Morphophysiological Responses of Lettuce to Irrigation Depths and Wastewater Sources with a Machine Learning Approach

Abstract

The increasing pressure on water resources has stimulated the use of treated wastewater in agricultural irrigation, although its effects on plant development remain uncertain. This study evaluated the effects of wastewater treatments and irrigation depths on the morphophysiological development of lettuce (Lactuca sativa L.). A split-plot experiment was conducted with crop cycles in the main plots and a factorial arrangement in the subplots, consisting of five water sources and five irrigation depths (50% to 150% ETc), with three replications. Seven variables were evaluated, including growth traits and water productivity. Irrigation depth significantly affected all variables (p ≤ 0.01) and was the main driver of vegetative growth, increasing shoot fresh mass, stem diameter, and plant height. In contrast, water sources showed smaller effects. Water productivity decreased with increasing irrigation depth and showed weak correlations with other variables (r ≤ 0.468). Machine learning models achieved moderate accuracy for irrigation depth prediction (≈55%), with confusion among adjacent classes, indicating detection of a gradient rather than precise classification. Prediction of water sources was low (<30%), confirming limited morphological differentiation. Plant height and stem diameter were the most informative variables. These results indicate that irrigation management has a stronger influence on lettuce growth than water source.

1. Introduction

Water scarcity is one of the major challenges for agriculture in the 21st century, stimulating the search for alternative water sources and more efficient use strategies. At a global scale, water demand is estimated to reach approximately 160% of the currently available volume by 2030, intensifying pressure on conventional water resources [1]. As about 70% of global freshwater consumption is attributed to agriculture, this scenario highlights the sector’s vulnerability to water scarcity and reinforces the need for alternative sources, such as treated wastewater [2,3,4]. In this context, the use of treated wastewater in agricultural irrigation emerges as a promising alternative, combining the appropriate disposal of effluents with the reduction in demand for conventional water resources [5,6]. However, the implementation of this practice requires a detailed understanding of its effects on plant development, considering both the chemical characteristics of the water and the conditions of application, particularly when multiple growth variables are evaluated simultaneously.

Previous studies have investigated the effects of wastewater on agricultural crops, often focusing on isolated growth parameters or specific aspects of water quality [7,8,9]. However, few studies have adopted an integrated approach that simultaneously considers multiple morphophysiological variables and their complex interactions. This limitation becomes more relevant when irrigation depth and water quality are evaluated together, since both directly influence physiological processes such as cell expansion, carbon assimilation, and water use efficiency. Thus, plant responses to environmental stresses rarely manifest in isolated parameters, but rather in integrated patterns of development [10,11].

Lettuce (Lactuca sativa L.) represents an appropriate model for such investigations due to its economic importance, short growth cycle, and sensitivity to cultivation conditions [12]. Additionally, its fresh consumption requires special care regarding irrigation water quality, making it an emblematic crop for studies involving alternative water sources [13]. Moreover, its rapid physiological response to variations in water availability and quality enables the detection of subtle changes in growth patterns, making it suitable for evaluating management strategies. Its wide cultivation under protected and open-field conditions further enhances the applicability and relevance of experimental findings to different production systems [6,9,10,11].

Recent studies have documented both agronomic benefits, such as increases in lettuce productivity that may exceed 80% with the use of treated wastewater, and associated challenges, particularly salt accumulation and the possible presence of microbiological contaminants in soil and plants [14,15]. These findings highlight the complexity of water reuse and the need for integrated approaches considering multiple dimensions of plant responses and the soil–water–plant ecosystem. In this context, understanding the balance between productivity gains and potential environmental risks is essential for the sustainable adoption of wastewater reuse in agricultural systems.

In recent years, advances in data analysis methods have expanded the capacity to interpret complex agricultural systems in which multiple biophysical variables interact simultaneously. Multivariate analysis techniques, correlation analysis, and data mining methods have been used to identify hidden patterns in agronomic datasets, allowing a better understanding of the relationships between plant morphological characteristics and management factors [16,17]. In particular, approaches based on correlation networks and attribute analysis have proven useful for revealing interdependence among physiological and productive variables [18,19]. These methods also help identify key indicators of crop performance under different environmental conditions. Together, they provide a systemic view of plant development, overcoming the limitations of traditional univariate analyses and supporting more robust interpretations of plant growth mechanisms under different management strategies.

The increasing availability of agricultural data, driven by sensors and monitoring technologies, has favored the use of machine learning (ML) techniques to explore complex patterns in agricultural production systems [20,21]. Unlike traditional statistical approaches, these methods allow simultaneous analysis of multiple variables and identification of nonlinear relationships between plant characteristics and cultivation conditions. In this context, classification and clustering algorithms have been used to identify phenotypic patterns related to water management and environmental conditions. In the present study, machine learning algorithms, including Naive Bayes, Random Forest, J48, and SMO, were employed to explore morphophysiological patterns of lettuce subjected to different combinations of irrigation depths and water sources.

Although previous studies have explored the effects of irrigation management and wastewater use on crop performance, most have focused on isolated variables or have relied on conventional statistical approaches, limiting the understanding of integrated plant responses. Additionally, studies applying machine learning in agriculture have often emphasized irrigation optimization, yield prediction, or decision support systems, with limited attention to the identification of morphophysiological response patterns under combined management factors such as irrigation depth and water quality [22,23]. Therefore, studies integrating controlled experimental data with advanced analytical approaches capable of capturing complex and nonlinear interactions among multiple plant traits are still needed. In this context, this study combines controlled experimentation with advanced analytical tools, including machine learning and correlation network analysis. These approaches were used to identify plant development patterns and evaluate the predictive capacity of morphophysiological variables under different irrigation and water source conditions.

The study is based on the hypothesis that lettuce responses are predominantly determined by the quantity and quality of irrigation water. It is further assumed that growth variables exhibit interdependence patterns and that part of this variability can be described using multivariate and machine learning approaches. In this context, this study proposes an integrated approach that combines controlled experimentation with advanced analytical tools, including machine learning and correlation network analysis. The objective was to characterize the morphophysiological response patterns of lettuce under different irrigation depths and water sources, evaluating the predictive capacity of agronomic variables and identifying the most informative attributes associated with the cultivation conditions.

2. Materials and Methods

2.1. Experimental Conditions

To conduct the experiment, lettuce (Lactuca sativa L.) seeds of the Saia Véia variety were used. This cultivar belongs to the group of curly leaf lettuces, presenting a rosette growth habit, green serrated leaves, and the absence of compact head formation [24]. In addition, this cultivar has a short growth cycle and is frequently used in agronomic studies due to its rapid response to management conditions and environmental variations.

Two cultivation cycles were carried out. In the first cycle, transplanting occurred on 25 June 2021, and harvesting was performed on 3 August 2021, totaling a cultivation period of 40 days. In the second cycle, transplanting occurred on 9 August 2021, and harvesting was performed on 15 September 2021, totaling a cultivation period of 38 days.

The experiment was conducted under protected environment conditions in a structure measuring 9.0 m in length, 6.5 m in width, and 3.0 m in height. The greenhouse was covered with transparent low-density polyethylene, 0.15 mm thick (150 µm), to reduce the impacts caused by storms and rainfall. The sides of the greenhouse were covered with protective mesh (anti-aphid screen). The structure is located at the Department of Agricultural Engineering (DEA) of the Federal University of Sergipe, in the municipality of São Cristóvão, Sergipe, Brazil. The geographical coordinates of the site are 10°55′46″ S latitude, 37°06′13″ W longitude, and an altitude of 8 m above sea level.

The climate of the region where the study was conducted is classified as As’, tropical climate according to the Köppen climate classification, characterized by winter–autumn rainfall [25]. The average daily air temperature and relative humidity were determined from measurements recorded throughout each day. Minimum and maximum values corresponded to the extremes observed during the same interval. Daily solar radiation was obtained by integrating records collected every 5 min over a 24 h period. Based on these variables, the daily reference evapotranspiration (ETo) was estimated using the Penman–Monteith equation [26] and accumulated for each cultivation cycle. The behavior of the meteorological variables is presented in Figure 1.

During the experimental cycles, the average daily air temperature, relative humidity, and solar radiation were 24.7 and 25.2 °C, 63.3 and 60.5%, and 12.3 and 17.3 MJ m⁻² day⁻¹ for cycles 1 and 2, respectively. At the end of each cycle, the accumulated reference evapotranspiration was 69.0 mm in cycle 1 and 99.6 mm in cycle 2.

Before transplanting, physical and chemical soil analyses were performed. The results indicated a field capacity of 8.38% and a permanent wilting point of 2.25%, expressed as gravimetric moisture on a dry basis. The soil texture was classified as sandy loam. Due to the observed acidity, liming was performed at a dose equivalent to 320 kg ha⁻¹, followed by soil saturation and an incubation period of approximately 90 days. Based on the chemical analysis, basal fertilization was applied at transplanting with 30 kg ha⁻¹ of N, 60 kg ha⁻¹ of P₂O₅, and 30 kg ha⁻¹ of K₂O, using urea, single superphosphate, and potassium chloride as sources. Topdressing fertilization was applied only in the treatments for which this practice was specified in the experimental design.

Each experimental unit consisted of three plants per replicate, all of which were used for the final evaluations. The cultivation system consisted of 15 L pots filled with soil. Irrigation was performed manually using watering cans, applying the corresponding irrigation depths according to the experimental treatments.

2.2. Treatments

The experiment was conducted in a split-plot design with three replications. The different crop cycles were evaluated in the main plots. In the subplots, a 5 × 5 factorial scheme was applied, consisting of water sources (WS) and irrigation depths (ID). Each experimental unit was composed of three plants, whose individual measurements were used to obtain a mean value, which was subsequently considered for the statistical analysis.

The tap water was supplied by the local water utility, and the treated wastewater originated from the Rosa Elze Wastewater Treatment Plant (WTP) located in São Cristóvão, Sergipe, Brazil. This WTP operates through a system of stabilization ponds arranged in series (two facultative and three maturation ponds), representing a secondary treatment process. The treated wastewater was collected from the same outlet point throughout the experimental period to ensure uniformity in its origin and treatment conditions. Although natural variations in chemical composition may occur over time, efforts were made to maintain consistency in the water source throughout the experiment. Part of the treated wastewater was subsequently subjected to a post-treatment filtration using a biochar-based system.

The biochar was produced from orange bagasse (Pêra variety) under controlled conditions involving drying, grinding, and carbonization in a muffle furnace at 550 °C for 60 min. This process resulted in a homogeneous material with suitable properties for filtration. This material was subsequently used as the filtering medium in the post-treatment system. The physicochemical and microbiological characteristics of the three water sources are presented in

Table 1. Physicochemical and microbiological characterization of the three water sources used in the experiment: tap water (TW), treated wastewater (TWW), and biochar-filtered treated wastewater (TWW + B), including electrical conductivity (EC), sodium adsorption ratio (SAR), dissolved oxygen (DO), total iron (Fe), total phosphorus (TP), and total coliforms (TC).

Parameters	TW	TWW	TWW + B
Electrical conductivity (EC, dS m−1)	0.2157 ± 0.0140	0.9750 ± 0.1791	1.1438 ± 0.0767
Sodium adsorption ratio (SAR)	2.9375 ± 0.8285	9.8750 ± 3.7908	9.6250 ± 3.8190
Dissolved oxygen (DO, mg L−1)	6.5895 ± 0.4320	6.4211 ± 0.9252	7.5526 ± 0.8630
Total iron (Fe, mg L−1)	0.0984 ± 0.0642	0.0975 ± 0.0249	0.0713 ± 0.0196
Total phosphorus (TP, mg L−1)	0.0748 ± 0.0230	3.0875 ± 0.5647	4.2250 ± 0.8245
Total coliforms (TC, CFU mL−1)	0.0 ± 0.0	2421.4 ± 1735.8	748.6 ± 367.9

.

The water source treatments were defined as follows: WS1) tap water with topdressing fertilization; WS2) treated wastewater from the treatment plant; WS3) treated wastewater filtered through biochar; WS4) treated wastewater with topdressing fertilization; WS5) treated wastewater with biochar filtration and topdressing fertilization. The water sources were used as the basis for the treatments as follows: tap water (TW) for WS1; treated wastewater (TWW) for WS2 and WS4; and biochar-filtered treated wastewater (TWW + B) for WS3 and WS5. Topdressing fertilization was applied only in treatments WS1, WS4, and WS5. Treatments WS2 and WS3 received exclusively the nutrients present in the treated wastewater or biochar-filtered wastewater.

The irrigation depths corresponded to percentages of crop evapotranspiration (ETc): ID₅₀) 50% ETc; ID₇₅) 75% ETc; ID₁₀₀) 100% ETc; ID₁₂₅) 125% ETc; ID₁₅₀) 150% ETc. Irrigation was performed daily. The irrigation depth corresponding to 100% ETc was determined based on soil moisture at the time of irrigation, calculating the volume of water required to restore soil moisture to field capacity. The other irrigation depths (50, 75, 125, and 150% ETc) were obtained as proportions of this standard irrigation depth.

It is important to clarify that, although the treatments are expressed as percentages of ETc, the 100% level in this study represents an operational reference based on soil water depletion rather than a classical climatic estimation of crop evapotranspiration. In practice, it corresponds to the amount of water required to restore soil moisture to field capacity at the time of irrigation. The remaining treatments were defined proportionally from this reference condition.

2.3. Evaluated Characteristics

Seven morphophysiological variables were evaluated: NL—number of leaves, PH—plant height (cm), SL—stem length (cm), SD—stem diameter (cm), SFM—shoot fresh mass (g), SDM—shoot dry mass (g), and WP—water productivity (g L⁻¹). Evaluations were performed at the end of each cultivation cycle, during harvest. Initially, the plants were cut at the soil surface to separate the shoot and root systems. Shoot fresh mass was determined immediately after harvest using a precision digital scale. Subsequently, the total number of fully developed leaves per plant was manually counted.

Plant height was measured using a graduated ruler, considering the distance between the soil surface and the highest point of the shoot. Stem length was determined using a millimeter ruler after removing the leaves, measuring the stem extension from the base to the apex. Stem diameter was measured at the median region using a digital caliper.

To determine shoot dry mass, samples were placed in paper bags and dried in a forced-air circulation oven at 65 °C for 72 h until constant mass was reached. They were then weighed on an analytical balance. Water productivity was calculated as the ratio between shoot fresh mass and the total irrigation water depth effectively applied to the experimental plots (net irrigation). This variable corresponded to the water volume supplied to the plants during the cultivation cycle and was expressed in g L⁻¹.

2.4. Data Analysis

The obtained data were subjected to preliminary analyses of normality and homogeneity of residuals using the Shapiro–Wilk and Bartlett tests, respectively. Analysis of variance (ANOVA) was performed using the F-test at significance levels of 0.01 and 0.05. For both significant and non-significant interactions among factors, mean comparisons were performed using Tukey’s post hoc test at a 5% significance level. All statistical analyses were carried out using R software (version 4.4.2).

The results were also subjected to correlation analysis among variables and the application of machine learning techniques using the Weka platform (version 3.8). Specifically, classification algorithms (Naive Bayes, Random Forest, J48, and SMO) were applied with 10-fold cross-validation, attribute importance analysis (InfoGainAttributeEval), and clustering using the Expectation–Maximization algorithm. The data were prepared in ARFF format, generating two primary datasets: one for irrigation depth prediction (excluding treatment) and another for treatment prediction (excluding irrigation depth). A third dataset combining treatment and depth into a single 25-class target was also generated. However, preliminary analyses showed no predictive signal above random chance; therefore, its results are not reported.

For the comparative statistical analysis of the algorithms and the generation of performance graphs in the Multivariate analysis and predictive modeling section, Python (version 3.9) was used in the Google Colab environment. In this stage, statistical analyses (including the Tukey test) were performed using the statsmodels (version 0.13.2) and SciPy (version 1.7.3) libraries. Data visualization was performed using Matplotlib (version 3.5.1) and Seaborn (version 0.11.2). Bar charts with significance letters were generated in high resolution (600 DPI) and exported in TIFF format, suitable for scientific publication. This procedure enabled the statistical analyses and the generation of the graphs used in this part of the results.

The dataset consisted of 150 observations (5 water sources × 5 irrigation depths × 3 replications × 2 cycles). Data from both cycles were combined into a single dataset, as preliminary analysis showed no significant cycle × treatment interaction for most variables. Before modeling, numerical attributes were standardized using z-score normalization (mean = 0, standard deviation = 1) to avoid scale bias, particularly in algorithms such as SMO. Class balancing was not applied because the experimental design ensured equal representation of each class (n = 30 per irrigation depth class; n = 30 per treatment class). Replications were kept independent and allocated to folds following the stratified 10-fold cross-validation procedure. Thus, each fold preserved the original class distribution and avoided mixing measurements from the same experimental unit across training and validation sets. To prevent information leakage, preprocessing parameters were computed only on the training folds and then applied to the validation fold. The choice of algorithms, Naive Bayes (simple probabilistic baseline), Random Forest (ensemble robust to small datasets), J48 (interpretable tree), and SMO (nonlinear SVM), was motivated by their complementary inductive biases and their common use in agronomic studies with moderate sample sizes [16,20].

3. Results

3.1. Agronomic Characteristics

Table 2. Mean squares and F-test (ANOVA) significance for the variables number of leaves, plant height, stem length, stem diameter, shoot fresh mass, shoot dry mass, and water productivity of lettuce grown under two cultivation cycles (C) and subjected to different water sources (WS) and irrigation depths (ID).

Factor	Mean Squares
	C	WS	ID	C × WS	C × ID	WS × ID	C × WS × ID
Number of leaves (ud pl−1)	4.59 × 101 ns	2.27 × 102 **	1.03 × 103 **	3.11 × 101 ns	1.24 × 102 *	6.61 × 101 ns	5.24 × 101 ns
Plant height (cm)	2.76 × 102 **	5.12 × 100 **	6.41 × 101 **	8.95 × 100 **	1.41 × 101 **	4.73 × 100 *	1.61 × 100 ns
Stem length (cm)	2.69 × 102 **	1.21 × 101 **	9.13 × 101 **	1.30 × 101 **	2.00 × 101 **	8.04 × 100 *	4.38 × 100 ns
Stem diameter (cm)	5.87 × 100 **	1.85 × 100 **	7.17 × 100 **	1.73 × 10−1 *	4.22 × 10−1 **	1.39 × 10−1 ns	8.82 × 10−2 ns
Shoot fresh mass (g pl−1)	1.03 × 106 **	6.15 × 104 **	2.61 × 105 **	2.13 × 104 **	5.17 × 104 **	1.09 × 105 ns	7.80 × 103 ns
Shoot dry mass (g pl−1)	2.75 × 102 ns	4.39 × 101 **	2.79 × 102 **	1.34 × 101 ns	3.26 × 101 **	1.88 × 101 **	1.05 × 101 ns
Water productivity (g L−1)	1.14 × 104 **	3.13 × 102 **	2.12 × 102 **	3.59 × 102 **	4.24 × 101 ns	9.29 × 101 ns	4.33 × 101 ns

C × WS × ID: triple interaction among cultivation cycles, water source treatments, and irrigation depths; * and **: significant at 5% and 1% probability, respectively, by the F test; ns: not significant.

presents the mean squares and F-test (ANOVA) significance levels for the variables number of leaves, plant height, stem length, stem diameter, shoot fresh mass, shoot dry mass, and water productivity of lettuce, evaluated as a function of cultivation cycles (C), water sources (WS), and irrigation depths (ID). In general, irrigation depths and water sources significantly affected most analyzed variables, highlighting the strong influence of water management and water quality on lettuce development. Irrigation depths had a highly significant effect (p ≤ 0.01) on all variables, indicating that the level of water replacement was a determining factor for lettuce growth and productivity.

The cultivation cycle showed a significant effect on most of the evaluated variables, except for number of leaves and shoot dry mass. Irrigation depths and water sources also significantly affected most variables. The interaction between water sources and irrigation depths (WS × ID) was significant for some variables, as was the interaction between cycle and irrigation depths (C × ID). However, the triple interaction (C × WS × ID) was not significant for any analyzed variable.

Table 3. Mean values followed by their respective standard deviations for number of leaves, plant height, stem length, stem diameter, shoot fresh mass, shoot dry mass, and water productivity of lettuce subjected to different irrigation depths (ID, % ETc) and water sources (WS). WS1—tap water + topdressing; WS2—treated wastewater; WS3—treated wastewater + biochar; WS4—treated wastewater + topdressing; WS5—treated wastewater + biochar + topdressing.

Factor	ID	WS1		WS2		WS3		WS4		WS5
Number of leaves (ud pl−1)	50	37.917 ± 3.264	Ab	42.333 ± 4.727	Aa	44.333 ± 3.625	Ab	33.750 ± 4.250	Ac	40.083 ± 6.643	Ac
	75	42.167 ± 11.955	Aab	49.833 ± 10.869	Aa	50.833 ± 3.517	Aab	44.333 ± 3.329	Abc	43.167 ± 4.410	Ac
	100	48.333 ± 8.083	ABab	48.167 ± 4.569	ABa	58.333 ± 4.517	Aa	49.083 ± 4.264	ABab	47.000 ± 3.073	Bbc
	125	49.000 ± 4.323	Ba	49.833 ± 6.566	ABa	56.333 ± 5.966	ABa	56.500 ± 2.971	ABa	59.667 ± 9.179	Aa
	150	52.833 ± 2.758	Aa	49.833 ± 6.838	Aa	55.667 ± 3.649	Aa	50.667 ± 6.215	Aab	54.833 ± 10.355	Aab
Plant height (cm)	50	13.363 ± 0.864	Bb	15.300 ± 1.641	ABb	15.892 ± 0.710	Aa	13.850 ± 0.400	ABd	15.388 ± 0.689	ABb
	75	16.250 ± 2.454	ABa	16.725 ± 0.666	ABab	18.075 ± 0.923	Aa	15.692 ± 0.500	Bcd	16.175 ± 1.093	ABb
	100	18.283 ± 2.233	Aa	16.867 ± 1.395	Aab	18.317 ± 0.851	Aa	16.808 ± 0.967	Abc	17.475 ± 0.988	Aab
	125	16.946 ± 1.254	Ba	18.283 ± 0.847	ABa	18.042 ± 1.402	ABa	19.842 ± 1.404	Aa	18.775 ± 1.204	ABa
	150	17.921 ± 1.529	Aa	17.925 ± 0.783	Aa	18.225 ± 0.216	Aa	18.958 ± 1.117	Aab	17.375 ± 2.599	Aab
Stem length (cm)	50	8.378 ± 2.265	ABb	10.832 ± 1.952	Aa	11.088 ± 1.688	Ab	7.398 ± 0.738	Bc	9.768 ± 0.915	ABb
	75	11.110 ± 2.093	Aab	11.033 ± 1.416	Aa	12.802 ± 1.150	Aab	10.892 ± 1.663	Ab	9.745 ± 2.074	Ab
	100	14.253 ± 1.343	Aa	11.127 ± 2.196	Aa	12.950 ± 1.200	Aab	12.137 ± 0.681	Aab	12.683 ± 1.287	Aab
	125	12.794 ± 1.749	Aa	12.817 ± 1.314	Aa	15.260 ± 3.504	Aa	14.447 ± 1.013	Aa	14.520 ± 2.132	Aa
	150	12.414 ± 1.985	Aa	12.627 ± 1.798	Aa	13.670 ± 1.725	Aab	14.287 ± 0.978	Aa	11.588 ± 3.236	Aab
Stem diameter (cm)	50	1.7867 ± 0.2127	ABb	2.2450 ± 0.1098	Ab	2.2433 ± 0.1206	Ac	1.7125 ± 0.1575	Bc	1.6458 ± 0.1782	Bc
	75	2.1517 ± 0.3005	Cb	2.6500 ± 0.3856	ABab	2.9700 ± 0.1180	Ab	2.2583 ± 0.1387	BCb	2.1567 ± 0.1219	Cbc
	100	2.7617 ± 0.4049	Ba	2.9383 ± 0.3521	ABa	3.4017 ± 0.2872	Aab	2.7842 ± 0.3325	Ba	2.5483 ± 0.1876	Bab
	125	2.8208 ± 0.3233	Ba	2.9333 ± 0.4317	Ba	3.4950 ± 0.3396	Aa	3.1417 ± 0.1426	ABa	2.9150 ± 0.1759	Ba
	150	2.9442 ± 0.1835	Aa	3.1217 ± 0.3104	Aa	3.1667 ± 0.2838	Aab	3.0917 ± 0.1630	Aa	2.9617 ± 0.7188	Aa
Shoot fresh mass (g pl−1)	50	142.17 ± 32.69	Ab	235.33 ± 115.32	Ab	240.33 ± 47.26	Ab	174.83 ± 59.75	Ab	166.17 ± 19.06	Ab
	75	245.50 ± 82.33	Bab	286.17 ± 45.69	ABab	395.67 ± 25.29	Aa	216.00 ± 19.22	Bb	230.67 ± 34.47	Bb
	100	321.50 ± 102.07	Ba	342.33 ± 112.07	Bab	483.00 ± 76.70	Aa	345.08 ± 87.55	Ba	375.67 ± 18.33	ABa
	125	314.33 ± 13.12	Ba	375.00 ± 77.84	ABa	458.17 ± 76.18	Aa	456.25 ± 54.87	Aa	421.17 ± 97.49	ABa
	150	340.75 ± 65.50	Ba	375.33 ± 59.47	ABa	418.50 ± 75.11	ABa	465.00 ± 82.10	Aa	406.33 ± 163.00	ABa
Shoot dry mass (g pl−1)	50	10.047 ± 2.123	ABb	13.935 ± 2.122	Aa	14.368 ± 1.263	Ab	9.078 ± 1.248	Bc	12.563 ± 3.815	ABb
	75	14.362 ± 3.617	ABab	16.343 ± 3.623	ABa	17.843 ± 2.146	Aab	14.155 ± 0.902	ABb	12.233 ± 2.749	Bb
	100	16.065 ± 3.316	ABa	14.539 ± 3.020	ABa	18.714 ± 1.409	Aab	14.289 ± 1.881	Bb	16.580 ± 0.778	ABab
	125	17.196 ± 3.553	Ba	18.367 ± 1.766	ABa	20.932 ± 4.480	ABa	21.836 ± 2.344	Aa	20.478 ± 2.499	ABa
	150	17.345 ± 1.155	Aa	16.432 ± 2.443	Aa	19.520 ± 2.385	Aa	20.587 ± 1.618	Aa	18.785 ± 6.246	Aa
Water productivity (g L−1)	50	21.132 ± 4.870	Ba	34.408 ± 16.596	Aa	31.984 ± 6.200	ABab	29.169 ± 9.994	ABa	26.761 ± 3.175	ABa
	75	26.611 ± 8.748	Aa	30.744 ± 4.915	Aa	37.836 ± 2.434	Aa	26.592 ± 2.408	Aa	27.360 ± 4.251	Aa
	100	27.590 ± 8.637	Aa	29.686 ± 8.825	Aa	36.269 ± 5.729	Aa	33.494 ± 8.560	Aa	35.388 ± 1.698	Aa
	125	22.299 ± 0.927	Ba	27.096 ± 5.133	ABa	28.350 ± 4.622	ABab	36.693 ± 4.346	Aa	32.744 ± 7.488	ABa
	150	20.452 ± 3.900	Ba	23.807 ± 3.631	ABa	22.166 ± 3.887	ABb	32.008 ± 5.613	Aa	27.098 ± 10.673	ABa

Means followed by the same uppercase letters in the column and lowercase letters in the row do not differ from each other by Tukey’s test (p < 0.05).

presents the mean values (±standard deviation) for number of leaves, plant height, stem length, stem diameter, shoot fresh mass, shoot dry mass, and water productivity of lettuce as a function of the interaction between water sources and irrigation depths. The unfolding was performed only for this interaction (WS × ID), as it represents the main focus of the study. The cycle factor was not detailed at this stage because it is associated with climatic variations. In addition, the second cycle was used to expand the dataset and support subsequent analyses.

In general, Table 3 shows that increasing irrigation depths resulted in higher values for growth variables, especially shoot fresh mass, stem diameter, and plant height. Regarding water sources, treatments with treated wastewater, alone or combined with biochar and/or fertilization, often showed higher shoot fresh and dry mass values. Water productivity presented similar or higher values in some treatments with alternative water sources compared to tap water. In addition, differences among treatments were more evident at intermediate and higher irrigation depths (100–150% ETc). At 50% ETc, these differences were less pronounced.

3.2. Multivariate Analysis and Predictive Modeling

The application of machine learning techniques allowed the evaluation of the predictability of cultivation conditions based on the agronomic characteristics of lettuce. Classification models for predicting irrigation depth achieved an accuracy of 54.67% using the Naive Bayes algorithm, a performance significantly higher than random chance (20%). The confusion matrix revealed a coherent pattern: adjacent irrigation depths were frequently confused, whereas the 50% ETc depth was correctly identified in 73.30% of the cases. In contrast, models for predicting water source showed limited performance, with a maximum accuracy of 30.67% obtained with Random Forest. This result indicates that the evaluated morphological characteristics do not efficiently discriminate among the different water types.

The confusion matrix revealed that most misclassifications occurred between neighboring irrigation depths (e.g., 75% vs. 100% ETc), whereas extreme classes (50% and 150% ETc) were more accurately identified. This pattern suggests that morphological responses to water availability are gradual. The model captures the general trend of increasing vegetative growth with higher irrigation depths but lacks resolution to reliably separate intermediate water regimes.

The attribute importance analysis corroborated these observations. For irrigation depth prediction, plant height stood out as the most informative variable (information gain 0.761 ± 0.113), followed by stem diameter (0.687 ± 0.026) and shoot fresh mass (0.634 ± 0.093). Notably, water productivity showed zero information gain, indicating independence from the other variables. For treatment prediction, all variables exhibited information gain values close to zero. This explains the low performance of the classifiers.

Clustering analyses identified patterns in the agronomic characteristics of lettuce. For irrigation depth, four clusters were identified corresponding to distinct growth patterns, with a clear association between plant size and the applied irrigation depth. For water source, the three identified clusters were mainly segregated by vegetative size, without a consistent association with water type. This result reinforces the idea that the treatments do not induce characteristic morphological patterns.

3.2.1. Correlation Analysis and Network Structure

To complement the predictive analysis and explore interrelationships among variables, Pearson correlation analysis and correlation networks were applied using different thresholds (r ≥ 0.3; r ≥ 0.5; r ≥ 0.7). The correlation matrix (Figure 2) showed positive and significant associations among most of the vegetative growth variables of the lettuce crop.

Shoot fresh mass, stem diameter, plant height, stem length, and number of leaves formed a strongly intercorrelated core, with coefficients frequently exceeding 0.70. In clear contrast, water productivity showed weak to moderate correlations (r ≤ 0.468) with all other variables, with its lowest association observed with shoot dry mass (r = 0.192).

The complete segregation of water productivity in the network with a threshold of 0.70, as well as its low connectivity at lower thresholds, reinforces the result obtained in the machine learning analysis. This finding indicates that this variable operates independently of the others. This network topology suggests that water productivity is not strongly associated with the morphological variables under the tested conditions. However, this pattern should be interpreted as a descriptive property of the current dataset rather than evidence of independent physiological mechanisms. Further experiments measuring physiological traits (e.g., stomatal conductance, transpiration efficiency) would be required to investigate functional dissociation.

3.2.2. Performance of Classification Algorithms for Prediction

The comparative analysis of machine learning algorithms revealed distinct patterns for predicting irrigation depth and water source (Figure 3). For irrigation depth prediction, the Naive Bayes algorithm showed the best performance (55.2% mean accuracy), followed by Random Forest (52.4%), SMO (48.2%), and J48 (46.2%). The Tukey statistical test indicated significant differences among all algorithms, with each belonging to a distinct statistical group (Figure 3A).

In contrast, algorithms for water source prediction showed lower performance and no statistical differences among them. Random Forest achieved the highest mean accuracy (28.9%), followed by J48 (26.9%), Naive Bayes (25.4%), and SMO (22.8%). The Tukey test did not detect statistically significant differences among the algorithms, with all being grouped in the same significance class (Figure 3B).

The superior performance of Naive Bayes may be attributed to the relatively small sample size, limited number of features, and substantial overlap between adjacent irrigation depth classes. Under these conditions, simpler models tend to be more robust and less prone to overfitting than more complex algorithms such as Random Forest, J48, or SMO.

4. Discussion

4.1. Agronomic Responses to Irrigation Depth and Water Source

The morphological responses of lettuce were more strongly influenced by irrigation depth than by water source, indicating that soil water availability was the main determinant of vegetative growth under the evaluated conditions. This behavior reinforces that the development of leafy vegetables depends directly on plant water status, since processes such as maintenance of cell turgor and tissue expansion are highly sensitive to water potential [27,28,29]. In this context, increasing water availability promotes growth, whereas deficit conditions limit leaf expansion and biomass accumulation. Similar responses have been widely reported in leafy vegetables [4,9], including lettuce under different irrigation management strategies [12,30,31].

However, a tendency toward reduced incremental gains was observed at higher irrigation depths, suggesting a stabilization of growth close to optimal water supply conditions. This plateau effect observed at 125% and 150% ETc suggests that the crop approached a physiological and morphological saturation point, where additional water supply no longer translated into proportional biomass gains.

This behavior may be associated with diminishing irrigation returns, in which the first increments in water availability strongly stimulate growth, while subsequent increments progressively reduce efficiency. In addition, the response may be constrained by limitations in leaf area expansion and canopy self-shading. These factors restrict further increases in light interception and biomass accumulation. Another contributing factor may be reduced root-zone aeration under higher water contents, which can limit oxygen availability, root activity, and nutrient uptake efficiency.

The interactions among factors indicate that the crop response is not exclusively additive. The interaction between water sources and irrigation depths indicates that irrigation management efficiency depends on water quality. Recent studies demonstrate that treated wastewater can influence plant growth through nutrient supply and other agronomic effects [32,33]. In addition, its use presents sanitary implications for fresh-consumed leafy vegetables [34] and potential risks related to the transfer of antimicrobial-resistant bacteria and genes to edible crops [35].

On the other hand, the absence of a triple interaction indicates no simultaneous dependence among cycle, water source, and irrigation depth. This suggests that the effects are predominantly explained by individual factors and two-way interactions. In this context, irrigation depth and water source emerge as the main determinants of the crop’s agronomic response.

Although differences among water sources were less pronounced than those observed for irrigation depths, there was a tendency for better performance in treatments with treated wastewater, either alone or combined with biochar and fertilization. This behavior may be associated with the additional nutrient supply promoted by treated wastewater, as reported in recent studies on its agricultural use [14,36].

The relatively weak differentiation among water sources suggests that the treated wastewater had a moderate and nutritionally balanced composition after treatment, reducing contrast with tap water. In addition, biochar filtration may have contributed to partial stabilization of water quality by reducing contaminant excesses while maintaining nutrient availability. However, these changes were insufficient to generate strong agronomic divergence among treatments. Recent studies also indicate that the agronomic effects of treated wastewater are often associated with its nutrient supply and relatively stable composition under controlled conditions [32,33]. Furthermore, topdressing fertilization appears to have played an important role in modulating plant responses, partially offsetting or masking differences among water sources. In several cases, topdressing fertilization was associated with reductions in certain growth indices. This response may indicate nutrient imbalance or interactions between mineral fertilization and nutrients supplied via treated wastewater, which may have altered nutrient uptake dynamics and reduced treatment contrasts. Additionally, greenhouse conditions likely buffered differences among water sources by minimizing environmental stress and reducing plant sensitivity to moderate variations in water quality.

This interpretation is supported by the physicochemical characterization of the water sources (Table 1), which shows higher total phosphorus and electrical conductivity in treated wastewater (TWW and TWW + B) compared to tap water (TW), indicating greater nutrient availability. Meanwhile, similar values of total iron and sodium adsorption ratio, along with moderate variation in dissolved oxygen, suggest partial chemical overlap between TWW and TWW + B. This reduced contrast helps explain the limited discrimination among treatments in univariate analyses and machine learning models.

In addition, the electrical conductivity (EC) and sodium adsorption ratio (SAR), presented in Table 1 of the manuscript, indicate a potential long-term risk of soil salinization and sodification when using treated wastewater (TWW) and biochar-filtered wastewater (TWW + B). Although no immediate negative effects on lettuce growth were observed under the controlled conditions of this study, EC values close to 1.0 dS m⁻¹ and SAR values around 10 suggest that continuous use may gradually affect soil structure, reduce infiltration capacity, and alter nutrient balance. These risks tend to be more pronounced under field conditions, where cumulative salt inputs and limited leaching may intensify salinity effects. Therefore, the agronomic benefits associated with nutrient supply from treated wastewater should be considered together with the need for careful monitoring and management of soil salinity and sodicity, as highlighted in recent studies [37].

Water productivity showed an inverse behavior to the applied water volume, with higher values observed under lower irrigation depths, highlighting the trade-off between productivity and water-use efficiency in irrigated systems [38]. This behavior indicates that, although increasing irrigation depth promotes growth, water conversion into biomass does not occur proportionally.

Additionally, water productivity is structurally a ratio variable (biomass per unit of applied water), which explains its weak association with morphological traits. Unlike growth variables, WP integrates both production and water input. Therefore, it is more sensitive to changes in irrigation efficiency than to plant size itself, which explains its independent behavior in correlation and predictive analyses.

Finally, differences among treatments were more evident at intermediate and higher irrigation depths. Under greater water availability, yield differences became more evident, allowing clearer differentiation among water sources. Under water-limited conditions, growth tends to be generally constrained regardless of the water source used.

4.2. Multivariate Analysis and Predictive Modeling

The classification models indicated a moderate capacity to predict irrigation depths based on the morphological variables of the plants [16,17]. The Naive Bayes algorithm showed the highest accuracy, close to 55%, which was higher than the performance expected by chance considering the number of evaluated classes. This moderate performance indicates that morphological variables respond to overall water availability, forming a continuum from deficit to excess water supply. However, the frequent confusion between adjacent depths shows that the model detects a gradient, not a sharp classifier. Thus, although plant height, stem diameter, and shoot fresh mass are informative of the general water regime, they are not sufficiently sensitive to precisely predict specific irrigation depths.

This performance suggests that the analyzed morphological characteristics present a measurable relationship with the applied water regime. Specifically, variables such as plant height, stem diameter, and shoot fresh mass directly reflect soil water availability because cell expansion and biomass accumulation depend on turgor maintenance and water-regulated physiological processes [10,27,28].

The analysis of the confusion matrix revealed that adjacent irrigation depths were frequently misclassified [16], suggesting that small differences in the amount of applied water do not generate sufficiently pronounced morphological changes to allow clear distinction among classes. This behavior indicates that the morphological response of lettuce to water management occurs gradually, producing partially overlapping patterns among nearby treatments. Agronomic studies report that lettuce growth often shows linear or quadratic responses to increasing irrigation depth. This behavior tends to reduce morphological separation among adjacent water regimes [12,39].

The prediction of water sources showed lower performance, with accuracy below 30% and no statistical differences among the evaluated algorithms. This result indicates that, under the tested conditions, treated wastewater and conventional water did not produce sufficiently distinct morphological differences in the plants to allow their classification. This outcome may be associated with the presence of essential nutrients in treated wastewater, such as nitrogen, phosphorus, and potassium, which can promote plant growth and reduce the need for complementary mineral fertilization. Thus, when properly treated, these water sources may promote vegetative development similar to that obtained with conventional water combined with mineral fertilization [40,41,42]. The findings of the present study are consistent with reports that properly managed treated wastewater does not negatively affect vegetable growth, reflecting treatment efficiency and the absence of relevant phytotoxicity [6,14].

The attribute importance analysis confirmed that variables associated with vegetative growth are the most informative for predicting irrigation depth. Plant height, stem diameter, and shoot fresh mass showed the highest information gains, indicating sensitivity to variations in water supply. On the other hand, water productivity showed information gain close to zero, indicating that its behavior is regulated independently from the measured morphological characteristics. This occurs because water productivity depends on the relationship between biomass production and total applied water volume. Therefore, it may vary even when plant morphological characteristics remain similar [7,11,30]. Thus, increases in irrigation depth may reduce water productivity when biomass gains do not occur proportionally, a phenomenon frequently observed in lettuce cultivation under high water regimes [12,39].

The correlation network revealed strong associations among shoot fresh mass, shoot dry mass, number of leaves, and plant dimensions, forming an interconnected core of variables that respond coordinately to vegetative growth. Within this structure, shoot fresh mass and stem diameter showed high centrality, indicating that these variables integrate the overall developmental status of the plants and act as structural elements within the correlation network. In contrast, the segregation of water productivity in the network reinforces that this variable operates through different physiological mechanisms, possibly related to transpiration efficiency and water redistribution within the plant. These processes are not directly reflected in observable morphological characteristics.

Taken together, the results of the multivariate analysis and predictive modeling indicate that irrigation depth management is the dominant factor determining lettuce morphology, whereas the type of water, under the evaluated experimental conditions, has minimal impact. The integration of predictive modeling, attribute importance analysis, and correlation networks made it possible to identify structural patterns in crop responses. Morphological variables were sufficiently sensitive to reflect changes in the water regime but showed limited capacity to discriminate management conditions that did not result in marked structural differences, as observed for treated wastewater use.

4.3. Limitations and Implications of the Study

A limitation of this study is the relatively small sample size (150 observations) for a multiclass classification problem involving five irrigation depths, five water sources, and their 25 combinations. Although cross-validation mitigates overfitting, the moderate accuracies reported (≈55% for depth prediction) should be interpreted as exploratory evidence of predictive signal rather than validation of a robust operational classifier. Future studies with larger datasets and independent field trials are needed to confirm the generalizability of these findings.

Another important limitation concerns the characterization of the treated wastewater. Although the effluent was consistently collected from the same treatment plant and outlet, ensuring uniformity in its origin and treatment process, its chemical composition may vary over time due to operational and environmental factors inherent to stabilization pond systems. Therefore, the results obtained in this study are associated with the specific conditions of the evaluated wastewater source. Caution is required when extrapolating these findings to other treatment systems or regions with different effluent characteristics.

The comparative performance of the algorithms should be interpreted with caution. The finding that Naive Bayes outperformed more complex classifiers (Random Forest, J48, SMO) likely reflects the specific characteristics of our dataset: moderate sample size, few predictive attributes, and considerable morphological overlap between neighboring irrigation treatments. Simpler models often generalize better under such conditions, whereas more complex models may capture noise or become unstable. Therefore, this result should not be generalized as evidence that Naive Bayes is always superior for this type of agronomic data. Instead, it highlights the importance of considering dataset scale and class separability when selecting machine learning methods.

Furthermore, the observed superiority of Naive Bayes over more complex algorithms may be specific to the sample size and feature set used here and should therefore be interpreted with caution.

5. Conclusions

This study demonstrated that irrigation depth was the main factor influencing the morphophysiological and agronomic responses of lettuce. Significant effects were observed for all evaluated variables, promoting higher growth values (shoot fresh and dry mass, plant height, stem diameter, and number of leaves), whereas water sources had a less pronounced influence. Water productivity showed an inverse relationship with irrigation depth, being higher under water deficit conditions, highlighting the trade-off between productivity and water use efficiency. In addition, treated wastewater, used alone or combined with biochar and/or fertilization, showed similar or superior performance compared with conventional water, with no adverse effects detected under the evaluated conditions.

Machine learning models showed moderate ability to predict irrigation depth, indicating that morphological variables capture general gradients of water availability. However, they showed limited capacity to distinguish adjacent irrigation levels. Plant height, stem diameter, and shoot fresh mass were identified as the most informative variables for predicting irrigation conditions. In contrast, prediction of water sources showed low accuracy, reinforcing that the different water sources did not induce sufficiently distinct morphological patterns.

Overall, the results indicate that irrigation management plays a dominant role in determining lettuce growth, whereas treated wastewater represents a viable alternative water source under the studied conditions. The integration of experimental and multivariate approaches allowed the identification of consistent response patterns, although further studies including additional plant and soil variables are recommended to better understand more subtle effects of water quality on crop performance.