes, such as irrigation, fertilization, and remediation. The model assumes a heterogeneous wireless communication infrastructure that minimizes delay and ensures efficient data communication via the multi-access control unit. Jitter is another challenge that impacts the effective processing of sensor data. The proposed model assumes an orchard area on a square plot with a side length of 300 m. The orchard field is systematically planted with fruit trees, with each tree spaced 5 m apart, and is irrigated via drop irrigation tubes in 60 rows and 60 columns. The sensor nodes (SN) are distributed throughout the field to collect the physical parameters of the climate, soil, and plants. The sensor node controller is a centralized node that is highly capable of collecting, analyzing, and deciding. The controller node is assumed to be an intelligent multi-access edge (iMEC) unit that can connect to multiple access links to ensure reliable connectivity to sensor nodes. The controller node is also assumed to be an intelligent edge and hosts the proposed agentic AI model. The system is considered and adapted after evaluating the forecasts of the data inputs and the system’s decision, as assessed by agricultural experts. The proposed model addresses this challenge by accurately predicting and classifying sensor data to support relevant process analysis. Additionally, decision-making is employed to enable the autonomous management of agricultural processes.

Data collection will be performed via the multi-access edge controller, which connects to available wireless access points. The proposed multi-access solution enables the intelligent edge controller to collect data from myriad sensors in a sustainable manner. Additionally, a data prediction process is proposed to mitigate potential data drops, delays, and outages by aiding in proactive processing. The proposed solution provides an ongoing check to synthesize the predicted data with real-time data collection. This synthesis process assesses the predicted data to ensure that they are as close as possible to the real data’s behavior. This capability is designed to provide proactive, real-time data processing. Moreover, it provides intelligent decision-making in managing multiple agricultural processes. This deliberate decision-making aims to enable the simultaneous operation of smart irrigation, smart fertilization, and innovative remediation.

4. Methodology

A flowchart of the steps of the proposed model’s process is shown in Figure 2. The model relies on five main processes: data collection, prediction, decision-making, evaluation, and adaptation. The model enables an innovative agricultural system that integrates seamlessly with heterogeneous wireless infrastructure.

This prerequisite step is useful for establishing a resilient and sustainable connectivity framework among the system elements. Next, the study implements the system to collect extensive sensor data from the agricultural environment. The study adopts an agricultural system comprising an embedded system board, agricultural sensors, actuators, and pumps with tanks. The system collects sensor data, analyzes them, and makes a relevant decision by actuating the relevant pump.

Agricultural sensors collect data on weather, soil, and plant health. The weather sensors measure temperature and humidity. The soil sensors collect the soil temperature and humidity. The plant health sensors collect UV index data and perception data. Agricultural data are essential for monitoring plants; however, they may be lost due to sensor disconnection or connectivity failures.

In the next step, the average arrival time of sensor data is calculated to identify any transmission delays. This analysis is vital for ensuring timely responses to agricultural needs. This data acquisition is maintained over a full planting season: three months. Moreover, the system predicts data bursts to support the evaluation of agricultural processes for abnormal behavior. The predicted data are compared with the empirical data, and the predictions are adapted as needed. These data bursts are then processed when the system detects abnormal sensor readings. Moreover, the system evaluates and adapts agricultural processes in the context of climate change, focusing on the SOC feature. Data prediction enables timely responses to processing and connectivity delays.

The model uses MLR to predict sensor data parameters, enabling more accurate forecasts of future conditions. The system then compares the predicted data with the actual sensor burst readings. These comparisons are performed one by one since the data are time-stamped. The comparison factor is computed as the average over the burst, yielding a substituted value. The value of the comparison factor is maintained in serial per time stamp and used to adapt the burst data.

Once the predictions are finalized, the data are systematically sorted and classified according to the specific agricultural tasks involved. This classification is performed using a decision-making algorithm that assigns priorities to actions, such as irrigation, fertilization, and remediation, based on the collected data. The system assesses this information, with particular attention to its relevance to ongoing farming activities. Based on insights derived from integrating predicted data with current operations, the system executes the necessary actions to improve agricultural outcomes. Finally, the effectiveness of the proposed forecasting method is thoroughly evaluated and compared with existing methods described in the literature to ensure it meets or surpasses current standards in agricultural data management and decision-making.

The system operates intelligently, prioritizing processes according to urgency and relevance. The normal agricultural process is applied every 15 min to the original dataset. The burst dataset is used to detect abnormal behavior, with data simplified every 10 min and heavy data every 5.4 min. This strategy maintains the agricultural process, responds only when necessary, extends device lifetime, and reduces power consumption. The study employs a DT method to support this decision-making process, ensuring that actions are taken in the most effective order. Ultimately, the proposed solution is evaluated and thoroughly validated through comparative analysis with similar methodologies in the existing literature.

5. Procedure

5.1. Sensor Data Prediction and Classification

The proposed agricultural system aims to continuously monitor and maintain optimal planting conditions from seeding through ripening and harvesting. The system collects heterogeneous sensor data to produce more accurate decisions. The study tests and validates the adopted sensors and provides the required calibration for the sampling rate. The system evaluated a strawberry plant over a 90-day cultivation period. The included dataset measures long-term environmental trends and burst data. Long-term data are collected at 15 min intervals and used in decision-making. These long-term data are collected over three months and averaged to obtain an average daily reading.

The burst high-frequency (20 Hz) data are simulated, and the predicted data are used for critical anomaly detection (e.g., fire, flooding, malnutrition, infection, insects). This burst data is split chronologically (60/40) to validate the model’s predictive power across different growth stages. The collected data include timestamps normalized to a maximum delay of 50 milliseconds [46]. The maximum delay value ensures automatic sampling, aligns with diverse sampling rates across sensors, and keeps monitoring for sudden changes. However, to mitigate the risk of data loss or the loss of relevant sensor information, data prediction is necessary. The sensor data prediction method also includes slope verification of the real and predicted data to improve prediction accuracy. Data prediction is followed by the proactive making of accurate decisions and the triggering of actions when crisis symptoms are detected. The proposed decision-making process is crucial for determining which agricultural processes should be prioritized for maintenance first. The decision-making process aims to execute multiple processes simultaneously. Several agronomic processes are considered; however, the proposed system addresses only irrigation, fertilization, and remediation. The classification method employs a normalization technique to efficiently process heterogeneous data. The sensor data are then classified and labeled according to the relevant process, as shown in Figure 3.

The system preserves a full monitoring dataset (3 months) and a burst dataset (the 150-sample high-frequency packets). The full monitoring dataset is used throughout the season to support autonomous agriculture. In contrast, the burst dataset is used specifically for transient event detection and to validate the Esp32’s ability to process environmental shifts such as fires or flooding. The system labels the datasets according to the nature of the agriculture process, either as a “normal” or “critical” state:

• Normal Labels: Samples were labeled Irrigation Required if soil moisture dropped below 60% and Fertilization Required if Electrical Conductivity (EC) fell below 1.5 dS/m.
• Critical Event Labels: These were synthetically or physically induced for demonstration.
  • Fire: Labeled by a rate-of-rise threshold.
  • Infection: Labeled via soil records through an inspection of fungal growth, correlated with high humidity for normal periods.

The probability and possibility of running the process are automatically maintained throughout the automated decision-making process. This study replicates a human judge, as incorporated in prior systems, to determine which of the three processes to run first. A human-dependent system, according to Fisher [47], has a probability of making a correct decision from x alternatives given by the following:

P(x) = 1/n

(1)

Since the system has three alternatives (irrigation, fertilization, and plant medicine), the probability of making a correct decision is 33.33%. In contrast, the likelihood of making a wrong decision is 66.67%. Furthermore, the decision-making process for selecting multiple alternatives concurrently becomes more sophisticated. The probability of making a correct decision drops to approximately 11.11%, or (1/9). This probability is derived from the possibility that human-dependent systems can make correct decisions. However, the autonomous system is better able to make decisions because it can process heterogeneous data. An autonomous system faces two limitations in achieving this: the operational error rate and the risk of interruption.

Offline edge computing ensures accuracy and avoids interruptions caused by outages in internet and cloud connectivity. A set of methods is adopted to provide a confined solution within an intelligent edge solution. The methodology adopts lightweight techniques, including machine learning, MLR, and DT algorithms.

5.2. Hybrid Regression Algorithm for Environmental Parameter Prediction

The DT categorizes all sensor data within the proposed agricultural system. The sensors used in this system, depicted in Figure 3, comprise weather- and soil-sensing stations. The weather sensors measure temperature, humidity, rainfall intensity, and light intensity. The soil sensors measure temperature, humidity, moisture, and pH levels. The irrigation process considers soil, air, and moisture temperatures. While executing the irrigation process, it also records timestamps and durations and determines the appropriate amount of irrigation. Fertilization also considers climate, humidity, and soil moisture, while tracking timestamps, type, and amount during execution. The remediation process also logs onto timetables and treatment parameter types.

Predicting relevant physical parameters is essential for maintaining the sustainability of the proposed agricultural system as a farming approach. Forecasting is crucial for continuous monitoring and processing, particularly during periods of delayed connectivity, connectivity failures, or when a sensing unit is unavailable. The methodology employs a hybrid regression framework to predict the most significant climate change phenomena, specifically SOC. The proposed hybrid regression models incorporate the climate-driven properties of temperature, humidity, and the UV index linearly to quantify the effects of external environmental variables on SOC. The proposed hybrid algorithm captures non-linear internal soil behavior (e.g., microbial activity and organic matter decomposition) via a regression coefficient that evolves into a time-varying polynomial function, thereby representing the non-linear soil response to global warming and long-term climate forcing. The hybrid regression model of the SOC may be expressed mathematically as follows:

$$\mathrm{SOC}(\mathrm{t}) = \mathrm{\alpha }(\mathrm{t}) + {\mathsf{\beta }}^{\mathrm{T}} X^T+\epsilon (\mathrm{t})$$

(2)

$$X^T=\left[{\mathrm{x}}_1,{\mathrm{x}}_2,\dots ,{\mathrm{x}}_{\mathrm{n}}\right]$$

(3)

$${\mathit{\beta }}^T=\left[{\mathrm{\beta }}_1,\mathrm{\beta },\dots ,{\mathrm{\beta }}_{\mathrm{n}}\right]$$

(4)

where

• $\mathrm{S}\mathrm{O}\mathrm{C}\left(\mathrm{t}\right)$: soil organic carbon at time $\mathrm{t}$
• ${\mathit{\beta }}^T$: linear regression coefficient transposed vector (climate-driven sensitivities)
• $\mathrm{\epsilon }\left(\mathrm{t}\right)$: model residual (white noise)
• X^T: predictor transposed vector (such as air temperature, air humidity, soil temperature, soil humidity, UV index)
• $\alpha \left(t\right)$: a non-linear adaptive regression constant which represents internal soil non-linear feedback to warming [48]. The non-linear adaptive regression constant $\alpha \left(t\right)$ can be expressed as follows:

$$\alpha (t)=c_0+c_{1}t+c_2{t}^2+c_3{e}^{\gamma T_{soil}(t)}$$

(5)

where

• $c_0,c_1,c_2,c_3$: polynomial coefficients (learned adaptively).
• $t$: normalized time index (0 for base year).
• $T_{soil}\left(t\right)$: soil temperature (captures direct thermal nonlinearity).
• $e^{\gamma T_{soil}\left(t\right)}$: exponential component reflecting accelerated decomposition and carbon loss due to warming.

The proposed hybrid regression algorithm models the effects of climate variables linearly (air/soil temperature, humidity, UV). It incorporates non-linear soil feedback from global warming into the adaptive regression constant to predict SOC behavior for the period of 2025–2050. The regression method predicts how soil responds to a physical climate parameter, helping maintain system stability even during wireless communication failures or delays. Additionally, it continuously adjusts the soil to the effects of climate change without processing delays.

5.3. The Proposed Hybrid Decision-Making Algorithm

The DT is a machine learning method commonly used for classification [49]. This study requires a decision-making method that uses forecasted data to support informed decisions. These decisions will determine which processes to execute first. The DT takes forecasted data from MLR as input and outputs the process to be performed. In this context, three methods are investigated: irrigation, fertilization, and plant remediation. Classification trees using the Gini index are used to verify the branching of decisions at the root. The root evaluates the measured data and considers branching into two main decision paths: irrigation or fertilization. It then applies the Gini coefficient to assess the decision and determine a new course of action, such as continuing irrigation or switching to fertilization. Branching and splitting are based on the Gini index, which evaluates decisions using the dataset and the decision itself. The Gini score ranges from 0 to 1 and provides an assessment based on specific criteria. A score closer to zero indicates an accurate decision, while a score nearer to one indicates an incorrect decision. The decision that receives the highest Gini score advances to the next split, while the other decision is terminated. This process is crucial for achieving the objectives of the proposed system by autonomously managing multiple agricultural operations simultaneously and triggering additional actions in response to new decisions. The pseudocode for the proposed agrarian system is presented in Algorithm 1.

Algorithm 1: SOC_Prediction_And_Agri_Decision_Support

Input: * Historical dataset D = {(Pi, ATi, AHi, LIi, STi, SOCi) | i = 1…n} //where: Pi = Precipitation, ATi = Air Temperature, AHi = Air Humidity, LIi = Light Intensity, STi = Soil Temperature, SOCi = Soil Organic Carbon at time ti. Future sensor readings F = {(Pt, ATt, AHt, LIt, STt) | t = 2025…2050} Output: SOC_forecasts = {SOCt | t = 2025…2050} Agri_decisions = {Decisiont | t = 2025…2050} 1.  Initialize: Define input features and target. X = [P, AT, AH, LI, ST] and 2.  target Y = [SOC]. 3 . Step 1: Train Multiple Linear Regression (MLR) Model 4.  Fit on dataset using least squares: 5 . Y = β0 + β1·P + β2·AT + β3·AH + β4·LI + β5·ST + ε 6.  Step 2: Forecast SOC for each year t from 2025 to 2050 7.       foreach year t in 2025 to 2050 do 8.           Extract sensor values from xt = (Pt, ATt, AHt, LIt, STt) from F 9.           Predict SOCt = MLR_Model.predict(xt) 10.         Add SOCt to SOC_forecasts 11. end 12. Step 3: Train Decision Tree Classifier 13. Define labeled training dataset A = {(Pi, ATi, AHi, LIi, STi, SOCi, Labeli) } 14. Train DecisionTree_Model on A using information gain or Gini index 15. Step 4: Predict Agricultural Decision 16.    foreach year t in 2025 to 2050 do 17.        Construct input vector zt = (Pt, ATt, AHt, LIt, STt, SOCt) 18.        Decisiont = DecisionTree_Model.predict(zt) 19.        Add Decisiont to Agri_decisions 20. end 21. return SOC_forecasts, Agri_decisions

The algorithm, as shown in the pseudocode provided through Algorithm 1, begins by collecting the perception dataset comprising air temperature, humidity, light intensity, and soil temperature. Later, it creates a new reading register to store environmental parameters predicted by MLR. This is performed to generate two categories in a frequently used dataset. In the first category (A), the algorithm runs on 150 datasets sampled at 50 ms and processes them every 10 min. In the second, the algorithm processes 6500 datasets sampled at 50 ms intervals, taking 5.4 min. However, the real dataset collected from the empirical design is processed for routine agricultural decision-making every 15 min. This agentic AI approach provides prediction and validation, as well as processing and adaptation within the agricultural process.

Next, the algorithm performs MLR to forecast weather variables for 2025–2050. It then stores the newly predicted values in each relevant register. The MLR also models the relationship between soil carbon and the predicted environmental parameters to forecast SOC values over the prediction period. The process involves four main steps:

• Step 1: Train an MLR model to assign each environmental parameter to a register, with SOC as the target output. Use the least-squares method to fit the relationship described in Equation (2) more accurately.
• Step 2: Forecast SOC for each year from 2025 to 2050. The algorithm runs a recursive loop to calculate SOC values for each set of environmental variables and stores these values.
• Step 3: Train a DT Classifier. In this step, the algorithm generates a new dataset that includes environmental variables and the predicted SOC value and assigns labels to the DT. These labels serve as indicators for decisions regarding irrigation, fertilization, medication, or no action. The decision model then trains on this dataset.
• Step 4: Predict agricultural decisions. The algorithm determines each decision for a data vector through a recursive loop spanning the years 2025 to 2050. It uses the Gini index to split decisions within a branch, a technique known as leaf-based processing. The option with the higher Gini value continues to divide, whereas the other stops. Finally, the algorithm lists and executes the decisions in the systematic order obtained.

This model aims to maintain an efficient DT that not only performs classification but also provides leaf-level management. This could be achieved by implementing a hybrid algorithm that combines DT with FAHP. DT is used to make individual decisions for each process (yes/no), and FAHP is employed to determine dosage/priority from several uncertain inputs (e.g., soil moisture slightly low and a high crop stage). In such a case, the decision is to perform moderate irrigation. FAHP is combined with DT for prioritization and dosing. It is worth noting that FAHP provides additional judgment at the decision-tree leaf level and contributes to the delivery of multiple process-management solutions, such as smart irrigation and fertilization. The leaf index value is also used as a control for the calibrated application rates of water, fertilizer, and medicine. The pseudocode of the hybrid decision-making algorithm is provided in Algorithm 2.

Algorithm 2: Decision Logic and Safety Execution

1.  // Decision tree prediction 2.  Set prob_irrigation=leaf_output(rrigation′) 3.  if prob_irrigation > 0.7 then 4.           action=‘irrigation’ 5.           actions. append(‘lopper’) 6.    end 7.    // Prioritize or schedule multiple actions 8.    actions=prioritize(actions). sensor_data 9.    foreach action in actions do 10.         if actions_prioritize(actions.sensor_data) then 11.             Execute a.action() 12.         end 13.     end 14.     // Execute actions safely 15.     foreach action in actions do 16.            if safety.check(action.sensor_data) then 17.                for safety.check(action.sections_holder_for_portions) do 18.                      skipped=compass=Win or skipped 19.                      i=uninterrupted_ever 20.                 end 21.            end 22.      end

A supervised DT classifier is trained to map the feature vector given by the algorithm represented by the following:

$$X^T=\left[x_1,x_2,\dots ,x_6 \right]$$

(6)

where

• $x_1$: Soil humidity (SH)
• $x_2$: Soil temperature (ST)
• $x_3$: Soil salinity (EC)
• $x_4$: Air temperature (AT)
• $x_5$: Air humidity (AH)
• $x_6$: Ultra-violet index (UVI)

The output vector $Y^T$ is given by the relationship formed as follows:

$$Y^T=[y\_1,y\_2,\dots ,y_7 ]$$

(7)

where

• y₁: irrigation decision
• y₂: fertilization decision
• y₃: plant medicine decision
• y₄: multiple decisions (y₁, y₂) == (y₂, y₁)
• y₅: multiple decision (y₁, y₃) == (y₃, y₁)
• y₆: multiple decision (y₂, y₃) == (y₃, y₂)
• y₇: no decision or alteration of the current process

The Gini index is used as the impurity criterion to select optimal splitting attributes. During training, the tree partitions the data space into decision regions corresponding to distinct agricultural conditions. At inference time, each new sensor input is propagated through the tree to reach a terminal leaf node Li, which gives the output $y_k$ with probability $P\left(y_k∣L_i\right)$ for all possible actions. For single classification, the final decision $y_{dec}$ is given by

$$y_{dec}=arg \underset{k}{max} P(y_k\mid L_i)$$

(8)

For multi-class (multi-label) conditions, multiple processes are triggered within Formula (9).

$$Y=\{y_k:P(y_k\mid L_i)>\tau \}$$

(9)

where $\tau$ is the confidence threshold (set to 0.76). The DT leaf is indexed using the Gini index [50], estimated using Formula (10).

$$\mathit{Gini}(L_k)=1-{\sum }_{i=1}^k{P(y_k\mid L_i)}^2$$

(10)

The closer the Gini index is to zero, the more it indicates that a final decision is warranted and that no further splitting is required. However, a higher fractional value near 1 indicates that greater splitting is needed, and this fraction will be used to maintain proportional action. The decision tree provides a possible decision, which is insufficient for execution and management, especially in cases involving multiple simultaneous actions. The proposed solution prioritizes actions using the FAHP. To address the uncertainty in agricultural methods, a secondary reasoning layer is applied. The FAHP module assigns priority scores ${\mathsf{\lambda }}_{\mathrm{i}}$ for deciding each action appropriately, giving an action priority as given by

$$\mathit{Action} \mathit{Priority}=\mathit{sort}\left(P\right(y_k\mid L_i)\times {\mathsf{\lambda }}_i)$$

(11)

Once the optimal set of processes and priorities is determined, the Execution Manager converts these decisions into actuator commands. This relevant layer includes the following:

• Safety Validation: Ensures no conflicting or unsafe actions occur (such as fertilization immediately after irrigation).
• Parameter Translation: Converts probabilities into control variables (such as irrigation time = base time × $P\left(\mathrm{Irrigation}\right)$).
• Scheduling: Executes single or multiple operations in sequence according to computed priorities.
• Feedback Monitoring: Continuously reads post-action sensor data to evaluate the effect of each operation. The simplified execution logic is illustrated in Algorithm 3.

Algorithm 3: Action Execution with Safety and Logging

foreach action in actions do     if safety_check(action,sensors) then         execute(action(intensity) = prob × priority[action])         monitor_and_log(action)     end end

Sensor feedback is stored following each execution for system evaluation. If post-action conditions deviate from expected outcomes, new labeled instances are added to the training dataset. The DT is retrained periodically to reflect changes in field behavior, enabling continuous improvement in decision accuracy. Over time, this closed-loop adaptive mechanism allows the system to self-optimize in response to environmental changes, crop types, and seasonal variations.

6. Results and Analysis

6.1. Measurements and Simulation

The simulation process involved forecasting sensor data using MLR, as described in Equation (2). The regression was performed in MATLAB 2024 on a Dell Latitude with a 2.45 GHz Core i7 processor, 16 GB of RAM, and a 250 GB SSD (Dell systems, Round Rock, TX, USA). Meanwhile, Python 3.8, along with the Scikit-learn (Version 1.8.0) and Pandas packages (Version 1.17.0), was used to develop code for classifying sensor data using a DT classifier. The study implemented a practical setup for the smart agriculture system, as shown in Figure 4. This system used an ESP32 controller with connected sensors and pumps (Espressif Systems, Shanghai, China). The weather sensor provided measurements of humidity, temperature, and precipitation. The light sensor assessed UV radiation. The soil sensor evaluated soil temperature and moisture. The implementation was carried out within an outdoor strawberry pot with clay soil in a semi-arid environment. The system recorded up to 300 measurements using the Chennai GST-3AFMPA8962J2Z5 Soil sensor (Concord Scientific Devices, Chennai, India), the DHT11 temperature and humidity sensor (Adafruit, New York, NY, USA), and the DC 3.3 V–5 V UV detection sensor (Adafruit, New York, NY, USA) [51]. The empirical measurements were collected as listed in

Table 1. The proposed smart agriculture systems sense physical parameters.

Entries	Perc. (%)	Air Temp (C°)	Relative Air Humidity (%)	Light Intensity (Luminous)	Soil Temp (C°)
1	55	14	68	790	23
2	39	15	67	750	23
3	18	20	67	800	23
4	13	26	67	850	23
5	7	30	67	890	23

using the following procedure:

• Install sensors for soil moisture, soil temperature (°C), air temperature (°C), and relative humidity (%).
• Acquire 300 time-series readings over the monitoring period. Each sample has an entry period < 50 ms, and the entries are combined into a CSV file containing up to 300 entries.
• Record timestamps, sensor readings, and ancillary metadata (e.g., irrigation event, fertilizer application). This is efficient for taking proper action over the relevant period.

Figure 5 compares the measured values with the weather.com [52] data (accessed on 15 July 2024). The difference between the exact and sensed data is within 0.02%, indicating reliable measurement.

A machine-learning technique is employed to maintain real-time processing. MLR is a machine learning technique that produces predictions with an error rate of 0.02 or less, matching the sensing device’s measurement error. The comparison between predicted environmental readings and actual measurements is shown in Figure 6.

The expected values closely match the absolute sensed data, as confirmed by an empirical study. This strong alignment could be an asset for enhancing real-time decision-making. The next step is to predict the effects of these values on SOC for 2025–2050 and to develop an appropriate decision-making strategy to address this issue continuously. Utilizing MLR techniques helps assess the impact of climate change on soil behavior. This study investigates the burst datasets frequently for clay soil to derive actionable agronomic insight as a quick sense and respond strategy for the following:

• A total of 150 heterogeneous sensor readings (soil moisture, soil temperature, air temperature, and humidity). The test sample consisted of 60 measurements obtained via a random split. Since the dataset is relatively uniform, with a sampling rate of 50 milliseconds, it was checked frequently, every 10 min.
• A total of 6500 heterogeneous sensor measurements (soil moisture, soil temperature, air temperature, and humidity). The test sample consisted of 2400 measurements obtained via a random split. Since the dataset is relatively uniform, with a sampling rate of 50 milliseconds and checked frequently every 5.4 min, it was checked frequently, every 5.4 min.

The system analyzes the predicted burst dataset to evaluate system performance under the SOC feature-reduction threat.

In Figure 7, we consider the SOC rate as an indicator of climate change; it is expected to reach its lowest point by 2050 [53]. This Figure shows that the regression model developed by the proposed system closely aligns with the forecast results. This forecast supports the development of a climate change adaptation plan through soil remediation. It also advocates for maintaining smart agriculture as a proactive system that continues soil remediation even when sensor readings are missing. To achieve this, the system uses the predicted burst dataset rather than the real-time dataset to preserve the agentic AI rule.

The agricultural process decisions are illustrated in Figure 8 as a DT-based classification that guides the proposed agricultural system’s decisions based on sensor readings.

The agricultural processes overseen by the proposed agriculture system are interconnected, making their integration essential for the planting system. Figure 8 illustrates the DT classification along with the relevant parameter values and classes. The root status is chosen based on humidity to guide a fertilization decision, with a Gini index of 0.653 suggesting it might not be optimal and could benefit from improved decision-tree splitting. The split results in two statuses: one in orange on the left-hand side (LHS), representing an irrigation decision with a Gini index of 0, indicating a perfect decision with no further splitting needed. Meanwhile, the right-hand side (RHS), in green, indicates an incomplete decision, with a Gini index of 0.469, suggesting additional fertilization decisions are necessary. The fertilization branch then splits further into an LHS decision labeled “plant remediation” (Gini index = 0) and an RHS decision for fertilization (Gini index = 0.408), indicating that further splitting is required. This additional split yields an RHS decision for plant remediation with a Gini index of 0, whereas the LHS yields a fertilization decision with a Gini index of 0.278. The LHS then branches into two final decisions that should be processed simultaneously: fertilization and plant remediation. The DT, therefore, follows a logical sequence of actions to maintain optimal planting conditions and SOC levels. The DT-FAHP not only supports classification but also provides action-management rules. This could be expressed as follows:

$$\mathit{Action}\_\mathit{Set}=f\left(P\right(y_k\mid L_i), \mathit{safety}, \mathit{priority})$$

(12)

The function f(.) maps the leaf’s output probabilities and external rules to executable decisions. Each leaf node represents a final decision outcome (the classification). The leaf decision contains the following:

• The predicted class label(s) such as “irrigation”, “fertilization”, “plant medicine”, or a mix of them.
• Statistics about how confident the model is (class probabilities).
• Feature conditions that led to that decision (the path of “if–then” rules).

So, the instant leaf (L_i) is designed to be the execution point where the system can accomplish the following:

• Identify which action(s) to take.
• Quantify how sure it is.
• Combine with FAHP or prioritize multiple possible actions.

The proposed hybrid decision-making approach is an effective means of enabling multiple processes to run concurrently. To illustrate the significance of the proposed hybrid decision-making process, a comparison of single vs. multiple classifications is described in

Table 2. Effectiveness of the hybrid decision-making system.

Mode	Purpose	Output at Leaf	Example
Single classification	One process per decision	One label (such as “irrigation”)	When soil moisture < 30% Then apply irrigation
Multiple classification (multi-label)	Several simultaneous processes are possible	Vector of labels or probabilities	When soil NPK is low and moisture is low apply irrigation and fertilization

.

The leaf enables parallel action detection and multi-process coordination. The proposed solution can also rank actions by probability and execute them in priority order.

6.2. Evaluation, Comparison, and Analysis

6.2.1. Prediction Model

This study investigates the reliability of the predicted burst data for both categories A (150) and B (6500). This investigation demonstrates its ability to detect reductions in the SOC feature, providing an effective early-warning system [54]. This assessment involves calculating accuracy, precision, and recall, as well as performing k-fold cross-validation. To compute these evaluation metrics, essential definitions such as true positive, false positive, false negative, and true negative are required.

• True positive TP: correctly predicted to have soil organic carbon reduction.
• False positive FP: incorrectly predicted to have soil organic carbon reduction.
• False negative FN: incorrectly predicted to have no soil organic carbon reduction.
• True negative TN: correctly predicted to have no soil organic carbon reduction.

Accuracy = (TP + TN)/(TP + FP + FN + TN)

(13)

Precision = TP/(TP + FP)

(14)

Recall = TP/(TP + FN)

(15)

The study employed Equation (2) to investigate the reduction in the SOC feature using real-time data parameters and then using predicted data parameters. The linear impact coefficient was determined as follows:

• Air temperature: −0.0598
• Relative humidity: +0.0280
• Soil moisture: +0.0954
• Soil temperature: +0.0283

This investigation resulted in two categories of true positive, false positive, true negative, and false negative scales, one for the real data and the other for the predicted data. The study then employs Equations (13)–(15) to deduce the system accuracy, precision, and recall for the first burst data category, with 150 measurements:

• True positive: 22 (Correctly predicted SOC reduction).
• True negative: 34 (Correctly predicted no SOC reduction).
• False positive: 3 (Incorrectly predicted a reduction—Type I Error).
• False negative: 1 (Incorrectly predicted stability when a reduction occurred—Type II Error).

The measured data versus the predicted values (56/60) achieved 93.33% accuracy; the training accuracy was 95%. Precision (22/25) was 88%, and recall (22/23) was 95.65%, indicating that the proposed system delivered high predictive performance.

To ensure that the system adapted correctly to varying environmental conditions, we applied K-Fold cross-validation (K = 5) across the entire 150-sample set.

• Mean K-Fold accuracy: 100% (indicating the DT successfully captured the rule-based logic of the system).
• Overfitting Analysis:

  ▪

  Training accuracy: 95%.

  ▪

  Testing accuracy: 93.33%

  ▪

  Observation: The slight 1.67% difference confirms that the model is well-generalized and not overfitting to the “burst” noise, allowing it to provide reliable early warnings.

For the second burst data category b, with 6500 measurements, we measured the following:

• True positive: 802 (Correctly identified SOC reduction events).
• True negative: 843 (Correctly identified stable or increasing SOC phases).
• False positive: 354 (Incorrectly flagged a reduction due to sensor noise, type I error).
• False negative: 401 (Failed to detect a real reduction during a 50-millisecond transition-type II error).

The measured data versus the predicted values (1645/2400) achieved 68.54% accuracy, training accuracy 73.9%, while the precision (802/1156) was 69.38% and the recall (802/1203) was 66.67%. The system maintained a high overall reliability, correctly classifying the SOC state in nearly 69% of high-speed measurements. The system captured two-thirds of all actual SOC reduction events within the high-frequency burst window.

K-Fold cross-validation (5-Fold)

The 6500 readings were processed through a 5-fold cross-validation to ensure the algorithm’s consistency across the entire 5.4 min monitoring window.

• Mean CV accuracy: 69.29%.
• Significance: The close alignment between the mean CV accuracy (69.29%) and the test accuracy (68.54%) proves that the model was statistically stable and provided reliable performance across different temporal segments of the burst.

A 95% Confidence Interval (CI): With a substantial test sample of 2400, the margin of error equals 1.86%. We are 95% confident that the true predictive accuracy for the high-speed detection system lies between 66.68% and 70.4%.

Overfitting Analysis

• Training accuracy (4100 samples): 73.90%.
• Testing accuracy (2400 samples): 68.54%.
• Gap: 5.36%.

The narrow gap indicates a robust model. The system successfully learned the underlying environmental patterns defined by the SOC Equation (2) without becoming overly sensitive to the specific noise profile of the training data. This balance is critical for real-time field deployment.

The move to 6500 readings provides a robust statistical baseline, confirming that the high-frequency 50 millisecond sampling strategy effectively tracked SOC micro-fluctuations with high confidence and minimal overfitting.

Additionally, evaluating the prediction system for climate change data and its effect on SOC levels is supported by calculating the Root Mean Squared Error (RMSE). RMSE measures the average magnitude of errors, and a lower RMSE indicates better accuracy. RMSE estimates the degree to which the data points align with the fitted regression line. The system achieved an RMSE value of 0.07.

$$\mathit{RMSE}=1/n\times {\sum }_{i=1}^n{(yi-\overline{yi})}^2$$

(16)

where

• n is the number of data points
• i is a counter from 0 to n
• y_i is the data value at the ith iteration
• $\overline{yi}$ is the average value for the ith data reading as $\overline{yi}$ = 1/n * ${\sum }_{i=1}^{n}yi$.

Mean Absolute Error (MAE) is a metric that measures the average size of errors in a set of predictions, ignoring their direction. MAE provides a clear indication of the extent to which predictions deviate from actual outcomes by averaging absolute errors. In this case, the system achieved a commendable MAE of 0.06, indicating high predictive accuracy.

$$\mathrm{MAE}=1/n* {\sum }_{i=1}^n{|yi-\overline{yi}|}^1$$

(17)

R-squared: The model explains the proportion of variance in SOC. A higher R-squared signifies a stronger regression model. The system achieved an R-squared value of 0.985.

$$R^2=1-[({\sum }_{i=1}^n{(yi-\widehat{yi})}^2)/({\sum }_{i=1}^n{(yi-\overline{yi})}^2)]$$

(18)

Table 3. Prediction model comparison with other techniques available in the literature.

	Proposed Solution	[55]	[56]	[57]	[58]
RMSE-Value	0.07	0.085	0.3	0.116	0.35
R-squared value	0.985	0.56	0.88	0.88	0.2

presents a detailed comparison of the system’s results with various techniques documented in the literature [55,56,57,58]. This comparison uses two key performance metrics: R-squared and RMSE. The R-squared measures the proportion of variance in the dependent variable explained by the independent variables. However, the RMSE quantifies the average magnitude of the errors between the predicted and observed values. These metrics provide valuable insights into the system’s effectiveness and accuracy relative to alternative methods.

The comparison highlights the effectiveness of the proposed solution, as evidenced by its achieving the highest R-squared and the lowest RMSE. This highlights the remarkable performance of the proposed system, which is summarized as follows:

• SOC stability: The regression constant (4.1823) serves as the baseline SOC level, effectively “confining” the complex, nonlinear soil variables that do not change as rapidly as air temperature.
• Climate sensitivity: Soil moisture has the highest linear impact (+0.0954) on SOC. This impact indicates that irrigation frequency is a primary adaptation to climate change.
• Early warning capability: Since the value is high (0.9376), the system can accurately predict expected SOC levels. Real-time measurements that deviate from the MLR prediction would trigger the “abnormal behavior” warning.
• System strategy efficiency: By running sensing every 10 min and the decision algorithm every 15 min, the system maintains a high-fidelity record for MLR validation. This approach also optimizes power for the computationally more-demanding FAHP validation steps.

6.2.2. Decision-Making

A rule-based labeling scheme was employed to train a DT classifier, which successfully categorized irrigation, fertilization, and plant protection needs with high accuracy. The resulting model exhibited clear, threshold-based decision boundaries. Such models enhance precision agriculture by converting raw sensor signals into interpretable, data-driven management decisions. The DT hyperparameters used in the simulation are max depth = 6, min sample split = 10, min sample leaf = 5, and random state = 42.

The confusion matrix for an autonomous controller node integrating a DT and the FAHP is not merely a metric of success but also a diagnostic tool for systemic reliability and resource efficiency.

The FAHP Hierarchy Levels are as follows:

• Goal: Accurately predict SOC micro-fluctuations.
• Alternatives: Medicine, Fertilization, Irrigation, and None (no further process).
• Criteria (Fuzzy Weights):
  • Soil Moisture.
  • Soil Temperature.
  • Air Temperature/Humidity.
  • Solar Exposure.

The matrix shown in Figure 9 visualizes where the model succeeds and where it might misclassify. The diagonal elements represent correct predictions.

The decision-making evaluation focuses on two critical dimensions:

1. Risk-Weighted Error Analysis (Type I vs. Type II Errors)

In smart agriculture, all misclassifications are not equal. The confusion matrix helps distinguish between errors that cost money and those that cost crops.

• Type I Error (False Positive): The system predicts “Medicine” when the plant is healthy, which is considered an Environmental Efficiency Failure. This results in chemical runoff and unnecessary costs, although the crop remains safe.
• Type II Error (False Negative): The system predicts “None” when “Irrigation” is critically needed. This is a Productivity Failure. This is considered more severe, as it directly impacts yield stability and food security.

2. The Hybrid Mitigation Factor (DT FAHP)

A key insight for the proposed system is the secondary validation provided by the FAHP.

• If the DT misclassifies a state (such as predicting “irrigation” due to a faulty humidity sensor reading), the FAHP layer acts as a mathematical filter.
• If other sensors (UV, air temp) do not support the “High Urgency” state, the FAHP weight () will be low.
• The final output will result in a negligible amount of water being dispensed, effectively mitigating the DT’s error.

The autonomous controller node’s ability to categorize environmental states was evaluated using a multi-class confusion matrix. The model demonstrated high classification robustness, particularly in the irrigation and fertilization classes. A critical finding of this study is the relationship between the DT’s classification certainty and the weight intensity assigned by the FAHP. While the DT provides a categorical “leaf-level” decision, the FAHP ensures that the magnitude of the action is proportional to the environmental stress level. This hybrid approach has a “dampening effect” on misclassification rates. For instance, if a sensor error leads to a false positive for irrigation but the FAHP determines the weight of the “temperature” and “soil moisture” criteria to be low (based on global consistency checks), the resulting water volume dispensed is minimal, thus preserving resource integrity despite a localized logic error. Moreover, the confusion matrix confirms that the system maintains a specificity of over 118 for the “None” (optimal) state. In a traditional threshold-based system, sensor noise often triggers unnecessary actuator activity. By employing the DT’s hierarchical logic, the controller node acts as a high-pass filter, ensuring that actuators are engaged only when a combination of factors (such as temperature, UV, and soil moisture) collectively indicates the need for action.

6.3. Model Performance Versus Deployment Feasibility

Another vital discussion of the rationale for selecting the DT algorithm involves comparing it with related methods in the literature. Figure 10 compares several machine learning techniques reported in the literature with the proposed hybrid decision-making system.

Figure 11 compares the proposed hybrid classification method with four supervised learning models in terms of precision. These techniques are used to make decisions based on sample data in MATLAB. The neural network model outperforms the others with 95% accuracy, consistent with findings in the literature on neural networks or CNN-based architectures in agriculture [59]. Random Forest achieves a strong balance, achieving 93% accuracy due to ensemble voting and robustness against overfitting [60]. SVM achieves 91% accuracy and performs well on high-dimensional decision boundaries [61]. The proposed DT trails at 89%, reflecting its susceptibility to overfitting and limitations in modeling complex feature interactions. Although DTs are less accurate, they remain relevant due to their low computational costs and high interpretability, making them ideal for real-time systems or edge deployment. However, the proposed hybrid decision-making approach achieved a 7% post, which is still higher than 96%, and is more lightweight. Moreover, the accuracy metric alone is insufficient for making deployment decisions in embedded systems; other factors such as latency, power, and interpretability are also important. Figure 11 illustrates the fundamental trade-off between accuracy and complexity, providing clearer insight into the superiority of the proposed hybrid DT-FAHP method, even relative to its native DT.

The results shown in Figure 11 illustrate the trade-off between deployment cost and performance gain, as reflected in the balance between model accuracy and computational complexity. DT achieves reasonable accuracy (0.87) with minimal computational complexity, making it suitable for low-power devices such as the ESP32 and Raspberry Pi, as well as for highly scalable sensing applications. Random Forest improves accuracy (0.94) and has moderate complexity. SVMs achieve substantial accuracy (0.90) but exhibit high training and inference complexity on larger datasets. Neural networks achieve the highest accuracy (0.96), although they are the most complex. The proposed hybrid decision-making achieved a higher score of 0.97. This hybrid decision-making approach is slightly more complex but remains lighter than alternatives. This analysis typically favors hybrid DT-FAHP for real-time, on-field decisions. Training was performed on a dataset of 60 samples from 150 entries and 2400 samples from 6500 entries, with 8 iterations. The results of the training runs using diverse machine learning techniques are summarized in

Table 4. Comparing the proposed model against prior ones according to complexity.

Model	CPU Usage (%)	RAM Usage (MB)	Training Time (S)	Interference Time (ms)	Interpretability	Scalability
Proposed hybrid DT-FAHP	15–25	50–70	0.5–2	1.9235	High	Moderate
Random Forest [60]	35–50	100–180	5–15	5–15	Medium	High
SVM [61]	50–70	150–250	10–30	10–25	Low	Low-Moderate
Neural Network [59]	60–90	250–500	20–90	3–10	Low	Very High

.

Table 4 offers insights into comparing different models using the following metrics:

▪

CPU Usage (%): Average processor load during training or inference on a typical edge device.

▪

RAM Usage (MB): Memory footprint during operation.

▪

Training Time: Duration needed to train the model on a medium dataset (90–150 samples), (4100/6500).

▪

Inference Time (ms): Time to generate a single prediction (critical for real-time systems).

▪

Interpretability: The human understanding level of the decision process.

▪

Scalability: How well the model adapts to large datasets or complex features.

The results presented in Table 4 demonstrate that rapid training makes DT-FAHP well suited to dynamic learning and on-site model retraining. Neural Networks and SVMs are better suited to offline training followed by edge deployment. Inference speed was measured for a single-sample prediction, as it is critical for real-time control in this context, particularly for irrigation and fertilization processes. DT-FAHP enables deployment at the microcontroller level, whereas neural networks can be optimized with TensorFlow Lite for real-time applications. The interpretability of decision trees is high, making it easy for users to understand decision paths. For the random forest technique, interpretability is moderate, and decisions are less clear. SVMs are challenging to interpret due to their high-dimensional kernel space. From a scalability perspective, DT-FAHP performs well on medium-sized datasets but struggles with noise and overfitting. Both Random Forests and neural networks scale efficiently through distributed training or GPU acceleration. However, SVMs quickly become impractical for large datasets due to their O(n²) memory and time complexity. This outstanding performance underscores the robustness and reliability of the proposed solution relative to a human-dependent system, as shown in

Table 5. Comparing the proposed model with human-dependent ones.

Model	Multiple Data Processing	Probability of a Wrong Decision	Comment
Proposed autonmous	Up to three	1.5–6%	A wrong decision is probable due to errors in data prediction
Human-dependent	Only one	66%	A wrong decision is probable due to inefficiency in providing the right decision for heterogeneous data

.

Table 5 offers insights for comparison with human-dependent systems, which lack the efficiency to provide accurate decisions on heterogeneous data. The results also highlight the effectiveness of our approach in mitigating the impacts of climate change on soil performance. The proposed prediction approach achieved 98.33% accuracy and an error rate of 1.67% on 150 bursts. The system also achieved 68.54% accuracy and an error rate of 5.36% on 6500 bursts. The slight 1.67–5.36% difference confirms that the model is not overfitting to the “burst” noise, thereby enabling reliable early warnings.

Moreover, our system effectively predicts essential physical parameters, thereby eliminating the delays typically associated with traditional data collection and analysis methods. This strong predictive ability, as illustrated in Figure 5, Figure 6 and Figure 7, offers a promising approach to regenerative agriculture. The results illustrated in Figure 8 confirm its ability to provide proactive action measures for regenerative agriculture strategies.

Based on the analyses in Figure 10 and Figure 11 and the comparative data in Table 4, we conclude that the hybrid-DT-FAHP is highly interpretable and lightweight. This lightweight merit supports deployment due to its low hardware requirements. This lightweight advantage also offers improved scalability and reliability, as well as making it an economic solution. The proposed model is promising as a scalable approach to distributed agentic AI decision-making across multiple 300 × 300 square meter areas. The high precision of the proposed solution makes it more efficient under normal conditions at minimizing water and energy losses. This accuracy enables adjustments of the irrigation period, fertilizer application rate, and precise medication treatment. Finally, the proposed solution maximizes the edge rule by autonomously managing multiple agricultural processes. This extends to serving as an early-warning system and a responder.

Compared with prior autonomous edge solutions, it achieved better results, as shown in

Table 6. Comparing the proposed solution against solutions from the literature.

Ref.	Sensing Parameter	Controller Type	AI Methods	Agriculture Processes	Climate Adaptation	Prediction	Human- Dependent Decision	Back-End System	Crop Type/ Specialty
Proposed solution	Air Temperature, Humidity, and UV Index. Soil Temperature and Moisture.	IoT Esp32	Agentic AI with DT, FAHP, and Regression.	Irrigation Fertilization Plant remediation	Adapt with respect to SOC reduction	Prediction model utilizing MLR	No totally autonomous	No totally autonomous	Orchard
Shah B. [62]	Air Temperature, Humidity, UV Index, and Rainfall. Soil Temperature and Moisture.	Hunter controller	Not applicable	Irrigation Fertilization	Not considered	Not used	Scheduling dependent	Web platform	Horticulture
Gómez-Flores JL [63]	Soil Temperature, Moisture, and Salinity. Airborne Imagery.	Hunter controller	Not applicable	Irrigation Fertilization	Not considered	Not used	Scheduling dependent	Not available	Orchard
Abo-Zahhad [64]	Air Temperature, Humidity, UV Index. Soil Moisture and Ph.	Esp32 connected to Mobile Application	KNN	Irrigation	Preset adaptation	Preset Weather Forecasting	Fully dependent	Web server	Cereal
Gueye M [65]	Air Temperature and Humidity, UV Index, and Rain. Soil Temperature, Moisture, EC, and pH. High-Resolution Camera.	Arduino Uno	Computer vision	Irrigation Fertilization Plant remediation	Not Considered	Not used	Data acquisition, especially for plant photos	Local Computer with an AWS cloud web connection	Horticulture

. The solution presented in [62] provided autonomous irrigation management using the Hunter controller. However, this solution relied on a human expert to evaluate sensor data and manually schedule irrigation. The solution provider acknowledged that it lacks AI and machine learning techniques to improve the analysis of heterogeneous data. This solution demonstrated that changing the irrigation strategy has a positive impact on agricultural production. However, this strategy is planned and imposed without consideration of prediction or climate change. The solution in [63] employed the same approach, extending it to fertilization. Fertilization was provided alongside irrigation in an automated manner based on manual scheduling by the controller. Beyond that, it provided ongoing soil monitoring and observation that informed irrigation and fertilization scheduling. This solution demonstrated the impact of strategy changes on agricultural outcomes by comparing its performance with that of conventional agriculture. However, this solution still did not account for climate change impacts or for data prediction, nor did it enable the fully orchestrated system operation that would allow the process to be managed autonomously without human intervention. The solution provided in [64] accounted for climate change and weather forecasting. In addition, this solution provided a mobile application for monitoring and controlling the agricultural process. However, the solution did not provide a comprehensive soil assessment and did not enable autonomous management; it remained human-dependent. Weather predictions were maintained in a preset menu rather than through a prediction mechanism. In addition to this human dependency, the solution relied on human factors to load the plant preset, weather data, and even the control. The worst drawback was that it still depended on the internet through a backend web service. The solution proposed in [65] relied more heavily on a local computer as a backend server. However, this solution still depended on AWS-based cloud services to maintain the agricultural process decisions. However, this solution provided management for irrigation, fertilization, and plant remediation. The capability for cloud AI management and data analysis, including plant images, underpinned this management capability. However, it lacked orchestration of system elements and relied on human judgment. Moreover, the solution did not provide predictions for data, climate change, or soil impacts.

The proposed solution enabled heterogeneous data analysis and the prediction of data bursts, thereby effectively assessing climate impacts on soil. In addition, ensuring autonomous real-time management is a remarkable achievement of the proposed solution.

7. Conclusions

This paper proposed an autonomous agricultural system as an early-warning tool for assessing the impacts of climate change and providing appropriate actions. The proposed solution examined the SOC reduction, considering both linear and non-linear behavior. The proposed solution incorporated this behavior into a predictive model by employing a modified MLR to predict SOC for the period of 2025–2050. The proposed solution developed a prediction model based on measured heterogeneous sensor parameters. Moreover, the proposed solution evaluated the prediction model parameters against real-time measurements to synthesize model results. The proposed solution then turned the predicted model results into an actionable process by employing a hybrid DT-FAHP model. The hybrid DT-FAHP model employed a DT at the leaf layer and assessed the leaf weight through the FAHP. This assessment through the FAHP provided a significant advantage, not only for managing multiple methods, such as irrigation, fertilization, and plant remediation, but also for supporting a quantitative action plan. The proposed solution is a new framework for the intelligent edge era that leverages AIoT to enable an agentic AI framework. The system supports both single- and multi-label classification, enabling the simultaneous or sequential execution of multiple processes based on the tree’s probabilistic outputs. The system evaluation achieved an overall accuracy of 93.33% for 150 burst-based predictions and 68.5% for 6500 burst-based predictions. In terms of precision, it gained a score of 0.5, indicating that it was effective at detecting true positives. A perfect recall rate of 100% confirmed its reliability in identifying all actual cases without missing any. Furthermore, the system’s predictive ability was excellent, with an R-squared of 0.985, indicating a near-perfect fit between predicted and observed outcomes. The model also had an RMSE of 0.070 and a mean absolute error (MAE) of 0.057, demonstrating its accuracy in estimating agricultural parameters. In addition, the system is robust in adapting to improve its prediction performance by resampling real-time data. This adaptability is automated as well, which ensures its agentic feature and its ability to adapt and sustain. The system is also agile, as it can adjust its configuration based on expert-level evaluation conducted on the farm. Both features, in addition to their IoT nature, make it highly adaptable and fast-convergent. Evaluating the DT-FAHP model indicated a balance between light weight and accuracy, making it suitable for devices with limited resources. The proposed solution is promising for achieving outstanding results in modern agricultural fields that adopt smart agriculture systems. The proposed solution is limited in scale to a multiple-unit area of 300 × 300 square meters, can be used for orchard planting, with a maximum of three agronomic processes, and can be implemented on a resource-constrained edge device. As a direction for future work, we propose assessing a more lightweight prediction mechanism in addition to assessing the edge device’s sustainable operation for full annual management. Moreover, the device’s scaling with multiple intelligent edge solutions that span large areas could be investigated.