Water Footprint Through an Analysis of Water Conservation Policy: Comparative Analysis of Water-Intensive and Water-Efficient Crops Using IoT-Driven ML Models

Abstract

Although economic profitability and food security often outweigh water conservation priorities in arid and semi-arid regions, this study investigates irrigation practices by evaluating water footprint and economic feasibility through a comparative analysis of water-intensive and water-efficient crops. In this context, an optimal irrigation disparity framework integrated with Internet of Things (IoT) and Machine Learning (ML) mechanisms is proposed to evaluate the effectiveness of water conservation, thereby assessing the potential for enhancing economic profitability. IoT-enabled components are employed to monitor real-time environmental—soil moisture, temperature, and weather—conditions between March and November 2023. This data is processed using a hybrid modeling approach that integrates KNN, GBT, and LSTM algorithms to predict both the duration of cultivation and the water requirements. Finally, the predicted parameters are incorporated into a multi-objective framework aimed at minimizing the disparity in water allocation per net benefit. The final results indicate that saffron required substantially less water—ranging from (19.87 to 28.65 ∗ 10⁶ m³)—compared to watermelon, which consumed (34.61 to 47.07 ∗ 10⁶ m³), while achieving a higher average net profit (33 ∗ 10⁹ IRR) relative to watermelon (31 ∗ 10⁹ IRR). Moreover, saffron consistently approached optimal values across disparity-based objective functions, averaging (0.404). These findings emphasize the dual advantages of saffron as a value-added, water-efficient crop in achieving substantial water conservation while enhancing profitability, offering actionable insights for authorities to incentivize water-efficient crop adoption through subsidies, market mechanisms, or regulatory frameworks. These strategies operationalize technical insights into actionable pathways for balancing food security, economic growth, and environmental resilience.

1. Introduction

Water serves as a vital element for agricultural irrigation, thereby significantly contributing to global food security [1,2]. In recent years, the persistent rise in the water requirement for non-agricultural purposes, including urban and industrial users, coupled with greater concerns regarding global climate change, has led to the strategic management of water requirement for irrigation [3,4] Due to advancements in computing mechanisms, sophisticated technologies, and enhanced equipment, meta-heuristic mechanisms have been highlighted within the domain of conservatism policies [5,6,7].

In agriculture, various studies have recently investigated approaches to integrate advanced IoT techniques for data transmission with various ML models to predict and analyze data, thereby facilitating the management of irrigation processes (Figure 1).

For instance, Godwin & Johnson [8] used field-based models to balance yield and water use, but did not include economic policy levers. Similarly, Kunt [9] and Pargo et al. [10] introduced efficient soil-moisture control, yet they omit crop-intensive footprint metrics and optimization modules. Lakhiar et al. [11] highlighted the potential of precision irrigation technologies to enhance water-use efficiency, boost crop yields, and minimize environmental impacts. The study emphasized the role of IoT, AI, and sensors in optimizing irrigation practices under variable climatic conditions, thereby contributing to sustainable water management. Wilberforce and Mwebaze [12] developed an IoT-enabled smart agriculture framework using Raspberry Pi and multi-sensor integration to provide real-time environmental monitoring for crop management in Uganda. Their work highlighted the value of sensor-driven architectures in addressing weather unpredictability and improving irrigation practices. Bernardo et al. [13] proposed an ANFIS (Adaptive Fuzzy Inference System) algorithm considering the use of the IoT not only to monitor and manage the irrigation system, but also to enhance agricultural yields by adapting the decision-making process of the automated irrigation system.

Recent advancements in precision irrigation have underscored the necessity of integrating spatial–temporal decision tools with crop-specific water-demand assessments to inform conservation policy. Despite the proliferation of ML and IoT frameworks for irrigation scheduling, relatively few studies have leveraged high-resolution predictive modeling to assess the distributive efficiency of water allocation across crops with divergent water-use profiles. In particular, the dynamic interplay between real-time irrigation responses and long-term water conservation outcomes remains underexplored in the context of value-based crop selection. While prior research has emphasized yield maximization or resource optimization in isolation, insufficient attention has been paid to evaluating water footprint trade-offs between water-intensive and water-efficient crops under climate change and market conditions.

To bridge this gap, the present study proposes a predictive framework that integrates IoT-enabled environmental monitoring with hybrid ML modeling to generate crop-specific irrigation forecasts. Specifically, to address the interrelated challenges of spatial variability in water distribution and temporal irrigation scheduling, this study utilizes three ML models, including K-Nearest Neighbors (KNN), Gradient Boosted Trees (GBT), and Long Short-Term Memory networks (LSTM). The KNN algorithm is employed to estimate localized environmental parameters by identifying nearest-neighbor relationships in the input space. The GBT model captures complex nonlinear interactions to enhance the accuracy of water requirement prediction at specific decision points. Meanwhile, the LSTM model, with its capacity to model sequential dependencies, is applied for time-series prediction of irrigation needs across different intervals. By leveraging the complementary strengths of these models, the proposed framework ensures robust and dynamic prediction of the aforementioned parameters in varying agricultural conditions. Subsequently, the predicted irrigation data are incorporated into an optimization framework that evaluates trade-offs between crop-specific water usage and economic profitability, enabling a comparative assessment of water-conservation policies under alternative cropping scenarios.

2. Materials and Methods

The correlation between water consumption and water conservation can be conceptualized as an inverse relationship [14,15]. A larger water footprint typically implies higher water use and potentially less conservation, while a smaller water footprint suggests better water conservation efforts [16,17]. In agricultural irrigation systems, water consumption in upstream regions can influence the availability of resources for downstream consumers, particularly in contexts where water management practices are suboptimal or unevenly enforced. Although various governance mechanisms and efficiency-enhancing strategies exist, their implementation may vary across regions, resulting in discrepancies in water-use efficiency. The increased water footprint upstream leads to over-extraction, reducing flow, and contributing to disputes over equitable water distribution in shared systems. Conversely, farmers prioritize their economic gains, often disregarding the critical water demand and the conflicts arising from irrigation timing. In this context, this study proposes a multi-objective approach to pursue comparative water footprint between competing crops, thereby evaluating water conservation strategy and economic benefits. To address the aforementioned conflicting factors, this research uses IoT and ML mechanisms to predict cultivation duration and the volume of water requirement. Next, the surrogate ML model is incorporated into the proposed multi-objective framework to analyze water footprint implications between water-intensive and water-efficient crops aimed at evaluating water conservation policy (Figure 2).

In this context, data collection and transmission are facilitated through Wireless Sensor Networks (WSN) for real-time monitoring. Additionally, data epochs are captured at intervals of 12 h, enabling month-wise predictive analysis based on the collected time-series data.

Then, ML models, namely, KNN, GBT, and LSTM, are compiled for data processing sourced from the IoT sensors to conduct predictive analysis of water requirements and duration of cultivation. This proposed method is implemented as follows:

• KNN (K-Nearest Neighbors) Algorithm

The KNN algorithm is employed to identify the closest values relative to the plant’s root with regard to predicted sensors. The KNN training dataset consists of features represented in a multidimensional space for each class of supervised label data. The input comprises the $K$ nearest values within the resource dataset. The predicted output class is established by consolidating the votes from its closest predictions. The equation representing the prediction of nearest neighbors with varying parameters is outlined as follows:

$$Y=Z_i^w(c_1,c_2,\dots c_n)$$

(1)

where $Y$ refers to the predicted output, $Z$ is the closest feature, $w$ refers to the weight, $i$ represents the number of features, and ($c_1$,$c_2$,...$c_n$) shows various input features. In this study, the prediction values are influenced by the nature of the features’ distances, whether continuous or discrete. The relationship between the predictions for different features is determined using Euclidean Distance:

$$D=\sqrt{{\displaystyle {\sum }_{i=0}^n(c_i-c_j)}}$$

(2)

where $D$ represents the distance,$n$ represents the number of data points associated with specific features, and the indices $i$ and $j$ denote the starting and ending points, respectively.

2.

LSTM (Long Short-Term Memory) Algorithm

The LSTM, as a deep learning framework, is employed for classifying and predicting data derived from time series facilitated by time-interval data. The LSTM architecture comprises an input gate, a forget gate, and an output gate. The cell retains the time interval, and the gates play a crucial role in regulating the flow of information obtained through input and output operations. However, the forget gate function is defined as follows:

$$f_t={\sigma }_t(w_f{c}_t+u_f{h}_{t-1}+b_t)$$

(3)

where $h_{t-1}$ denotes the hidden state vector from the previous time step ($t-1$), $c_t$ is the input at the current time step, $b_t$ is the bias parameter, and ${\sigma }_t$ represents the sigmoid function. The sigmoid function outputs values between 0 and 1, with a value of 1 indicating that the information is retained, and a value of 0 indicating that the information is forgotten:${\sigma }_t={\displaystyle \frac{1}{1+e^t}}$.

3.

GBT (Gradient Boosted Tree) Algorithm

The GBT regression represents a machine learning algorithm utilized for both classification and prediction, generating a real prediction tree. Gradient boosting employs decision trees as base learners, wherein it combines multiple weak and strong learner bases through an iteration process. The $GBT$ utilizes a decision tree at the $m-th$ step, denoted as $d_m(c)$, to the pseudo-residual. The decision tree is divided into leaves, each comprising distinct spaces $(S_1,S_2,\dots S_{Nm})$, along with prediction values assigned to each space. The output of $d_m(c)$ for diverse input features ($v$) is as follows:

$$d_m(c)={\displaystyle \sum _1^{Nm}{b}_m{S}_m(c)}$$

(4)

where $b_m$ is the predicted value of each space $S_m$. The $GBT$ model implements a stage-wise fashion, iteratively enhancing the optimized model at each stage.

Subsequently, the integration of these three algorithms forms a hybrid model provided by Vianny et al. [7] to generate predictions for water requirements and irrigation scheduling. This approach optimally combines the strengths of each algorithm to improve prediction accuracy and reliability. The proposed hybrid workflow structure simultaneously employs the three specified algorithms to analyze and predict non-parametric values. Next, Spearman’s rank correlation is employed to validate the data by examining the correlation coefficient ratio (Figure 3).

In this context, the predicted values generated by the $KNN, GBT$, and $LSTM$ models are denoted as $KN{N}_{SR}^n$, $GB{T}_{SR}^n$, and $LST{M}_{SR}^n$, respectively, where $n$ represents the number of input features utilized in each prediction. The corresponding observed data series is referred to as $R_n$. Furthermore, beginning from the second 12 h interval, the model executes a repeated selection process across each hourly loop ($hr$). For validation, Spearman correlation coefficients are computed between each model’s prediction and the observed data. The model achieving the highest correlation is selected as $Bes{t}_{SR}^n$, and its output is subsequently fed into the hybrid prediction framework. Additionally, an interception term $I_{hr}$, derived from the correlation context, is incorporated into the decision framework to iteratively refine the hybrid model’s output sequence over time.

Input: $KNN, GBT$, $LSTM$, $R_n$, $and parameters$ (n).

Output: Hybrid prediction values.

$$\begin{array}{l}Begin\\ Interval\leftarrow 1\\ While(Interval\le 12)do\\ KN{N}_{SR}^n\leftarrow Spearman(KN{N}_{SR}^n,R^n)\\ GB{T}_{SR}^n\leftarrow Spearman(GB{T}_{SR}^n,R^n)\\ LST{M}_{SR}^n\leftarrow Spearman(LST{M}_{SR}^n,R^n)\\ R_{SR}\leftarrow \left\{KN{N}_{SR}^n,GB{T}_{SR}^n,LST{M}_{SR}^n,R^n\right\}\\ Bes{t}_{SR}^n\leftarrow Max\left\{KN{N}_{SR}^n,GB{T}_{SR}^n,LST{M}_{SR}^n,R^n\right\}\\ I_{hr}\leftarrow \left\{Bes{t}_{SR}^n,R_{SR}\right\}\\ hr\leftarrow hr+1\\ While(hr\le 12)do\\ Hybri{d}_{hr}^n\leftarrow I_{hr}*Bes{t}_{SR}^n\\ End\end{array}$$

Next, an optimal multi-objective framework is developed to ensure the efficiency and sustainability of water distribution, thereby comparing how water footprint impacts water conservation policies. In the following, three objective functions are implemented:

OF1: The first objective function seeks to minimize the disparity of water supply per unit of net benefit within an irrigation system [2]:

$$\underset{X_j^t}{\min } U^1={\displaystyle \sum _{j=1}^m{\displaystyle \sum _{j^{\prime }=2}^m\left|\frac{X_j^t{A}_j^t}{E_j^t}-\frac{X_{j^{\prime }}^t{A}_{j^{\prime }}^t}{E_{j^{\prime }}^t}\right|}}$$

(5)

where $j$ is the number of the crop, $X_j^t$ is the volume of water supply (decision variable), and $A_j^t$ is the assigned area for crop $j$ at time $t$. The decrease in the objective function $U^1$ towards zero indicates the equitable allocation of water per unit of net benefit among sub-areas. Also, ${\displaystyle \frac{X_j^t{A}_j^t}{E_j^t}}$ shows the simulation of achieving the optimal supply pattern to “grow more crop with less water”, which refers to the amount of water supplied per unit of economic profit. $E_j^t$ shows the net benefit with regard to the parabola model, as follows [18,19]:

$$\begin{array}{l}{E}_j^t={\displaystyle \sum _{j=1}^m({\mu }_j.P_j^t.A_j^t-\rho .X_j^t.A_j^t)}\\ t=1,2,\dots T\end{array}$$

(6)

where ${\mu }_j$ is a benefit parameter for crop $j$ per unit of allocated water, $P_j^t$ is the quantity of yield per unit of acreage, and $\rho$ refers to the price of water.

OF2: The second objective function aims to minimize the disparity of the gap between water requirement and water supply per unit of water requirement between multi-crops [20]:

$$\underset{X_j^t}{\min }\hspace{0.17em}{U}^2={\displaystyle \frac{1}{T}}{\displaystyle \sum _{j=1}^m{\displaystyle \sum _{j^{\prime }=2}^m\left|\frac{\max (R_j^t{A}_j^t-X_j^t{A}_j^t,0)}{R_j^t}-\frac{\max (R_{j^{\prime }}^t{A}_{j^{\prime }}^t-X_{j^{\prime }}^t{A}_{j^{\prime }}^t,0)}{R_{j^{\prime }}^t}\right|}}\ast 100\%$$

(7)

where $R_j^t$ refers to the water requirement for crop $j$ at time $t$, and $T$ is the total operation year. However, the value of $U^2$ fluctuates between 0 and 1, with a tendency towards 0 indicating a lack of scarcity and an equitable irrigation system.

OF3: The third objective aims to minimize the inequality in the ratio of the total distribution of water per ratio of available water resources [21]:

$$\underset{X_j^t}{\min U^3}={\displaystyle \sum _{j=1}^m{\displaystyle \sum _{j^{\prime }=2}^m\left|\frac{X_j^t}{W_j^t}-\frac{X_{j^{\prime }}^t}{W_{j^{\prime }}^t}\right|}}\ast 100\%$$

(8)

where $W_j^t$ is the volume of available water for crop $j$ at time $t$. In total, the tendency of $U^3$ to 0 indicates the least disparity.

Constraints:

The amount of water stored in the reservoir $R$ is calculated based on the previous period’s water level and the inflow range $I$, as expressed below:

$$R_i^{i+1}=\min [I_i^t+R_i^t-{\displaystyle \sum _i{X}_i^t},\overline{R_i]}$$

(9)

The rate of available water in the reservoir must be between the minimum and maximum storage of the reservoir:

$$R_{\min }^t\langle R^t\langle R_{\max }^t$$

(10)

The amount of water distributed to multi-crops cannot exceed the rate of available water:

$$0\le {\displaystyle \sum _{i=1}^n{X}_i^t}\le R^t$$

(11)

Lastly, Evolutionary Algorithms (EAs) are effective due to their capability to address complicated models characterized by features like discontinuities, multi-modality, disjoint feasible spaces, and noise during evaluating functions. Indeed, these algorithms aim to generate a Pareto-approximate set of solutions, serving as a reference for decision-making in multi-criteria problems [22]. Thus, this research applies (EA) algorithms, including the non-dominated general sorting algorithm II (NSGA II) and the multi-objective evolutionary algorithm based on decomposition (MOEA/D) as optimizers to generate the best feasible optimization outputs [23,24].

3. Case Study

The Neyriz Basin (29°11′39.62″ N to 54°19′11.60″ E) in Fars Province, Iran (as shown in Figure 4), is primarily fed by the Kor River [25]. The Kor River is the main water source in the region and plays a crucial role in supporting agriculture, especially around the city of Marvdasht [26]. The Neyriz region is situated near Bakhtegan Lake, which is one of the largest lakes in Fars Province and was historically fed by the Kor River [27]. However, the situation for both the river and the lake has been deteriorating for years due to a combination of factors, including water over-extraction for agriculture, reduced rainfall, and mismanagement of water resources [28]. The water resources in Fars Province, particularly within the Neyriz Basin, face significant challenges [29]. The region experiences widespread water scarcity attributed to excessive groundwater extraction, the construction of dams along the Kor River, and a persistent decline in rainfall. These factors have collectively resulted in severe agricultural impacts and notable ecological degradation [30].

This study analyzes the cultivation of essential crops like watermelon to support food systems in the study area, alongside recent consideration by decision-makers to incorporate value-added, water-efficient crops, such as saffron, to assess potential revenue gains. The study compares these two crops in terms of water conservation strategies and the economic benefits they yield.

Data for this study was collected through continuous IoT-based monitoring from two active cultivation zones (Marvdasht and Neyriz) between March and November 2023. Rather than a fixed-plot experimental design with replications, the framework relies on time-series environmental data collected at 12 h intervals across heterogeneous farm conditions (

Table 1. Basic requirements of watermelon as well as saffron.

Sr No	Species Name	Values Watermelon/Saffron	Sr No	Species Name	Values Watermelon/ Saffron
1	Max. water requirement	600/300 mm per growing season	5	Number of nodes	3000/500,000 per hectare
2	Duration of Planting and Harvesting Period	Mar/June 2023	6	Coefficients in water production function (a, b, c)	−0.0199, 29.690, −5185/−0.1303, 63.18, −6817
3	Max. water requirement/month	150/65 Liters	7	Price (109∗IRR/kg)	0.018/0.66
4	Watering Schedule Duration	6/20 days

and

Table 2. Summary of various parameters in terms of predictive analysis.

	Mar Wat/Saf	Apr Wat/Saf	May Wat/Saf	June Wat/Saf	July Wat/Saf	Aug Wat/Saf	Sep Wat/Saf	Oct Wat/Saf	Nov Wat/Saf
Root moisture (%)	0.84/0.51	0.81/0.53	0.88/0.56	0.74/0.57	0.82/0.5	0.73/0.5	0.78/0.49	0.73/0.43	0.7/0.47
Ambient moisture (%)	0.63/0.38	0.59/0.32	0.65/0.41	0.51/0.39	0.64/0.47	0.58/0.43	0.62/0.51	0.53/0.46	0.51/0.30
Humidity (%)	0.31/0.29	0.35/0.32	0.38/0.39	0.38/0.27	0.25/0.31	0.27/0.36	0.18/0.27	0.21/0.28	0.25/0.32
Rainfall (mm)	49/37	34/41	22/38	28/24	18/26	16/19	21/13	26/17	19/11
Maximum watering (Lt)	35/10	35/10	35/10	40/10	40/5	35/10	25/5	25/10	25/5

).

4. Results and Analysis

4.1. Analysis of Data Pre-Processing

To predict the required data for irrigation scheduling, key parameters incorporated into the analysis included weather patterns, sub-area environmental conditions, soil temperature, and soil moisture levels. Figure 5 presents the relationship between the predicted and observed irrigation requirements, based on these four primary input parameters. All input variables are normalized using Z-score standardization before model training to ensure consistent scaling. Although the model demonstrates a strong fit for soil-related variables—specifically soil moisture and temperature—with an $R^2$ value exceeding 0.90, a higher degree of dispersion is evident in predictions driven by weather pattern data, as indicated by a comparatively lower $R^2$ value. This indicates a potentially weaker or more complex association between weather variables and short-term irrigation demand, underscoring the non-linear and variable nature of climatic influences on irrigation requirements within the predictive modeling framework.

The data pre-processing involved data collection, training, testing, and validation. Data were gathered at 12 h intervals using LSTM from two sub-areas of the cultivation field. For each crop, the Spearman correlation coefficients (ρ) vary across models (Figure 6), reflecting each algorithm’s performance in capturing the ranking order of water requirements under different environmental conditions.

For watermelon, the KNN model shows a moderate correlation with actual values (ρ $=0.86$), indicating some ability to maintain relative water requirements but limited by potential sensitivity to local variations in data. GBT exhibits the lowest Spearman correlation (ρ $=0.83$), suggesting that it captures complex patterns in environmental and soil-related variables, making it better suited for rank-order consistency in water predictions. LSTM demonstrates the highest correlation (ρ = 0.91), indicating strong performance in detecting temporal patterns. This model’s ability to consider sequences in data makes it advantageous for dynamic irrigation needs, where water requirements fluctuate over time. For saffron, the results follow a similar trend, with the LSTM model achieving the highest (ρ = $0.96$), underscoring its proficiency in managing data with temporal dependencies. The KNN model again outperforms GBT (ρ = $0.90$), though the gap between models is smaller, indicating that saffron water requirements might have fewer complex dependencies compared to watermelon.

To incorporate these complementary figures, the variance and performance of the most consolidated data were evaluated using $R^2$ and $RMSE$ metrics. The correlation coefficient averaged 0.70 for watermelon and 0.75 for saffron over a 12 h interval, indicating strong predictive accuracy. As the number of epochs increased, values for $R^2, RMSE$, and $PT$ also improved, stabilizing with fewer epochs. Overall, there were no substantial differences in terms of variance across sub-areas (

Table 3. Performance of time metrics for GBT.

Epochs	Zone1 S.C Wat/Saf	R2 Wat/Saf	RMSE Wat/Saf	PT(s) Wat/Saf	Zone2 S.C Wat/Saf	R2 Wat/Saf	RMSE Wat/Saf	PT(s) Wat/Saf
12	0.72/0.78	0.94/0.89	0.439/0.689	13/12	0.74/0.69	0.94/0.91	0.354/0.587	15/12
24	0.71/0.81	0.95/0.93	0.367/0.721	14/14	0.72/0.87	0.92/0.91	0.248/0.604	14/18
36	0.71/0.69	0.93/0.91	0.432/0.584	22/26	0.73/0.84	0.91/0.94	0.531/0.467	27/23
48	0.72/0.76	0.97/0.91	0.512/0.609	36/32	0.77/0.67	0.95/0.87	0.543/0.671	28/31
60	0.71/0.75	0.96/0.91	0.722/0.612	34/35	0.75/0.71	0.94/0.89	0.615/0.74	26/24
72	0.68/0.79	0.99/0.92	0.457/0.719	35/30	0.69/0.76	0.98/0.91	0.551/0.498	38/29

Note: R²: coeffiecient of determination, S.C: spearman correlation, RMSE: Root Mean Square Error.

).

4.2. Analysis of Water Requirements Variability for Crops Across Time Frame

Figure 7 evaluates water requirement predictions for watermelon and saffron crops across 200 solutions and multiple time periods using GBT, LSTM, and a hybrid model. The water requirement rate exhibits substantial fluctuations at the beginning and end of the period. For instance, the initial average water requirement for watermelon in zone $1$ is less $(3\ast {10}^6{\mathrm{m}}^3)$ under LSTM and hybrid models, whereas the corresponding volume under GBT alone reaches approximately $(3.4\ast {10}^6{\mathrm{m}}^3).$ In contrast, saffron shows bimodal water requirement with peaks in April and October, ranging from $(2.0\ast {10}^6{\mathrm{m}}^3)$ to $(3.5\ast {10}^6{\mathrm{m}}^3)$, aligning with its unique growth cycle. The LSTM model captures temporal variability more dynamically for both crops, while the hybrid model refines predictions, reduces noise, and maintains stability. Overall, watermelon exhibits greater temporal consistency, while saffron highlights higher variability due to its seasonal growth.

Figure 8 illustrates the month-wise water requirements for watermelon and saffron crops during their cultivation period. According to the results, watermelon shows a consistent rise in water requirements, peaking in July $\left(3.8\ast {10}^6{\mathrm{m}}^3\right),$ while saffron exhibits bimodal peaks in April and October $\left(3.4–3.6\ast {10}^6{\mathrm{m}}^3\right)$. Real requirements fluctuate significantly, whereas optimized means provide smoother estimates with reduced extremes. Zone 2 consistently shows slightly higher water requirements than Zone 1 due to local environmental variations. In Zone 1, the shaded blue region represents the range of variability in optimized predictions, which is narrower during the mid-season (June–August) but widens in the early (March) and late (October) periods. Conversely, in Zone 2, variability is more pronounced during April and October, highlighting the challenges associated with these transitional months. Overall, watermelon exhibits a consistent increase in water requirements, reaching its peak during the mid-season (June–August), which corresponds to its heightened water demand during the peak growth phase. In contrast, saffron displays a bimodal pattern, with water demand peaking in spring (April) and autumn (October), reflecting its distinct growth cycle characteristics.

4.3. Analysis of Water Footprint with Regard to Economic Benefits

Figure 9 illustrates the Pareto efficiency analysis, showcasing the trade-offs between three objectives ($U^1$, $U^2$, and $U^3$) for watermelon and saffron. For $U^1$, watermelon shows variability at $\left(0.3–0.8\right)$, while saffron remains stable at $\left(0.2–0.5\right)$, highlighting its balanced water efficiency relative to economic benefit. In terms of $U^2$, watermelon points are distributed across higher values ($0.6–0.9$), reflecting substantial discrepancies between water requirements and supply, highlighting its dependency on high water volumes. Conversely, saffron points are concentrated in the lower $U^2$ range ($0.2–0.4$), demonstrating its efficient water use. For $U^3,$ saffron achieves values primarily within $\left(0.2–0.4\right)$, whereas watermelon exhibits a broader range of $\left(0.4–0.7\right)$, underscoring saffron’s suitability for water-constrained systems. This analysis underscores saffron’s distinct advantage in water conservation strategies, whereas watermelon, despite its economic returns, places a significantly higher strain on water resources.

Additionally,

Table 4. Comprehensive analysis of optimal water distribution across two crops.

		Watermelon	Saffron	${\mathit{U}}^1$	${\mathit{U}}^2$	${\mathit{U}}^3$
Zone 1	$X_i\left({10}^6{m}^3\right)$	34.61	19.87	0.716/0.494	0.802/0.351	0.318/0.495
	$E \left({10}^{9}IRR\right)$ Harvesting ($kg/ha)$	30.09 6792	34.88 6
Zone 2	$X_i\left({10}^6{\mathrm{m}}^3\right)$	41.07	28.65	0.483/0.359	0.743/0.398	0.508/0.327
	$E \left({10}^{9}IRR\right)$ Harvesting ($kg/ha)$	32.40 7104	39.71 8

provides a detailed comparison of water allocation and associated outcomes for watermelon and saffron across two zones under a seasonal cropping pattern time frame. Based on the results, water allocation for watermelon in Zone 1 was recorded at $\left(34.61\ast {10}^6{\mathrm{m}}^3\right)$, increasing to $(47.07\ast {10}^6{\mathrm{m}}^3)$ in Zone 2. In comparison, saffron received significantly lower water volumes of $\left(19.87\ast {10}^6 {\mathrm{m}}^3\right)$ in Zone 1 and $\left(28.65\ast {10}^6 {\mathrm{m}}^3\right)$ in Zone 2. Moreover, saffron achieved the highest average net profit across the two areas, with an average value of $\left(33\ast {10}^{9}IRR\right)$, whereas watermelon’s average net profit was slightly lower, remaining below $\left(31\ast {10}^{9}IRR\right)$. In terms of $U^i,$ saffron has nearly reached the optimal values across all three objective functions, averaging ($0.404$).

In summary, saffron demonstrated superior water efficiency, economic returns, and lower objective function values across zones, making it a more sustainable choice for water-scarce regions. Watermelon, while achieving higher productivity, imposed a substantial strain on water resources. This analysis reinforced saffron’s suitability for regions prioritizing water conservation and economic benefits.

5. Discussion

The findings of this study emphasize the critical role of water allocation strategies in achieving sustainable agricultural productivity, particularly when balancing water-intensive and water-efficient crops. The observed water footprint of saffron ($19.87 to 28.65\ast {10}^6{\mathrm{m}}^3$) was 42–46% lower than that of watermelon ($34.61 to 47.07\ast {10}^6{\mathrm{m}}^3$), while achieving a higher average net profit (33 $\ast {10}^{9}IRR)$ relative to watermelon (31 $\ast {10}^9$ $IRR$). Watermelon, a summer crop, showed peak irrigation demand in July, corresponding to its fruit bulking stage and the region’s highest temperatures, which drive evapotranspiration. Saffron, a cool-season crop, displayed twin demand peaks in April and October. The October peak aligns with its flowering period post-sprouting, while the April peak reflects post-dormancy vegetative growth after winter dormancy, both requiring supplemental irrigation due to limited rainfall. These patterns underscore the importance of synchronizing agricultural practices with regional hydro-climatic dynamics to enhance water-use efficiency. This aligns with prior empirical findings by Sepaskhah and Kamgar-Haghighi [30], who highlighted saffron’s lower water demand in semi-arid region of Iran, but the present study extends this by investigating dynamic trade-offs using an integrated IoT-ML-optimization framework. Watermelon’s high irrigation requirement corroborates Chapagain and Hoekstra [31], but the proposed model further reveals that under water-stressed scenarios, its economic viability deteriorates significantly, as evidenced by a disparity score ($U^2$ > 0.7) with only a 10% supply reduction.

From a methodological perspective, the hybrid ML model $(KNN, GBT, \mathrm{and} LSTM$) achieved high predictive performance ($\rho =0.91–0.96$), outperforming similar IoT-ML systems such as those by Vianny et al. [7], which lacked economic integration, and Alibabaei et al. [32], which did not resolve spatial heterogeneity. The innovation lies in the dynamic model selection mechanism, where $LSTM$ excelled in temporal prediction, while $KNN$ improved spatial accuracy, thereby reducing volume errors by 8–10% compared to individual models. The Pareto-optimal outcomes (saffron’s average objective function value of 0.404) support strategies like water credit trading systems and dynamic subsidy allocation. Consistent with Dalin et al. [33], this study highlights that high-value, low-water crops like saffron can simultaneously improve farmer income and reduce basin-wide water stress. Unlike top-down conservation models [18], the proposed framework enables decentralized, adaptive resource allocation using real-time feedback.

Following the preceding analysis, implementing some enhancement approaches must be taken into account:

5.1. Approach 1: Establish a Water Credit Trading System for Farmers

Introduce a market-based water credit system where farmers earn credits for saving water or adopting water-efficient practices, such as switching to crops like saffron. These credits can be traded within a community or with industries in need of water offsets. This system incentivizes water conservation while creating a financial mechanism for equitable water distribution. Policies for monitoring, validation, and enforcement are critical to the success of this innovative approach.

5.2. Approach 2: Create Digital Water Conservation Platforms with Seasonal Water Balancing

Developing dynamic crop rotation plans that alternate water-intensive crops like watermelon with low-water crops such as saffron or chickpeas can significantly strengthen water conservation efforts. These plans should account for seasonal rainfall patterns and groundwater resources to balance annual water consumption. Additionally, equipping farmers with user-friendly tools to implement these strategies can improve adoption rates and alleviate water stress. Such tools can forecast drought risks, identify water scarcity trends, and recommend optimal sowing times and irrigation schedules. By integrating predictive analytics into agricultural decision-making, farmers can proactively address water challenges, minimize waste, and ensure crop success.

5.3. Approach 3: Promote Vertical Farming for High-Value, Low-Water Crops

Encourage the adoption of vertical farming for crops like saffron in water-scarce areas. This method significantly reduces water usage through hydroponic or aeroponic systems while ensuring higher yields per unit area. Government subsidies or low-interest loans can make the initial investment affordable for small-scale farmers. Demonstrating the economic and water-saving benefits of vertical farming can serve as a compelling case for broader adoption.

Although this study is based on the Neyriz Basin in Fars Province, which is defined by arid climate conditions, limited precipitation, and high sensitivity to drought, the underlying framework is structured to be applicable to water-scarce regions globally. The transferability of the results requires local calibration of input parameters such as soil properties, crop coefficients, and weather patterns. This approach allows irrigation prediction and optimization results to be aligned with the specific environmental and agronomic characteristics of each target region. Accordingly, the proposed methodology holds broad potential for application in other water-scarce areas, provided that regional datasets are integrated appropriately. In general, several limitations must be acknowledged: (1) Environmental factors such as soil type, rainfall variability, and socio-economic characteristics may vary significantly across regions, potentially affecting model transferability. (2) While the present study focused on watermelon and saffron due to their contrasting water demands and economic profiles, the proposed framework is designed to be adaptable to other crop types. For instance, water-intensive crops such as rice or water-efficient cereals like wheat can be incorporated into the same decision structure, thereby providing their crop-specific phenology and local irrigation parameters in the training data. The framework’s modular nature allows for the integration of new crops by updating environmental, economic, and irrigation input datasets. (3) The computational complexity of the hybrid ML model (KNN-GBT-LSTM) may pose challenges for real-time implementation in resource-constrained regions. Furthermore, the economic profitability calculations are based on the assumption of stable market conditions, neglecting potential fluctuations in crop prices or external shocks.

6. Conclusions

This study provides a comprehensive evaluation of water footprint optimization by employing an innovative IoT-ML integrated framework. By focusing on watermelon, a water-intensive crop, and saffron, a water-efficient and economically valuable crop, this research highlights the trade-offs and synergies between water conservation and agricultural productivity in water-scarce regions such as Fars Province, Iran.

According to the results, saffron performance across multiple objective functions confirmed its optimality, with an average objective function value of $0.404$, reflecting balanced contributions to water conservation, economic profitability, and agricultural sustainability. In contrast, watermelon, while delivering higher productivity, exerted significant pressure on water resources, rendering it less favorable for long-term sustainability. These findings emphasize the potential of saffron cultivation as a strategic crop choice to address water scarcity challenges in arid and semi-arid regions. By combining water-efficient agricultural practices with advanced IoT-ML technologies, policymakers and stakeholders can devise strategies that promote sustainable water management while ensuring economic resilience. This study reinforces the need for prioritizing value-added, water-efficient crops like saffron in regional agricultural planning to achieve long-term sustainability and resilience in water-scarce ecosystems.

To advance research in this field, optimal frameworks could be developed that incorporate ensemble and unsupervised learning algorithms to improve prediction accuracy and decision-making. These frameworks could introduce deep prediction techniques to address issues related to signal noise and other external interferences.