Enhancing Surface Water Quality Parameter Prediction Using Deep Learning and Feature Augmentation Methods

Abstract

Water quality monitoring is crucial for public health and environmental protection, but traditional methods lack real-time accuracy. This study addresses this gap by combining feature augmentation methods (e.g., sliding window) with Principal Component Analysis (PCA) and a tailored Two-Layer Regularized Gated Recurrent Unit (TLR-GRU) model for efficient water quality prediction. Results demonstrate that the proposed framework significantly improves prediction accuracy, with R² increased by 7.78%, RMSE decreased by 27.36%, MAE by 36.71%, and MAPE by 45.08%. This approach offers a novel technical pathway for real-time environmental monitoring and water resource management.

1. Introduction

Water quality monitoring is a critical practice for safeguarding public health, protecting the environment, and ensuring sustainable water resource management. Globally, water pollution is an escalating concern, and the deterioration of water quality poses serious threats to ecosystem stability and human well-being [1]. Real-time and accurate monitoring and prediction of water quality are therefore essential. Effective monitoring systems not only facilitate the timely detection of and response to contamination events but also supply scientific data to support water resource management and policy-making [2]. With increasing industrialization and climate change impacts worldwide, the significance of robust water quality monitoring continues to grow [3].

In recent years, water quality prediction has emerged as a vital research focus, with growing emphasis on leveraging advanced computational approaches particularly deep learning to enhance predictive accuracy and operational efficiency. For instance, Lee et al. utilized Long Short-Term Memory (LSTM) models to improve early warnings for harmful algal blooms in major South Korean rivers [4]. Similarly, Liu et al. demonstrated the feasibility of LSTM-based deep neural networks for water quality prediction in IoT contexts, stressing the importance of high-quality input data [5]. Hu et al. introduced a deep LSTM approach to predict pH and temperature in smart mariculture, overcoming limitations of traditional forecasting techniques [6]. Other studies have explored soft sensing of 5-day biochemical oxygen demand (BOD) in urban waters using deep learning [7], compared multiple machine learning models for surface water quality prediction using large datasets [8], and developed real-time probabilistic forecasting frameworks combining deep learning with post-processing [9]. Hybrid architectures, such as CNN-LSTM models, have also shown promise for short-term water quality forecasting [10], while AI-driven index prediction and classification aid water management decisions. Beyond water quality, deep reinforcement learning has been applied in related areas such as agricultural yield prediction [11], and comparative analyzes confirm the superiority of deep learning over conventional machine learning in groundwater quality prediction [12]. Collectively, these studies underscore the potential of deep learning in this domain.

However, existing research has two key limitations: (1) a predominant focus on individual model architectures without systematic integration of feature augmentation strategies, leading to suboptimal handling of temporal dependencies; and (2) limited attention to computational efficiency and data redundancy reduction in high-dimensional water quality datasets. These gaps restrict the practical deployment of real-time prediction systems.

Conventional water quality monitoring typically depends on field sampling and subsequent laboratory analysis. Although accurate, these approaches are time-consuming, labor-intensive, and inadequate for real-time monitoring requirements [13]. Recent advances in sensor technologies and information systems have promoted the adoption of online monitoring and automated data acquisition [14]. Nevertheless, challenges such as data instability and limited spatial coverage persist.

In terms of prediction, classical statistical techniques—including regression analysis and time series models—are frequently applied to water quality data, yet they often struggle to capture complex nonlinear dynamics [15]. The emergence of artificial intelligence and machine learning has opened new avenues for improving predictive performance. Methods such as Support Vector Machines (SVMs), Random Forests (RF), and Artificial Neural Networks (ANNs) have demonstrated improved results [16]. However, these techniques generally depend on multi-dimensional datasets, which can be difficult to obtain; they often entail considerable model complexity, extended training times, and limited prediction accuracy [17].

To address these challenges, this study introduces a novel framework that combines multiple feature augmentation techniques with Principal Component Analysis (PCA) and a tailored deep learning model for efficient water quality parameter prediction, as illustrated in Figure 1 [18]. The TLR-GRU model was selected for its superior ability to handle long-term dependencies with lower computational complexity than LSTM, while integrated dropout and batch normalization layers mitigate overfitting. We expect the proposed framework to improve predictive accuracy significantly, with preliminary analyses indicating potential gains of over 5% in R² and 30% reduction in MAPE.

The key innovations are:

• Systematic integration of five feature augmentation methods (time-based, periodic, differenced, lagged, and sliding-window) with PCA for dimensionality reduction, enhancing temporal pattern capture while reducing redundancy.
• Development of a Two-Layer Regularized GRU (TLR-GRU) model optimized for water quality time series data, balancing computational efficiency and generalization performance.
• Comprehensive evaluation of feature augmentation and model configurations, validating accuracy improvements and operational cost reductions through multi-fidelity validation.

By integrating these components, this work provides a new technical pathway for water quality parameter prediction and offers scientific support for environmental monitoring. The results demonstrate that the proposed methodology enhances predictive performance while maintaining practical viability, underscoring its theoretical and practical relevance.

2. Materials and Methods

2.1. Dataset

The water quality dataset utilized in this study was systematically collected from China’s National Surface Water Quality Automatic Monitoring Real-Time Data Release System from January 2020 to December 2023. The dataset encompasses comprehensive monitoring records with a precisely controlled 4 h sampling frequency across 5 strategically distributed monitoring stations. This longitudinal sampling strategy resulted in a substantial dataset of 25,000 samples, providing robust statistical power for analysis.

The monitored parameters include temperature (Temp), pH, dissolved oxygen (DO), permanganate index (CODMn), ammonia nitrogen (NH₃-N), total phosphorus (TP), total nitrogen (TN), conductivity (Cond), and turbidity (Turb). These parameters provide comprehensive characterization of the physical, chemical, and biological properties of water bodies, serving as critical indicators for water quality assessment [19]. The summary statistics of water quality parameters is shown in

Table 1. Summary statistics of water quality parameters.

Parameter	Unit	Min	Max	Mean	Median	Std Dev
Temp	°C	2.75	29.05	18.63	18.82	6.37
pH	–	6.56	8.69	7.57	7.58	0.24
DO	mg/L	2.01	12.34	8.14	8.20	1.61
CODMn	mg/L	0.56	8.12	3.14	3.05	1.38
NH3-N	mg/L	0.01	1.25	0.23	0.20	0.18
TP	mg/L	0.01	0.45	0.11	0.10	0.07
TN	mg/L	0.31	4.57	1.68	1.57	0.74
Cond	µS/cm	56	13,856	2042	1790	1367
Turb	NTU	0.1	133	11.78	6.10	15.45

.

Data quality assurance was implemented through a rigorous multi-stage preprocessing pipeline. Missing values were systematically detected using threshold-based methods, with records containing more than 20% missing data automatically removed to ensure data integrity. The remaining missing values were meticulously addressed using linear interpolation, which preserves temporal continuity while minimizing artificial bias. Anomalous values were identified through statistical outlier detection methods and appropriately handled through deletion strategies. To mitigate potential scale effects and ensure numerical stability during model training, all parameters were normalized to the [0, 1] range using MinMaxScaler. These comprehensive preprocessing procedures guaranteed data reliability and consistency for subsequent feature engineering and model training processes [20].

2.2. Feature Augmentation

Feature augmentation enhances the feature space by generating derived variables, thereby improving the predictive performance and robustness of machine learning models [21]. In water quality prediction, feature augmentation helps capture complex temporal dependencies, periodic patterns, and nonlinear relationships, ultimately enhancing model accuracy [22]. This study systematically evaluates five feature augmentation techniques: differenced features, sliding window features, lag features, time-related features, and periodic features.

2.2.1. Differenced Features

Differenced features capture temporal variations by computing differences between current values and previous observations. This approach helps eliminate trends and seasonal components while enhancing sensitivity to short-term fluctuations [23]. For each water quality parameter “X”, differences were calculated at lags of 1, 6, 42, and 180 time steps:

$$X_{\mathrm{diff}\_\mathrm{n}}=X_t-X_{t-n}$$

(1)

where “n” represents the lag interval. This method effectively highlights short-term variations and anomalous patterns in the data.

2.2.2. Sliding Window Features

Sliding window features capture the local change trends of data by calculating statistical measures (such as mean and standard deviation) within a moving window [24]. The sliding window provides statistical information over specific time intervals, helping the model better understand the local dynamic changes in the data [25]. Select the size of the sliding window, such as 1 h, 3 h, 6 h, 12 h, and 24 h. For each water quality parameter X, we compute the mean and standard deviation within the sliding window:

$$X_{rolling\_mean\_window}=mean(X_{t-window+1},...,X_t)$$

(2)

$$X_{rolling\_std\_window}=std(X_{t-window+1},...,X_t)$$

(3)

The advantage of this method is that it provides dynamic change information over different time periods, helping the model capture the local trends and fluctuation characteristics of the data [26].

2.2.3. Lag Features

Lag features incorporate historical observations as predictor variables, enabling the model to capture temporal dependencies. Both short-term and long-term relationships can be revealed through appropriate lag selection [27]. Lags of 1, 3, 6, 12, and 24 time steps were used:

$$X_{\mathrm{lag}\_\mathrm{n}}=X_{t-n}$$

(4)

This method effectively preserves temporal dependencies and trend information, facilitating better modeling of time series patterns [28].

2.2.4. Time-Related Features

Time-related features extract temporal information from timestamps to capture periodic and seasonal patterns. These features help model diurnal, weekly, and annual cycles in water quality parameters [29]. The following features were extracted: year, month, day, hour, minute, and day of the week:

$$\{\mathrm{Year},\mathrm{Month},\mathrm{Day},\mathrm{Hour},\mathrm{Minute},\mathrm{Weekday}\}$$

(5)

For example, the hour feature is extracted as:

$$\mathrm{Hour}=\mathrm{timestamp}.\mathrm{hour}$$

(6)

These features enhance the model’s ability to identify temporal patterns and seasonal variations [30].

2.2.5. Periodic Features

Periodic features transform temporal variables into cyclic representations using trigonometric functions, effectively capturing recurring patterns. This transformation ensures continuity at period boundaries, improving modeling of cyclical phenomena [31]. Sine and cosine transformations were applied to hour, weekday, and month features:

$$T_{sin}=sin\left({\displaystyle \frac{2\pi T}{P}}\right),T_{cos}=cos\left({\displaystyle \frac{2\pi T}{P}}\right)$$

(7)

where “P” denotes the period (24 for hour, 7 for weekday, 12 for month). This approach enhances the model’s capacity to recognize cyclical patterns in water quality data [32].

2.3. Principal Component Analysis

Principal Component Analysis (PCA) was employed for dimensionality reduction and feature extraction. PCA projects the original data onto orthogonal principal components, retaining maximal variance while reducing redundancy [33]. This improves computational efficiency and model stability in water quality prediction tasks [34].

The PCA procedure involved: (1) standardizing the augmented feature set to zero mean and unit variance; (2) computing the covariance matrix; (3) performing eigenvalue decomposition to obtain eigenvectors and eigenvalues; (4) selecting principal components capturing 95% cumulative variance; and (5) projecting the original data onto the selected components. This process reduced the feature dimensionality from 90 to 30 principal components, achieving a 67% reduction while preserving 95% of the original variance. This yielded a reduced-dimensionality dataset preserving essential information while minimizing redundancy [35]. The retained components were validated to ensure they effectively represented the original feature space [36].

2.4. Prediction Models

This study evaluates seven deep learning architectures for water quality prediction, categorized into recurrent networks, attention mechanisms, and convolutional approaches. The comprehensive model specifications are summarized in

Table 2. Architecture specifications of evaluated deep learning models.

Model	Architecture	Layers	Regularization
TLR-GRU	2-layer GRU	128 × 2, Dense(64)	L2(0.001) + BN + Drop(0.3)
TLD-LSTM	2-layer LSTM	64 × 2	Dropout(0.2)
TLD-Transformer	2-layer Transformer	4 heads × 2, Dense(64)	Drop(0.3) + LayerNorm
DeepAR	Probabilistic LSTM	64 × 2	DistributionLambda
Bi-TLD-LSTM	Bidirectional LSTM	64 × 2	Dropout(0.2)
WaveNet	Dilated CNN	8 res blocks, 64 filters	Causal conv
CNN	2-layer Conv1D	64→128 filters, Dense(64)	L2(0.001) + BN + Drop(0.3)

.

2.4.1. Two-Layer Regularized GRU (TLR-GRU)

The TLR-GRU architecture, as illustrated in Figure 2, employs a sophisticated gated recurrent unit design with comprehensive regularization. The model consists of two GRU layers (128 units each) with sequential processing through batch normalization and dropout (30%) after each recurrent layer. A dense layer with 64 ReLU-activated units processes the temporal features before final regression output. This architecture effectively balances model capacity with regularization strategies, ensuring robust temporal pattern capture while maintaining generalization performance.

2.4.2. Comparative Model Architectures

The remaining six models implement distinct architectural paradigms for temporal data processing. Recurrent-based approaches include: TLD-LSTM (two-layer LSTM with 64 units and 20% dropout), DeepAR (probabilistic forecasting with distributional outputs), and Bi-TLD-LSTM (bidirectional processing with 64 units per layer). Attention mechanisms are represented by TLD-Transformer, employing two multi-head attention layers (4 heads each) with layer normalization. Convolutional architectures comprise WaveNet, using dilated causal convolutions with 8 residual blocks, and a standard CNN with two 1D convolutional layers (64 → 128 filters).

All models were trained using consistent experimental protocols: Adam optimization (learning rate 0.0001, except DeepAR at 0.001), batch size 32, and early stopping (patience = 20) based on validation loss. The dataset was temporally split (70% training, 15% validation, 15% testing) with comprehensive resource monitoring during training.

2.5. Performance Evaluation

Model performance was assessed using multiple metrics to comprehensively evaluate prediction accuracy and error characteristics as shown in

Table 3. Evaluation metrics for model performance assessment.

Metric	Formula	Units/Scale
Coefficient of Determination ($R^2$)	$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\hat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}$	Unitless
	Description: Measures the proportion of variance in the dependent variable explained by the model, indicating goodness of fit.
Root Mean Squared Error (RMSE)	$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(y_i-{\hat{y}}_i)}^2}$	mg/L
	Description: Sensitive to large errors, giving higher weight to significant deviations; useful for emphasizing prediction outliers.
Mean Absolute Error (MAE)	$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_i-{\hat{y}}_i\right|$	mg/L
	Description: Measures average magnitude of absolute errors between predicted and actual values, providing straightforward accuracy interpretation.
Mean Absolute Percentage Error (MAPE)	$\mathrm{MAPE}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|{\displaystyle \frac{y_i-{\hat{y}}_i}{y_i}}\right|\times 100$	%
	Description: Expresses error as percentage of actual values, offering insight into relative error size.

:

Among all the above formulas, the $y_i$ represents the i-th observed value, the ${\hat{y}}_i$ represents the i-th predicted value, and the $\overline{y}$ represents the mean of the observed values.

Computational efficiency was evaluated by monitoring training duration (seconds), memory utilization (GB), CPU usage (%), and GPU consumption (%) at 1-s intervals. This comprehensive assessment enables balanced consideration of prediction accuracy and resource requirements.

All experiments were conducted on a system equipped with an Intel(R) Core(TM) i7-10700F processor and NVIDIA RTX 2060 GPU, using TensorFlow-GPU 2.9 and Python 3.8. To ensure statistical reliability, each configuration was executed 10 times with different random seeds (1–10), with results reported as averages across repetitions.

3. Results and Discussion

This section presents a comprehensive evaluation of the proposed deep learning models for water quality prediction. The analysis compares model performance across different feature augmentation methods, examines computational efficiency, and identifies the optimal configuration based on both accuracy and resource utilization metrics.

3.1. Performance Comparison of Models Under Different Feature Augmentation Methods

Table 4. The average performance evaluation of models with different feature augmentation methods.

Model	Method	Training				Test
		R2	RMSE	MAE	MAPE	R2	RMSE	MAE	MAPE
TLR-GRU	Basic Data	0.913	115.793	85.228	12.991	0.873	148.549	108.948	14.534
	Time Related	0.857	213.280	167.096	17.267	0.840	240.840	173.693	16.527
	Lag	0.938	101.354	69.365	9.925	0.885	122.630	81.648	12.104
	Periodic	0.911	146.558	107.251	12.776	0.890	144.792	103.293	13.115
	Differenced	0.888	147.560	106.400	15.454	0.848	158.227	107.603	16.776
	Sliding Window	0.951	89.735	61.101	9.124	0.941	107.952	68.936	7.979
TLD-LSTM	Basic Data	0.913	113.366	82.992	14.496	0.849	168.109	131.360	18.478
	Time Related	0.813	152.115	110.422	17.932	0.724	202.461	148.094	21.593
	Lag	0.870	170.577	118.930	15.892	0.691	241.940	163.430	22.327
	Periodic	0.901	132.132	92.218	14.022	0.833	224.452	158.547	17.825
	Differenced	0.909	112.010	80.532	13.587	0.807	172.298	128.077	23.113
	Sliding Window	0.931	129.125	90.127	12.646	0.826	214.571	142.890	16.496
TLD-Transf	Basic Data	0.814	149.007	110.862	21.826	0.598	222.514	157.933	25.007
	Time Related	0.803	162.229	124.920	25.340	0.627	254.836	189.363	28.103
	Lag	0.882	145.314	106.409	17.438	0.559	342.915	234.964	32.347
	Periodic	0.872	155.103	118.590	21.471	0.704	282.455	196.928	24.705
	Differenced	0.936	80.819	61.646	13.662	0.782	169.483	129.038	23.298
	Sliding Window	0.936	117.145	87.993	14.286	0.759	262.812	183.256	22.437

presents a comprehensive comparison of three deep learning architectures with six feature augmentation methods. The TLR-GRU model with sliding window features demonstrates superior performance, achieving the highest test R² (0.941) and lowest error metrics. This configuration reduces RMSE by 27.36% and MAE by 36.71% compared to basic data processing.

The sliding window approach consistently outperforms other feature augmentation methods across all models due to its ability to capture local temporal dynamics and short-term fluctuations in water quality data. By maintaining a moving window of historical observations, this method effectively preserves temporal dependencies while reducing noise and outliers that can negatively impact prediction accuracy.

Analysis of training versus test performance reveals important insights about model generalization. The TLR-GRU model shows minimal overfitting, with only a 1.05% decrease in R² from training to test sets. In contrast, the TLD-Transformer exhibits significant overfitting, with a 18.9% performance drop, indicating poor generalization capability. This suggests that the regularization strategies in TLR-GRU (batch normalization and dropout) effectively prevent overfitting while maintaining model capacity.

Computational efficiency analysis (

Table 5. Resource usage of models under different feature augmentation methods.

Models	Methods	Training Time (s)	Memory Usage (MB)	CPU Usage (%)	GPU Usage (%)
TLR-GRU	Basic Data	633.755	66.493	35.919	29.850
	Time Related	538.393	65.729	34.977	30.569
	Lag	512.195	58.090	27.576	31.615
	Periodic	696.379	71.916	55.205	26.961
	Differenced	510.216	64.993	30.144	30.875
	Sliding Window	663.451	70.331	34.571	29.953
TLD-LSTM	Basic Data	250.255	59.758	50.007	31.795
	Time Related	129.588	57.668	38.882	25.475
	Lag	118.028	44.908	30.980	28.357
	Periodic	145.433	55.977	32.454	25.941
	Differenced	111.557	55.573	32.620	24.662
	Sliding Window	205.591	57.227	40.876	28.784
TLD-Transf	Basic Data	155.236	48.928	26.290	27.181
	Time Related	108.376	50.175	25.513	26.825
	Lag	106.475	45.717	25.116	28.724
	Periodic	199.640	45.266	32.118	28.931
	Differenced	115.815	47.025	23.541	26.403
	Sliding Window	116.975	45.649	23.461	26.372

) shows that the sliding window method imposes moderate resource demands. The TLR-GRU implementation requires 663.451 s training time with 70.331 MB memory utilization, representing a favorable balance between computational cost and predictive performance.

Visual analysis of model performance (Figure 3) confirms the superiority of the sliding window approach across all evaluation dimensions. The consistency of improvements across accuracy metrics (R² improvement: 6.70–8.79%) and error reduction (RMSE reduction: 8.55–27.36%; MAE reduction: 0.79–36.71%; MAPE reduction: 19.15–45.08%) demonstrates the robustness of this feature augmentation method.

3.2. Comparison of Other Models

Evaluation of seven deep learning architectures (

Table 6. Performance evaluation of different models on training and test sets.

Models	Training Set				Test Set
	R2	RMSE	MAE	MAPE	R2	RMSE	MAE	MAPE
TLR-GRU	0.951	89.735	61.101	9.124	0.941	107.952	68.936	7.979
TLD-LSTM	0.931	129.125	90.127	12.646	0.826	214.571	142.890	16.496
TLD-Transformer	0.936	117.145	87.993	14.286	0.759	262.812	183.256	22.437
DeepAR	0.943	81.738	54.152	9.252	0.927	115.403	72.691	8.835
Bi_TLD-LSTM	0.934	104.747	72.628	11.227	0.825	212.820	137.285	14.902
WaveNet	0.975	59.007	44.978	8.793	0.870	157.036	107.024	13.991
CNN	0.950	92.607	65.517	9.681	0.792	196.517	122.681	14.577

) reveals significant performance variations. The TLR-GRU model achieves superior generalization with the highest test R² (0.941) and lowest error metrics. The WaveNet architecture, while exhibiting excellent training performance (R²: 0.975), shows reduced test effectiveness (R²: 0.870), indicating potential overfitting.

Computational efficiency analysis (

Table 7. Resource usage of different models.

Models	Training Time (s)	Memory Usage (MB)	CPU Usage (%)	GPU Usage (%)
TLR-GRU	663.451	70.331	34.571	29.953
TLD-LSTM	205.591	57.227	40.876	28.784
TLD-Transformer	116.975	45.649	23.461	26.372
DeepAR	127.748	55.588	36.376	29.472
Bi_TLD-LSTM	201.957	58.715	36.795	26.535
WaveNet	599.714	57.784	8.361	23.930
CNN	737.959	64.610	4.044	27.405

) shows substantial variation in resource requirements. The TLD-Transformer achieves the shortest training time (116.975 s), while the TLR-GRU model occupies an intermediate position with 663.451 s, representing an optimal balance between computational overhead and predictive performance.

Detailed examination of CODMn and TN predictions (Figure 4 and Figure 5) reveals that the TLR-GRU model maintains excellent agreement between predicted and observed values. The model achieves R² values of 0.92 (training) and 0.88 (validation) for CODMn, and 0.95/0.90 for TN, demonstrating robust generalization across different water quality parameters.

The convergence behavior (Figure 6) confirms the TLR-GRU’s training stability, with smooth loss reduction observed across iterations. Comparative prediction curves (Figure 7) further validate the model’s capacity to accurately capture temporal dynamics, particularly in tracking sudden changes and trend variations.

3.3. Summary of Optimal Model

Based on the comprehensive evaluation, the TLR-GRU model with sliding window feature augmentation emerges as the optimal configuration for water quality prediction. This combination achieves the best balance between predictive accuracy (test R²: 0.941) and computational efficiency (training time: 663.451 s). The model’s superior performance stems from its effective handling of temporal dependencies through GRU mechanisms, combined with robust regularization strategies that prevent overfitting while maintaining generalization capability across different water quality parameters.

4. Conclusions

This study develops an integrated framework combining feature augmentation strategies with deep learning to address critical limitations in water quality prediction. The proposed methodology demonstrates significant improvements in predictive accuracy while maintaining computational efficiency, offering a practical solution for real-time environmental monitoring.

The key innovation lies in the synergistic integration of multiple feature augmentation methods—particularly the sliding window approach—with PCA dimensionality reduction and a tailored TLR-GRU architecture. This comprehensive approach effectively captures complex temporal patterns while mitigating overfitting through sophisticated regularization techniques. Experimental results confirm substantial performance enhancements, with the optimal configuration achieving a 7.78% increase in R² and simultaneous error reductions of 27.36% in RMSE, 36.71% in MAE, and 45.08% in MAPE compared to conventional methods.

For practical applications, this research provides water resource managers with an efficient tool for real-time quality monitoring and early warning systems. The framework’s ability to accurately predict key parameters enables proactive intervention and supports data-driven decision-making for sustainable water management. The methodology is readily deployable in existing monitoring infrastructure without substantial additional costs.

Future work will explore extending the framework to incorporate meteorological and hydrological data, developing adaptive feature selection mechanisms, and implementing transfer learning approaches for enhanced generalization across diverse conditions. These advancements will further strengthen the framework’s applicability in real-world environmental monitoring scenarios.