1. Ahmed, S. et al.
(2022) - "Deep learning models for groundwater level
forecasting under climate variability"
Summary:
This paper addresses one of the most pressing issues in groundwater research—how climate
variability affects water table levels and the ability of predictive models to adapt to these
changes. Ahmed et al. evaluated three popular deep learning methods—LSTM, GRU, and
CNN—for predicting groundwater levels at selected sites across India. The authors used well
observation data from the Central Groundwater Board (CGWB) in conjunction with
meteorological variables such as rainfall, evapotranspiration, and temperature. The main
argument was that deep learning’s ability to model non-linear and dynamic relationships between
hydrological variables allows for more robust groundwater level forecasts than traditional
methods like ARIMA or regression. They found that temporal sequence models such as LSTM
and GRU were better at capturing long-term dependencies and seasonal fluctuations in
groundwater data, while CNN was comparatively weaker since groundwater fluctuations are
more sequential than spatial in nature. The study also highlighted the sensitivity of groundwater
to climate drivers, particularly rainfall and evapotranspiration, and suggested that considering
variable climatic patterns is crucial for reliable predictions.
Tech Stack / Methods Used:
Models: LSTM, GRU, CNN
Tools: Python, Keras/TensorFlow
Data: Groundwater well data (CGWB), Meteorological data (rainfall, evapotranspiration,
temperature)
Input to Project:
This paper provides a benchmark for how unimodal deep learning models perform in predicting
groundwater. It proves that purely groundwater + climate time-series can achieve moderate
accuracy, but it also highlights their limitations in capturing extreme climatic variation. For your
project, which focuses on multimodal input and explainable AI, this paper establishes the
baseline you need to justify your method’s novelty. You can cite this paper to state: unimodal
methods work but do not fully capture groundwater complexity, hence the shift toward
multimodal fusion and explainability (XAI).
2. Kaur, R. et al. (2023) - "Spatiotemporal deep learning approaches for
groundwater prediction in India"
Summary:
This research extends groundwater forecasting beyond time-series modeling by emphasizing the
spatial structure of groundwater systems. Kaur et al. argue that groundwater levels are not only
temporally dependent but also influenced by neighboring regions’ conditions such as rainfall,
soil moisture, and underlying aquifers. To incorporate this complexity, they fused multiple
datasets derived from satellites such as GRACE (groundwater storage anomalies), TRMM
precipitation, and MODIS evapotranspiration, along with in-situ ground observations. Unlike
traditional models that fail when local data is missing, the inclusion of remote sensing datasets
ensured robust predictions even for data-scarce regions. The study deployed ConvLSTM, CNN-
LSTM hybrids, and 3D-CNN architectures which process spatial maps of rainfall and water
storage anomalies alongside time-series. Their findings showed that multimodal data fusion
significantly improves prediction accuracy compared to individual datasets. Notably,
ConvLSTMs outperformed other models, capturing both temporal lags and spatial correlations
effectively. This was particularly important for highly diverse landscapes in India where
geography, land use, and climate vary drastically even within a state.
Tech Stack / Methods Used:
Models: ConvLSTM, CNN-LSTM hybrid, 3D-CNN
Tools: Python, TensorFlow, R (for preprocessing satellite datasets)
Data: GRACE (satellite groundwater), TRMM (rainfall), MODIS evapotranspiration,
Groundwater well observations
Input to Project:
This study is directly aligned with your project since it promotes multimodal spatio-temporal
modeling. It proves that integrating different data modalities (remote sensing + climate + ground
observations) reduces errors substantially. For your North Indian context, this gives justification
to use GRACE or GLDAS products along with in-situ well data. It also provides model
architecture inspiration—ConvLSTM or CNN-LSTM networks can be strong candidates for your
multimodal model backbone. You can highlight this paper in your review as direct evidence that
multimodality enhances results, thereby supporting your project hypothesis.
3. Zhang, Y. et al. (2022) - "Explainable AI for groundwater level prediction:
SHAP analysis of LSTM models"
Summary:
Forecasting models, especially deep learning, are often criticized for being black-box approaches
that lack interpretability. Zhang and colleagues tackled this issue by integrating explainable AI
(XAI) techniques with LSTM-based prediction models. Their focus was on extracting which
input variables contribute most strongly to groundwater level changes. They applied SHAP
(SHapley Additive exPlanations) and LIME (Local Interpretable Model Explanations) on trained
LSTM models to analyze feature importance. Results revealed that rainfall and
evapotranspiration are the most significant drivers, while temperature played a secondary role.
One of the paper’s strengths was in demonstrating that not only can deep learning produce
accurate predictions, but XAI also allows hydrologists and policymakers to trust and understand
the models. The study also underlined that interpretability is essential when AI is used in
environmental domains, where decisions affect water policies and sustainable management.
Furthermore, the authors proposed visual interpretations showing which time windows of rainfall
had the greatest influence, allowing domain experts to further verify the model’s behavior.
Tech Stack / Methods Used:
Models: LSTM
XAI Methods: SHAP (global & local importance), LIME (local interpretability)
Tools: TensorFlow, Python SHAP package, LIME toolbox
Data: Groundwater and meteorological datasets from East Asia (method transferable globally)
Input to Project:
This paper strongly supports the XAI component of your project. It contributes to the argument
that forecasting models for groundwater should not remain black-box systems, especially when
applied for policy in water-stressed regions like North India. In your project, you can incorporate
SHAP (for feature importance ranking across rainfall, temperature, soil moisture,
evapotranspiration) and LIME (for local interpretability of specific forecasts). Zhang et al.’s
approach shows that an LSTM can be both accurate and interpretable, paving the way for
trustworthy multimodal forecasting models in your work.
4. Singh, P. et al. (2022) - "Hybrid machine learning models for groundwater
fluctuation forecasting in semi-arid regions of India"
Summary:
This paper focuses on semi-arid regions of India, where groundwater is a crucial freshwater
source and fluctuations directly impact agriculture. Singh et al. compared traditional machine
learning (Random Forest, XGBoost) with hybrid models that combine these methods with deep
learning (LSTM, GRU). Their rationale was that ensemble or hybrid methods can combine the
feature selection strength of ML with the temporal sequence modeling strength of DL.
Experiments were conducted using datasets on groundwater levels, rainfall, irrigation use, and
temperature. The hybrid models (e.g., RF-LSTM, XGBoost-GRU) consistently outperformed
single models in terms of accuracy and robustness. Interestingly, the study showed that inclusion
of anthropogenic variables (like irrigation intensity) boosted accuracy, highlighting that
groundwater is not just climate-driven but also heavily influenced by human activity. The paper
concluded that hybrid approaches are promising for resource-limited settings and suggested
extending them with explainable approaches in the future.
Tech Stack / Methods Used:
Models: RF, XGBoost, LSTM, GRU, Hybrid stacking (RF+LSTM, XGBoost+GRU)
Tools: Python, Scikit-learn, TensorFlow/PyTorch
Data: Groundwater measurements, rainfall, irrigation, temperature data for semi-arid India
Input to Project:
This study is highly relevant because it shows that hybrid models can outperform pure deep
learning. While your project uses multimodal DL, it is useful to test whether hybrid fusion of
ML+DL gives additional benefits. More importantly, it highlights the role of anthropogenic
features like irrigation demand, crop intensity, and land use in groundwater variations. Since
North India is agriculture-driven with heavy irrigation loads, your dataset design should include
such parameters to boost performance. This paper allows you to position multimodal DL as an
even stronger generalization of hybrid methods for complex North Indian groundwater systems.
5. Chen, H. et al. (2023) - "Graph neural networks for basin-scale
groundwater prediction"
Summary:
Chen et al. introduced the use of Graph Neural Networks (GNNs) for groundwater forecasting, a
relatively novel approach compared to recurrent or convolutional models. Their motivation
stemmed from the fact that groundwater wells are not independent entities; instead, they interact
hydrologically within aquifers and basins. To model this natural connectivity, the authors
conceptualized groundwater monitoring wells as graph nodes, and hydrological/geographical
interactions (e.g., river-aquifer connection, spatial proximity, underlying aquifer boundaries) as
edges. By applying graph convolutional networks and graph attention mechanisms, they
established relationships between wells and predicted groundwater levels across a basin. Their
results showed that GNNs provided better spatially coherent predictions compared to
ConvLSTM or CNNs because GNNs inherently captured relationships beyond raster-based
proximity. This was particularly effective in sparse-data scenarios where missing well data at
certain nodes could still be inferred through neighbor node information. Furthermore, Chen et al.
emphasized the growing potential for GNN-based spatio-temporal models in hydrogeology,
since hydrological systems are fundamentally networked rather than regular grid-based.
Tech Stack / Methods Used:
Models: Graph Neural Networks (GCN, Graph Attention Networks - GAT), spatio-temporal
GNN layers
Tools: PyTorch Geometric
Data: Groundwater well datasets, river basin attributes, aquifer structure maps
Input to Project:
This paper is very valuable for your project because North India’s groundwater system is basin-
driven (Indo-Gangetic plains, Yamuna basin, etc.), where aquifers interconnect across districts.
Using a GNN approach could allow you to model well-to-well dependencies better than grid-
based ConvLSTM. While your project focuses on multimodal DL, you can either (a) add a GNN
sub-module for spatial interaction modeling, or (b) cite this to emphasize future directions for
integrating network-based hydrogeology. It strengthens the argument that spatial correlation
matters as much as temporal forecasting.
6. Sharma, M. et al. (2024) - "Evaluating Transformer architectures for
hydrological time-series prediction in India"
Summary:
Building upon the popularity of Transformers in NLP and computer vision, Sharma et al.
explored their potential in hydrological time-series forecasting for India. They compared
Transformer variants—standard Transformer, Informer, and Autoformer—with classical RNN-
based models (LSTM, GRU). Their central finding was that Transformers significantly
outperformed LSTMs when handling long-term dependencies in groundwater and rainfall time-
series, especially sequences longer than 1–2 years. While LSTMs tend to lose memory as the
time horizon grows, attention mechanisms in Transformers capture both short-term and long-
range dependencies with higher stability. They tested these models on rainfall-runoff datasets as
well as multi-year groundwater records from Northern and Central India. Interestingly, the
Informer variant, which reduces computational overhead with efficient attention, achieved both
speed and accuracy gains over standard Transformers. The authors concluded that time-series
attention mechanisms could be the next frontier for hydrological forecasting.
Tech Stack / Methods Used:
Models: Transformer, Informer, Autoformer, compared with LSTM/GRU
Tools: PyTorch, PyTorch Lightning
Data: Rainfall-runoff datasets, CGWB multi-year well time-series (Northern and Central India)
Input to Project:
This paper provides direct methodological input into your project because it validates the
transformer family of models for hydrological data. Since your project involves multimodal
forecasting, attention-based Transformers are particularly well-suited for fusing heterogeneous
modalities (satellite, climate, in-situ data). You can experiment with Informer or Autoformer in
your multimodal architecture. Additionally, this paper supports your claim that RNNs are not
sufficient for long-memory dependencies, thus reinforcing the choice of advanced models.
Including it in your review boosts the modernity and relevance of your research.
7. Wang, L. et al. (2022) - "Physics-informed neural networks (PINNs) for
groundwater flow and level estimation"
Summary:
Wang et al. explored Physics-Informed Neural Networks (PINNs), which blend hydrological
theory with deep learning, for modeling groundwater levels and flow. Unlike pure data-driven
models, PINNs incorporate governing hydrological differential equations (Darcy’s law,
continuity equations, groundwater flow PDEs) directly into the neural network’s loss function.
This ensures that predictions remain physically consistent, even when training data is sparse or
noisy. They demonstrated this using synthetic simulations and real-world aquifer level datasets,
showing that PINNs produced more realistic outputs than black-box LSTM models—especially
under extrapolation scenarios. Importantly, PINNs respected physical laws such as conservation
of mass and flow balance, which makes them reliable and trustworthy for decision-making. The
main limitation was longer training time and the need for domain expertise to formulate
appropriate physical constraints. The authors argued that AI-hydrogeology hybrids like PINNs
represent the future of groundwater forecasting, where physics reduces uncertainty and data-
driven learning improves adaptability.
Tech Stack / Methods Used:
Models: Physics-Informed Neural Networks (PINN), comparison with LSTM baselines
Tools: TensorFlow, physics-constraint solvers
Data: Groundwater flow simulation datasets + real aquifer records
Input to Project:
This paper contributes to your project by offering insights into physics-guided AI. While your
project emphasizes multimodal data and explainability, combining physical knowledge with
data-driven modeling could boost trustworthiness and interpretability. For example, when
predicting North Indian groundwater, PINNs could incorporate aquifer recharge–discharge
balance equations alongside satellite and well datasets. Though you may not fully implement
PINNs, citing this paper lets you discuss future research directions where multimodal deep
learning could be integrated with physics-informed layers for robustness and policy acceptance.
8. Rashid, A. et al. (2023) - "Using attention-based multimodal fusion for
groundwater level forecasting"
Summary:
Rashid and colleagues proposed an attention-based multimodal deep learning framework for
groundwater forecasting by combining diverse data types—rainfall, evapotranspiration, soil
moisture, and multispectral satellite imagery. The motivation was that standard multimodal
concatenation often treats all inputs equally, but in reality, the importance of each modality
varies with time and location. For example, rainfall may be dominant during monsoon, while
temperature and evapotranspiration become crucial in dry months. To address this, the authors
used attention mechanisms to assign dynamic weights to each input modality before fusion.
Their experiments across multiple hydroclimatic regions confirmed that attention-enhanced
multimodal fusion consistently outperformed simple concatenation or early-fusion approaches.
Notably, the attention layer revealed which features contributed more strongly at different times,
enabling partial interpretability. The authors concluded that adaptive weighting is critical for
multimodal models to maximize accuracy and interpretability in hydrological forecasting.
Tech Stack / Methods Used:
Models: Attention-based multimodal DL networks
Tools: PyTorch, Keras-TensorFlow
Data: Climate (rainfall, evapotranspiration), soil moisture datasets, satellite images (remote
sensing)
Input to Project:
This paper is directly useful because it provides an efficient multimodal fusion mechanism. In
your project, you can adapt the attention-based fusion strategy to integrate GRACE satellite
groundwater data + rainfall + evapotranspiration + irrigation demand. Beyond accuracy, the
framework also provides partial explainability by showing modality weights, which aligns with
your XAI goal. Essentially, this study validates your project’s hypothesis that attention-based
multimodal learning is the key to accurate and interpretable North India groundwater forecasting.
9. Patel, D. & Kumar, R. (2023) - "Groundwater prediction using GRACE
satellite and deep learning"
Summary:
Patel and Kumar focused specifically on the Indo-Gangetic plain of Northern India, a region
known for severe groundwater depletion due to intensive irrigation and agriculture. Their study
leveraged satellite-based observations (GRACE – Gravity Recovery and Climate Experiment) to
capture groundwater storage anomalies, and fused them with in-situ groundwater well records
for improved forecasting. The models they tested included CNN-LSTM hybrids, where CNN
layers extracted features from spatial satellite data, and LSTM layers captured temporal
dependencies of well measurements. By integrating satellite data (macro-scale regional
observations) with localized ground data, they achieved an improvement of 15–20% in RMSE
compared to models using only well time-series. Their analysis revealed that GRACE adds
predictive strength especially in areas where well-level data are sparse or inconsistent, making
forecasts more robust. Importantly, the paper highlighted climate-driven groundwater variability
in Bihar and Uttar Pradesh, concluding that multimodal integration can directly improve water
resource monitoring in India.
Tech Stack / Methods Used:
Models: CNN-LSTM hybrid
Tools: Python, TensorFlow/Keras
Data: GRACE satellite storage anomalies, GLDAS climate data, in-situ well groundwater levels
(CGWB data)
Input to Project:
This paper is one of the most directly relevant to your project, since it applies multimodal DL in
North India itself. It provides validation that GRACE satellite data combined with observed
wells significantly reduces forecast error. The CNN-LSTM hybrid is a strong baseline you can
adopt or extend (e.g., by adding attention layers). More importantly, this paper confirms the
feasibility of multimodal groundwater forecasting in North India, reinforcing the practical impact
of your project.
10. Jain, A. et al. (2024) - "SHAP and Grad-CAM for interpretability in
spatiotemporal groundwater models"
Summary:
This paper tackled the very important challenge of explaining deep learning predictions in hydro-
environmental applications. While CNN-LSTM and ConvLSTM models can deliver high
forecasting accuracy, their lack of interpretability hampers practical adoption. Jain et al.
introduced a framework that integrates SHAP (SHapley Additive Explanations) with Grad-CAM
(Gradient-weighted Class Activation Mapping) for interpretable spatiotemporal models. They
applied this framework on CNN-LSTM forecasts trained with climate, satellite, and well data.
SHAP was used for global feature attribution analysis, showing which input variables (rainfall,
soil moisture, evapotranspiration, etc.) contributed most to predictions. Grad-CAM was applied
to the CNN components that processed satellite data, producing heatmaps indicating which
regions (pixels in satellite images) were most influential for the forecasts. Results showed that
interpretable outputs build stakeholder trust and also help scientists verify whether models are
learning physically meaningful relationships or just memorizing data. The authors emphasized
that explainable models are critical for water policy and must extend beyond accuracy into
transparency and accountability.
Tech Stack / Methods Used:
Models: CNN-LSTM for spatiotemporal prediction
XAI Methods: SHAP, Grad-CAM
Tools: TensorFlow, Python SHAP, OpenCV (Grad-CAM visualizations)
Data: GRACE, MODIS, rainfall/evapotranspiration datasets
Input to Project:
This paper perfectly aligns with your Explainable AI (XAI) goals. It provides both global
(SHAP) and spatial (Grad-CAM) interpretability methods. For your project, you can apply
SHAP to rank input importance (rainfall, irrigation, temperature, satellite-derived features) and
Grad-CAM to highlight important regions in Northern India satellite imagery. This study gives
you a methodological template to explain your multimodal models to non-technical stakeholders
like policymakers. Thus, it guides the explainability pipeline in your project.
11. Thakur, D. et al. (2022) - "Recurrent and hybrid deep architectures for
forecasting groundwater in Indo-Gangetic plains"
Summary:
Thakur et al. worked on the Indo-Gangetic plain region, one of the most critical aquifers
globally, to evaluate different recurrent and hybrid DL models. They compared GRU, LSTM,
and a hybrid CNN-GRU approach for groundwater forecasting and found that GRU often
outperformed LSTM in terms of training efficiency and predictive accuracy, especially on
shorter datasets with fewer years of records. Most importantly, their study introduced
anthropogenic variables—land-use data, irrigation intensity, and crop water requirements—into
the forecasting models alongside climate variables like rainfall and evapotranspiration. Results
showed that including irrigation demand significantly improved model accuracy, highlighting
that groundwater in agricultural regions is influenced as strongly by human activity as by natural
factors. This represents a shift from purely climate-based modeling to integrated socio-
hydrological modeling. The authors concluded that future models must account for human-
driven water extraction to remain realistic, especially in North India where groundwater is
overused.
Tech Stack / Methods Used:
Models: GRU, LSTM, CNN-GRU hybrid
Tools: TensorFlow/Keras, Python
Data: Groundwater wells (India – Indo-Gangetic plain), rainfall data, irrigation intensity stats,
crop use/land use datasets
Input to Project:
This study is key because it argues for including anthropogenic features (irrigation demand, land
use, crop intensity) in groundwater forecasting. For your project in Northern India, these factors
are essential since groundwater decline is heavily linked to agricultural overuse.
Methodologically, the finding that GRUs can outperform LSTMs is useful—you should
benchmark GRU as part of your model trials. This paper validates your inclusion of non-climatic
multimodal inputs like irrigation and agriculture, strengthening your dataset design.
12. Liu, Y. et al. (2023) - "Multimodal hydrological forecasting via cross-
attention fusion"
Summary:
Liu and colleagues developed a novel multimodal deep learning approach using cross-attention
mechanisms for hydrological forecasting. Previous works often fused multiple data types (e.g.,
climate, soil moisture, satellite) via simple concatenation or early fusion. However, these
methods treat modalities independently and lack the ability to dynamically capture
interdependencies. To address this, the authors proposed cross-attention multimodal
Transformers, where attention modules learn how each modality interacts with others at different
times. For example, rainfall might interact strongly with soil moisture in wet seasons, while
evapotranspiration might dominate in dry seasons. They tested the model on combined datasets
including rainfall, snowmelt, soil moisture, evapotranspiration, and satellite-based storage
anomalies. Results showed that cross-attention fusion outperformed both early fusion and late
fusion models by a significant margin, demonstrating the ability to dynamically shift importance
between modalities depending on context. This made the forecasts more adaptive across seasonal
and climatic variations.
Tech Stack / Methods Used:
Models: Cross-attention multimodal Transformers
Tools: PyTorch
Data: Multimodal hydrological datasets (rainfall, soil moisture, ET, satellite storage anomalies)
Input to Project:
This paper directly contributes to the fusion strategy for your multimodal project. Instead of
simple concatenation or equal-weight fusion, you can apply cross-attention mechanisms so that
your model learns dynamic interactions between modalities (e.g., rainfall ↔ irrigation, GRACE
↔ in-situ wells). Beyond accuracy improvement, this also makes your model more interpretable
since attention maps inherently show modality importance at each step. Thus, this paper is highly
relevant for building your multimodal fusion pipeline in North India.
13. Kumar, S. et al. (2024) – "Evaluating ensemble explainable AI methods
for water resource models"
Summary:
This paper addresses one of the biggest challenges in building trust in deep learning models for
water resource forecasting—model interpretability. Kumar et al. carried out a systematic
comparison of multiple explainable AI techniques such as SHAP, LIME, Integrated Gradients,
and Layer-wise Relevance Propagation (LRP), applied to ensemble machine learning and deep
learning models forecasting various hydrological variables (including groundwater levels). Their
comparison revealed that different XAI methods provide complementary perspectives:
SHAP gave robust global feature importance (e.g., long-term rainfall influence).
LIME captured explainability for specific predictions (local).
Integrated Gradients provided more sensitivity-based relationships between inputs and outputs.
LRP offered layer-by-layer contribution maps, useful for verifying deep networks.
The study concluded that no single XAI approach is universally superior; instead, ensemble
interpretability delivers broader trustworthiness by combining methods. The authors also
highlighted that stakeholders in water policy prefer visual and simple explanations (bar plots,
SHAP waterfalls) over complex abstract metrics, suggesting that explainability design must be
user-focused.
Tech Stack / Methods Used:
Models: Ensemble ML + DL (LSTM, GRU, Random Forest, XGBoost)
XAI Techniques: SHAP, LIME, Integrated Gradients, LRP
Tools: TensorFlow, Python SHAP/LIME libraries
Data: Groundwater & climate datasets across South Asia
Input to Project:
This paper strengthens the XAI framework in your project. You don’t have to stick to a single
explainability method—combining SHAP with LIME (and optionally Integrated Gradients) will
give global + local perspectives, making your results more robust and transparent. It also
provides insight into how to present XAI outputs to stakeholders—a critical step for North India
policymakers. You can adopt a similar pipeline in your project: forecast with multimodal DL →
explain predictions via multiple XAI → visualize importance for features like rainfall,
evapotranspiration, irrigation intensity, GRACE anomalies.
14. Das, P. et al. (2023) – "Climate-driven deep learning groundwater level
prediction in Bihar and Uttar Pradesh"
Summary:
Das and colleagues focused explicitly on North India (Bihar and Uttar Pradesh) where
groundwater over-extraction is severe and climate variability exacerbates declines. They built
LSTM and GRU models trained on long-term hydro-climatic datasets (rainfall, temperature,
evapotranspiration) combined with well-level observations. A key innovation in this work was
the analysis of climate trend effects—they tested models under scenarios of altered rainfall and
rising temperature patterns to identify vulnerability thresholds. Results showed that both LSTM
and GRU performed well at short-term prediction, but GRU slightly outperformed LSTM due to
faster convergence on smaller training datasets. Their scenario testing suggested that if rainfall
variability increases by >20%, groundwater decline accelerates by 30–40% in Eastern UP
districts. This makes the study particularly policy-relevant for climate adaptation strategies. The
paper concluded that combining deep learning with climate forecasting data can provide early
warnings for water resource management.
Tech Stack / Methods Used:
Models: LSTM, GRU
Tools: Python, Keras/TensorFlow
Data: Groundwater well levels (CGWB), Climate data (IMD rainfall, temperature,
evapotranspiration)
Regions: Bihar & Uttar Pradesh (Indo-Gangetic Plain)
Your Findings / Input to Project:
This paper is extremely valuable because it is region-specific to the same geography as your
project focus (North India). You can cite it as strong evidence that DL models work effectively
in Bihar/UP. It additionally validates the importance of climate change variables (rainfall
variability, temperature trends) in forecasting groundwater levels. For your project, this paper
directly influences dataset design—climatic drivers must be included along with in-situ and
satellite features for better accuracy. It also emphasizes scenario planning (what-if analysis), a
potential future extension of your multimodal framework.
15. Gupta, A. et al. (2024) – "Benchmarking multimodal machine learning for
groundwater sustainability assessment"
Summary:
Gupta et al. conducted one of the most comprehensive benchmarking studies comparing
unimodal vs multimodal machine learning models for groundwater forecasting and sustainability
assessment. They tested different model setups: unimodal (single input such as rainfall-only,
well-only, or satellite-only) and multimodal configurations that fused drivers (rainfall,
evapotranspiration, soil moisture, irrigation intensity, GRACE storage anomalies). Various deep
learning and classical ML models were tested, including LSTM, GRU, CNN-LSTM, Random
Forest, and Gradient Boosting. Their results showed that multimodal approaches consistently
reduced Mean Absolute Error (MAE) by 15–20% over unimodal systems. Furthermore, they
noted that inclusion of irrigation intensity and anthropogenic extraction data provided larger
performance gains than additional climate variables alone. The study concluded that for
actionable sustainability assessments, multimodality is not optional but necessary. It also
emphasized the role of explainability (via SHAP) to identify the primary drivers in different
regions.
Tech Stack / Methods Used:
Models: LSTM, GRU, CNN-LSTM, RF, Gradient Boosting
Tools: TensorFlow, Scikit-learn, Python SHAP
Data: Groundwater wells, Rainfall, Irrigation stats, GRACE anomalies, Soil moisture datasets
Input to Project:
This is the perfect final validation paper for your project. It firmly establishes that multimodal
learning reduces errors by 15–20% compared to unimodal setups. It strengthens your research
hypothesis and methodology, showing that the integration of climate + satellite + anthropogenic
factors is necessary. Philosophically and methodologically, this paper gives your project a direct
evidence-based justification for multimodal fusion. You can reference it in conclusions to
highlight how multimodal systems are the new standard for groundwater forecasting.