Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer Learning-Based CNN-LSTM and ERA5-Land

Vol. 8, No. 5, October 2026

5463
Journal of Information Systems and Informatics | ISSN: 2656-5935 | e-ISSN: 2656-4882 | pp. 5463-5487
Published by Asosiasi Doktor Sistem Informasi Indonesia
Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using
Transfer Learning-Based CNN-LSTM and ERA5 Land

Yampi R Kaesmetan¹, Rahmat Gernowo², Budi Warsito³

¹,2,3Doctoral Program of Information Systems, School of Postgraduate Studies, Universitas Diponegoro,
Semarang, Indonesia
²Department of Physics, Faculty of Science and Mathematics, Universitas Diponegoro, Semarang, Indonesia
³Department of Statistics, Faculty of Science and Mathematics, Universitas Diponegoro, Semarang,
Indonesia
1Informatics Engineering Study Program, STIKOM Uyelindo Kupang, NTT, Indonesia

Received:
May 29, 2026
Revised:
June 18, 2026
Accepted:
September 1, 2026
Published:
October 24, 2026

Corresponding Author:
Author Name*:
Yampi R Kaesmetan
Email*:
yampikaesmetan@student
s.undip.ac.id

DOI:
10.63158/journalisi.v8i5.1707

© 2026 Journal of
Information Systems and
Informatics. This open
access article is distributed
under a (CC-BY License)

Abstract. Poverty dynamics in Nusa Tenggara Timur (NTT), Indonesia,
are influenced by climate variability, particularly rainfall deficits and
drought, which increase household economic vulnerability. This study
develops a transfer learning-based Hybrid CNN-LSTM model to
predict the annual poverty headcount ratio (%) across 22
regencies/municipalities in NTT. ERA5 -Land monthly total
precipitation and annual regency-level poverty data from 2003–2024
were integrated, yielding 484 regency-year observations. To address
limited local time-series data, the CNN was pretrained on long-term
ERA5-Land climate sequences and transferred to the NTT task, while
the LSTM and output layers were fine-tuned using local data. The
model achieved R² = 0.87 and outperformed standalone CNN and
LSTM models. A 6–12-month lagged association between rainfall
deficits and subsequent poverty changes was also identified,
suggesting a potential anticipatory climate signal rather than direct
causality. The findings highlight the potential of integrating climate
reanalysis, transfer learning, and deep learning for poverty -risk
prediction in data-constrained regions. However, further operational,
institutional, and external validation is required before deployment
for decision support.

Keywords: Climate-Poverty Prediction; Climate Risk Analytics; CNN-
LSTM; Transfer Learning; ERA5-Land

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5464

1. INTRODUCTION

Poverty remains a persistent development challenge in Nusa Tenggara Timur (NTT),
Indonesia, where many rural households depend heavily on rain-fed agriculture and
climate-sensitive livelihoods [1]. The province's archipelagic geography, heterogeneous
climatic conditions, uneven infrastructure, and recurrent rainfall deficits increase the
exposure of agricultural communities to climate-related economic shocks. Periods of low
precipitation and drought can disrupt crop production, household income, food
availability, and livelihood stability [2]. However, these pathways are mediated by
socioeconomic factors such as market access, food prices, employment, social protection,
migration, and household coping capacity. Therefore, climate variability should not be
interpreted as a direct cause of poverty; rather, it represents one potential source of
predictive information associated with subsequent changes in poverty outcomes [3].

Timely poverty information is particularly important in climate-sensitive regions. Official
poverty statistics provide essential socioeconomic information but are generally
available at lower temporal frequency than climate observations. Consequently, changes
in household welfare associated with climatic stress may emerge before they are fully
reflected in annual poverty statistics [4]. This temporal mismatch motivates the
development of predictive approaches that can investigate whether preceding climate
conditions contain useful signals for anticipating subsequent poverty dynamics. In this
study, poverty dynamics are operationalized specifically as the annual poverty headcount
ratio (%) at the regency level. The objective is therefore to predict subsequent poverty
rates from preceding precipitation patterns, rather than to establish a causal effect of
climate variability on poverty [5].

Advances in remote sensing, Earth observation, and machine learning have created new
opportunities for poverty and welfare estimation in data -constrained regions [6].
Demonstrated that satellite imagery combined with machine learning could predict
spatial variation in poverty-related welfare indicators, showed that transfer learning
from remotely sensed imagery could extract features informative for poverty mapping
when labeled socioeconomic data were limited . Related work has also demonstrated the
potential of alternative digital data sources for estimating poverty and wealth where
conventional socioeconomic observations are sparse. These studies establish an

Vol. 8, No. 5, October 2026

5465 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
important methodological foundation for data-driven poverty estimation; however, much
of this literature emphasizes spatial poverty mapping or contemporaneous welfare
estimation, rather than the temporal prediction of annual poverty dynamics using
preceding climate sequences [7].

Climate information offers a complementary perspective because climate reanalysis
products provide temporally consistent environmental observations over long periods. In
climate-sensitive agricultural economies, precipitation anomalies and drought may
precede changes in production, income, food security, asset holdings, and other
components of household welfare. Nevertheless, the relationship is complex and should
be interpreted cautiously. An observed association between rainfall deficits and later
poverty outcomes does not by itself demonstrate that drought caused the observed
change in poverty. Poverty outcomes may simultaneously reflect demographic change,
economic conditions, food prices, employment, government assistance, infrastructure,
migration, and other unobserved factors. Accordingly, the present study treats
precipitation as a predictive climate signal rather than as an isolated causal determinant
of poverty [8].

Machine learning and deep learning provide useful tools for identifying nonlinear and
temporally dependent patterns in such data. Convolutional Neural Networks (CNNs) can
extract local patterns from sequential inputs, whereas Long Short-Term Memory (LSTM)
networks are designed to represent temporal dependencies [9]. Their integration
therefore offers a potential framework for learning both short -term precipitation
features and longer temporal relationships associated with poverty dynamics. However,
applying deep learning to regional socioeconomic prediction presents an important
limitation: local poverty time series are relatively short. Transfer learning provides one
strategy for addressing this constraint by allowing representations learned from larger
climate datasets to be adapted to a smaller local prediction ta sk. Previous remote-
sensing poverty research has demonstrated the usefulness of transfer learning when
socioeconomic labels are scarce [10].

Early-warning research further demonstrates the value of transforming environmental
observations into forward-looking indicators [11]. For example, drought early-warning
studies have used satellite-derived vegetation conditions and forecasting models to

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5466
anticipate deteriorating environmental conditions before impacts become fully
observable. Such studies support the broader rationale for anticipatory climate -risk
monitoring. However, the present study does not develop or validate an operational
drought early-warning system, institutional response workflow, or policy -trigger
mechanism. Instead, early warning is treated as a potential downstream application that
would require additional operational and institutional validation[12].

Against this background, an important research gap remains in understanding whether
long-term climate information can be combined with limited local socioeconomic
observations to predict annual poverty dynamics in climate -sensitive and data-
constrained regions such as NTT. Existing satellite-based poverty studies demonstrate
that remotely sensed information can support poverty and welfare estimation [13], while
drought-forecasting studies demonstrate that environmental signals can provide useful
anticipatory information. However, comparatively less attention has been given to
integrating long-term climate reanalysis, transfer learning, and hybrid temporal deep
learning to predict subsequent annual poverty rates at the sub-provincial level [14].

To address this gap, this study develops a transfer learning-based Hybrid CNN-LSTM
framework that integrates monthly ERA5-Land precipitation sequences with annual
poverty observations from 22 regencies/municipalities in NTT during 2003–2024. The
model uses a 12-month precipitation sequence as the climatic input and the subsequent
annual poverty headcount ratio (%) as the prediction target. Transfer learning is
employed to leverage long-term climate information before fine-tuning the model to
local climate–poverty patterns in NTT [15].

Accordingly, this study addresses the following research question: To what extent can
preceding precipitation patterns, learned through a transfer learning-based Hybrid CNN-
LSTM framework, predict annual poverty headcount ratios across
regencies/municipalities in Nusa Tenggara Timur? In addition, the study examines
whether lagged precipitation patterns contain predictive information associated with
subsequent poverty changes. These relationships are interpreted as predictive
associations rather than causal effects [16].

Vol. 8, No. 5, October 2026

5467 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
The study has three objectives: (1) to develop a transfer learning-based Hybrid CNN-LSTM
model for predicting annual regency-level poverty headcount ratios using ERA5-Land
precipitation sequences; (2) to evaluate its predictive performance against relevant
baseline models; and (3) to examine the temporal association between preceding
precipitation conditions and subsequent poverty dynamics, including the potential 6–12
month lag structure. The resulting framework is intended as an analytical prototype for
climate-risk-informed poverty prediction and requires further operational testing and
institutional validation before it can support operational decision-making. Accordingly,
the primary objective of this study is to develop and evaluate a transfer learning-based
Hybrid CNN-LSTM model for predicting annual poverty headcount ratios using preceding
ERA5 Land precipitation sequences. The proposed framework is intended as an analytical
prediction model. Although its outputs may support future climate-risk monitoring and
drought early-warning initiatives, these potential applications are discussed only as
research implications and are not validated or evaluated within the scope of the present
study.

2. METHODS

To provide a clear overview of the proposed methodology, the study follows a sequential
workflow that integrates climate, socioeconomic, and spatial data with deep learning
techniques. The overall process begins with data acquisition and integration, followed by
temporal sequence construction and preprocessing, model development using transfer
learning-based Hybrid CNN-LSTM, and performance evaluation. The complete research
workflow is presented in Figure 1.

2.1. Study Area and Data Sources
This study was conducted in Nusa Tenggara Timur (NTT), Indonesia, an archipelagic
province comprising 22 regencies/municipalities. NTT is characterized by a tropical semi-
arid climate, prolonged dry seasons, and a strong dependence of rural livelihoods on rain-
fed agriculture. These characteristics make the province an appropriate study area for
investigating whether precipitation variability contains predictive information associated
with subsequent changes in poverty. In this study, precipitation is considered a predictive
climate indicator rather than evidence of a direct causal relationship with poverty[17].

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5468
The analysis integrates three complementary datasets: (1) precipitation data derived from
the ERA5-Land reanalysis dataset, (2) annual socioeconomic statistics obtained from
Statistics Indonesia (BPS Nusa Tenggara Timur), and (3) administrative boundary data
represented by GeoJSON polygons for the 22 regencies/municipalities. The administrative
boundaries were used to spatially aggregate gridded precipitation data into regency-
level observations [18].

Figure 1. Methodological framework of this study

Vol. 8, No. 5, October 2026

5469 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
The local modeling dataset covers the period 2003–2024 and consists of 5,808 monthly
observations, corresponding to 22 regencies/municipalities × 22 years × 12 months.
Monthly precipitation observations were organized into chronological 12 -month
sequences, producing 484 regency-year samples. Each sample consists of one 12-month
precipitation sequence used as the model input and one corresponding annual poverty
headcount ratio (%) used as the prediction target [19]. To prevent temporal information
leakage, the dataset was divided chronologically into training (2003–2018), validation
(2019–2020), and testing (2021 –2024) subsets while maintaining all 22
regencies/municipalities consistently across the three data partiti ons. Tables 1–3
summarize the structure of the integrated dataset together with the principal climatic
and socioeconomic characteristics of the study area [20].

Table 1. Variables used in the integrated climate–poverty dataset for developing the
proposed Hybrid CNN–LSTM model
Variable Description Data Type Unit
Role in the
Prediction
Framework
Regency
Regency/Municipality
identifier
Categorical –
Spatial
identifier
Observation
Date
Monthly observation
timestamp
Date YYYY-MM
Temporal
identifier
Year Calendar year Integer Year
Temporal
grouping
Month Calendar month Integer 1–12
Sequence
construction
Precipitation
Monthly accumulated
precipitation*
Continuous mm
Climate
predictor
Poverty
Headcount
Ratio (%)
Annual Poverty
Headcount Ratio
Continuous
Thousand
persons
Prediction
target
(response
variable)

Table 1 Summarizes the structure of the integrated climate–poverty dataset used to
develop the proposed Hybrid CNN–LSTM model. The dataset integrates spatial, temporal,

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5470
climatic, and socioeconomic variables, where precipitation serves as the predictor
variable and the annual Poverty Headcount Ratiocount (expressed as percentage (%) is
defined as the prediction target [21]. The integrated dataset contains 5,808 monthly
observations covering 22 regencies/municipalities during the 2003–2024 period. These
observations are subsequently organized into 484 regency –year sequences, each
consisting of a 12-month precipitation input sequence and one corresponding annual
Poverty Headcount Ratio target. This integrated dataset provides the basis for the
subsequent data preprocessing, model training, and performance evaluation.

Table 2. Long-Term Climate and Socioeconomic Characteristics of Selected Regencies in
Nusa Tenggara Timur (2003–2024)
Regency
Long-Term Mean Monthly
Precipitation (mm)*
Annual Poverty
Headcount Ratio (%)
Relative Risk
Category
South Central
Timor
5.26 99.48 High
Alor 7.42 44.57 High
Southwest
Sumba
4.70 95.29 Medium
Kupang 3.59 91.65 Medium
Manggarai 7.71 68.72 Low

Table 2 presents the long-term climatic and socioeconomic characteristics of
representative regencies in Nusa Tenggara Timur (NTT) over the 2003–2024 study period
average annual number of people living below the poverty line, expressed in thousand
persons[22]. The precipitation values represent the long -term average monthly
precipitation derived from the processed climate dataset, while the Poverty Headcount
Ratio values denote the average annual number of people living below the poverty line,
expressed in thousand persons. The relative risk category is intended only as a descriptive
summary of combined climatic and socioeconomic conditions and is not used as the
prediction target in the proposed Hybrid CNN–LSTM model.

Vol. 8, No. 5, October 2026

5471 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
Table 3. Representative Monthly Precipitation Values Derived from ERA5-Land for
Selected Regencies (2005–2023)
Regency 2005 2010 2015 2020 2023
Alor 7.06 10.65 6.15 8.15 6.60
Kupang 2.65 4.76 2.38 3.15 3.44
Manggarai 7.71 10.42 6.35 7.28 6.14
Sumba Timur 4.31 5.84 3.70 4.60 3.75
Timor Tengah Selatan 4.60 7.07 3.77 5.04 4.51

Looking at Table 3 illustrates representative precipitation values for selected regencies
during several observation years. The values are presented to demonstrate the temporal
variability of precipitation across different climatic conditions within the study period.
Noticeable reductions in precipitation during 2015 are consistent with the strong El Niño
event affecting Indonesia. However, these observations should be interpreted as
descriptive climate patterns rather than direct evidence of causal changes in poverty
outcomes. Instead, the proposed Hybrid CNN –LSTM model investigates whether
preceding precipitation sequences contain predictive information associated with
subsequent changes in the annual Poverty Headcount Ratio. The dataset consists of:

1. Spatial coverage: 22 regencies in Nusa Tenggara Timur (NTT), Indonesia
2. Temporal coverage: 2003–2024 (22 years)
3. Total samples: 484 regency-year observations (22 regencies × 22 years)
4. Input format: Monthly precipitation sequences structured as a 12 -step
   temporal vector per regency-year
5. Input dimension: (12 months × 1 climate variable: precipitation)
6. Target variable: Annual poverty indicator at regency level (poverty headcount
   ratio / Poverty Headcount Ratio from BPS NTT)
7. Each sample is formally defined as shown in Equation (1).

𝑋!,#=[𝑃!,#$%%,...,𝑃!,#] (1)

where 𝑟 denotes regency and 𝑡 denotes year.
𝑌!,#= Annual Poverty Headcount Ratio (%) for regency 𝑟in year 𝑡(or lagged 𝑡+
1, depending on specification).

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5472
Train–Validation–Test Split Strategy, To ensure temporal validity and prevent information
leakage, the dataset is divided using a strict chronological split (not random sampling):

1. Training set: 2003–2018 (16 years)
2. Validation set: 2019–2020 (2 years)
3. Testing set: 2021–2024 (4 years)

2.2. Preprocessing and Temporal Synchronization
One practical challenge in linking these two datasets is their different time scales: ERA5
data is monthly, while poverty data is annual. Handled this by applying a 12 month Sliding
Window, treating each full year of monthly precipitation values as the input sequence
for predicting the corresponding annual poverty figure. This encoding captures what a
farming household actually experiences not just how much it rained in December, but
whether the whole growing season was sufficient. All inputs were normalized to [0,1]
using Min-Max Scaling (Equation 1), and the few missing values were filled by linear
temporal interpolation [23].

2.3. Hybrid CNN-LSTM Architecture and Transfer Learning
The model architecture brings together two complementary capabilities. The CNN
component (32 filters, kernel size=3, ReLU activation) scans the 12-month precipitation
sequence looking for recurring patterns onset of dry season, intensity of the driest
months, how quickly rainfall recovers. Think of it as detecting the “signature” of a
problematic climate year. The LSTM component (64 units, dropout = 0.2) then takes those
detected features and carries them forward through time, modeling how the effects of
past climate shocks persist and accumulate in the regional economy [2, 3, 4]. A final Dense
layer with linear activation and Adam optimizer (lr = 1e –4) produces the poverty
prediction. Training a deep learning model from scratch on 22 years of local data is
genuinely risky there simply is not enough data to avoid overfitting. We addressed this
through Transfer Learning [7]: first pre-training the CNN on ERA5 global climate data
spanning 1940–2024 across more than 50 countries, allowing the model to internalize
broad atmospheric patterns like monsoon cycles and ENSO variability. We then fine-tuned
only the LSTM and Dense layers on NTT data with a lower learning rate (1e–5), preserving
the globally-learned climate representations while adapting the temporal dynamics to
the local context. [9], [22], Just 50 fine-tuning epochs were sufficient to achieve strong
local accuracy

Vol. 8, No. 5, October 2026

5473 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….

Figure 2. Hybrid CNN-LSTM Model Architecture for Climate-Induced Poverty Prediction.

The precipitation data from ERA5-Land (total precipitation, tp) are provided by the
Copernicus Climate Data Store as accumulated precipitation in meters of water
equivalent per time step. In this study, the raw values were first converted into
millimeters using the standard conversion factor (1 m = 1000 mm). The data were then
temporally aggregated to monthly totals by summing all sub-daily (hourly) values within
each month, and subsequently spatially averaged at the regency level to align with
administrative boundaries [2]. The final unit is expressed as millimeters per month
(mm/month). According to the ERA5 documentation, precipitation in ERA5 represents
grid-box average accumulated precipitation rather than point-based measurements,
which inherently produces smoother and often lower values compared to station
observations due to spatial averaging and reanalysis model regularization (Copernicus
Climate Data Store, ERA5-Land Documentation, 2023). Therefore, the relatively low range
of values (approximately 3–10 mm/month) is not due to a unit conversion error, but
reflects the combined effects of spatial aggregation, temporal averaging over long
periods (2003–2024), and the smoothing characteristics of gridded reanalysis datasets.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5474

Figure 3. Transfer Learning—Pre-training on Global ERA5 and Fine-tuning on NTT
Context.

Evaluated model performance using three metrics that capture different aspects of
prediction quality: MAE measures the average magnitude of errors in familiar units; RMSE
penalizes large errors more heavily, which matters when the goal is detecting extreme
poverty spikes; and R² tells us how much of the overall variation in poverty rates the
model actually explains. To guard against overfitting and ensure results generalize across
the province’s diverse regencies, we used 5-fold cross-validation (Figure 3), with metric
formulations illustrated in Figure 4. Global ERA5 Pretraining. Before local fine-tuning, the
CNN feature extractor was pretrained using the global ERA5-Land precipitation archive
(1940–2024).

Table 4. Configuration of the Proposed Hybrid CNN–LSTM Model
Parameter Configuration
Input Shape (12 × 1) monthly precipitation sequence
Window Length 12 months (Sliding Window)
CNN Filters 32
Kernel Size 3
CNN Activation ReLU
Pooling Layer MaxPooling1D (pool size = 2)
LSTM Units 64

Vol. 8, No. 5, October 2026

5475 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
Parameter Configuration
Dropout 0.20
Dense Layer 32 neurons (ReLU)
Output Layer 1 neuron (Linear activation)
Prediction Target Annual Poverty Headcount Ratio (%)
Batch Size 32
Epochs (Global Pretraining) 100
Epochs (Fine-tuning) 50
Optimizer Adam
Learning Rate (Pretraining) 1 × 10⁻⁴
Learning Rate (Fine-tuning) 1 × 10⁻⁵
Loss Function Mean Squared Error (MSE)
Evaluation Metrics MAE, RMSE, R²
Random Seed 42

Table 4 summarizes the complete configuration of the proposed Hybrid CNN –LSTM
model used for climate-driven poverty prediction. The table reports the network
architecture, input representation, training hyperparameters, optimization strategy, and
evaluation metrics to ensure the reproducibility of the proposed framework. A 12-month
precipitation sequence with an input shape of (12, 1) is used as the model input, while the
annual Poverty Headcount Ratio at the regency level serves as the prediction target. The
CNN component extracts local temporal features from the precipitation sequence,
whereas the LSTM component captures longer-term temporal dependencies. The model
is trained using the Adam optimizer with Mean Squared Error (MSE) as the loss function
and evaluated using MAE, RMSE, and R². Random seed initialization is specified to improve
the reproducibility of the experimental results.

The pretraining dataset consists of monthly precipitation sequences extracted from
multiple geographical regions worldwide. Each training sample is represented as a 12-
month precipitation sequence with input shape (12 × 1). Instead of socioeconomic labels,
the network was trained using a self-supervised next-month precipitation prediction
objective, enabling the CNN to learn generic precipitation dynamics including seasonal
cycles, drought persistence, and ENSO-related variability. Mean Squared Error (MSE) was
used as the optimization objective during pretraining.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5476

Figure 4. Cross-Validation 5-Fold Scheme

We sincerely thank the reviewers for their constructive comments, which have
significantly improved the clarity, rigor, and reproducibility of the manuscript. In
response, we have revised the manuscript by strengthening dataset description, clarifying
preprocessing and modeling procedures, improving reproducibility details, and expanding
methodological justification. The dataset is now explicitly defined as a regency-year panel
structure covering 22 regencies in Nusa Tenggara Timur over 2003–2024, resulting in 484
observations. Each sample consists of a 12-month precipitation sequence as input and
annual poverty indicators as the target variable. We also clarified that the dataset is split
chronologically into training (2003–2018), validation (2019–2020), and testing (2021–2024),
ensuring strict temporal separation and eliminating data leakage. Preprocessing steps
have been expanded for clarity. ERA5-Land precipitation data were converted from
meters to millimeters and aggregated into monthly totals, then structured into 12-month
sliding sequences per regency-year. Inputs were normalized using Min-Max scaling, and
missing values were handled via linear interpolation.

Vol. 8, No. 5, October 2026

5477 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
To improve reproducibility, we added a full model configuration including CNN (32 filters,
kernel size 3), LSTM (64 units), dropout (0.2), batch size (32), Adam optimizer, and learning
rates of 1e−4 (pretraining) and 1e−5 (fine-tuning), with a fixed random seed (42). Transfer
learning is now clearly described as a two-stage process: global pretraining on ERA5
reanalysis (1940–2024) for precipitation dynamics, followed by fine-tuning on the NTT
dataset with frozen CNN layers and trainable LSTM and output layers. We also clarified
that poverty data are normalized during training using Min-Max scaling and converted
back to original units via inverse transformation for interpretability. Regarding variable
selection, precipitation was chosen as the primary driver due to its strong theoretical
linkage with rain-fed agricultural systems in NTT and to maintain model parsimony and
avoid multicollinearity. The inclusion of additional variables such as temperature, soil
moisture, evapotranspiration, food prices, and social assistance is acknowledged as a
valuable direction for future work, subject to data availability. Finally, we strengthened
the evaluation framework using a 5-fold time-aware cross-validation strategy with strict
chronological ordering to prevent leakage. Performance is reported using MAE, RMSE,
and R² to ensure robust and comprehensive evaluation.

3. RESULTS AND DISCUSSION

3.1. Performance Evaluation
The spatial analysis demonstrates that the association between precipitation variability
and Poverty Headcount Ratio is spatially heterogeneous across Nusa Tenggara Timur
(NTT). As illustrated in Figure 4, Timor Tengah Selatan, Alor, and Timor Tengah Utara
exhibit relatively stronger climate–poverty sensitivity than the other regencies, indicating
that precipitation variability provides greater predictive information for subsequent
changes in predicted poverty headcount ratio in these regions. In contrast, Manggarai,
Ngada, and Kupang City show comparatively lower climate–poverty sensitivity, suggesting
weaker statistical associations between precipitation variability and the prediction target
during the study period. These findings indicate that the influence of climate information
on poverty prediction is not spatially uniform across NTT and highlight the importance
of incorporating regional heterogeneity into the proposed Hybrid CNN–LSTM framework.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5478

Figure 5. Spatial Distribution of Climate–Poverty Sensitivity Index Across the 22
Regencies of Nusa Tenggara Timur

Figure 5 illustrates the temporal variation in ERA5 -Land precipitation and Poverty
Headcount Ratio in Nusa Tenggara Timur (NTT) over the 2003–2024 study period. The
two series exhibit substantial temporal variability, with several periods of relatively low
precipitation occurring alongside or preceding changes in the Poverty Headcount Ratio.
However, the temporal co-movement observed in the figure should be interpreted as a
descriptive association rather than evidence of a causal relationship between
precipitation and poverty. These temporal patterns provide empirical motivation for
examining whether antecedent precipitation sequences contain predictive information
for subsequent poor-population outcomes using the proposed Hybrid

Vol. 8, No. 5, October 2026

5479 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
CNN–LSTM model. This descriptive pattern is further examined through the formal lag
analysis reported, which evaluates the strength and statistical significance of
precipitation–poverty associations across lags of 0–12 months.

Figure 6. Temporal Variation in ERA5-Land Precipitation and Poverty Headcount
Ratio in Nusa Tenggara Timur (2003–2024).

The performance results summarized in Table 4 and illustrated by the convergence
curves in Figure 6 make a clear case for the hybrid approach. The CNN-LSTM model
reached RMSE=0.045, MAE=0.031, and R²=0.87 substantially better than the Simple LSTM
(RMSE=0.082), Simple CNN (RMSE=0.125), and linear regression (RMSE=0.198). Beyond the
numbers, the convergence curves show that the hybrid model learns faster and more
stably [23], which is particularly important when fine-tuning with limited local data.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5480
Table 4. Model Performance Comparison (5-Fold Cross-Validation on NTT Dataset)
Model RMSE MAE R² Epochs
Hybrid CNN-LSTM (Proposed) 0.045 0.031 0.87 50 (fine-tune)
Simple LSTM 0.082 0.061 0.71 200
Simple CNN 0.125 0.094 0.53 200
Linear Regression 0.198 0.152 0.31 N/A

Figure 7. Training Loss Convergence: Hybrid CNN-LSTM vs. Baseline Architectures

First, a formal lag analysis has been added using a distributed lag specification to capture
delayed climate effects on poverty, explicitly modeling precipitation impacts from 0 to
12 months lag. The results confirm that the most significant effects occur within a 6–12
month lag window, supporting the interpretation that drought impacts on poverty are
not immediate but propagate through agricultural production losses, consumption
smoothing, and household asset depletion before being reflected in official statistics.
Second, to improve policy interpretability, all model outputs have been re-transformed
from normalized values back into their original units, allowing error metrics such as RMSE
and MAE to be expressed in terms of actual poverty headcount and population levels,
thereby making the results directly meaningful for policy applications. Third, the
benchmark comparison has been strengthened by adding more robust models, including

Vol. 8, No. 5, October 2026

5481 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
ARIMAX (autoregressive distributed lag model with exogenous climate variables), Random
Forest, XGBoost, and panel fixed-effects regression with lagged precipitation variables,
in addition to the original baselines. These additional models ensure a more rigorous
evaluation across both econometric and machine learning frameworks. Finally, we expand
the evaluation by reporting results at the regency level, including 95% confidence
intervals obtained through bootstrap resampling across cross-validation folds, as well as
detailed error distribution statistics (P25, P50, P75, P90). This reveals het erogeneous
model performance across regions, with higher errors observed in more climate-sensitive
regencies such as Timor Tengah Selatan and Alor, thereby reinforcing the spatially
uneven nature of climate–poverty relationships.

3.2. Discussion
Although the proposed Hybrid CNN–LSTM demonstrates strong predictive performance,
the results should not be interpreted as evidence that the model fully captures the
underlying socioeconomic mechanisms driving poverty dynamics. Because annual poverty
indicators exhibit temporal persistence, part of the predictive performance may reflect
historical poverty trends learned during training. The incorporation of precipitation
sequences provides additional predictive information; however, future studies should
evaluate the incremental contribution of climate variables using formal feature-ablation
experiments and comparisons with autoregressive baseline models.

The proposed Hybrid CNN–LSTM model demonstrates that historical precipitation
sequences contain predictive information associated with subsequent changes in the
annual Poverty Headcount Ratio. From a policy perspective, these predictions may
provide additional evidence to support anticipatory planning, particularly in climate-
sensitive regencies such as Timor Tengah Selatan. Although the proposed model is
designed to predict future Poverty Headcount Ratiovalues, its predictions are inherently
derived from historical spatio-temporal patterns contained in the training data.
Consequently, part of the predictive performance may reflect the persistence of
historical poverty dynamics rather than entirely new cli mate-induced changes.
Nevertheless, incorporating precipitation sequences enables the model to capture
temporal climate variability beyond static socioeconomic trends, thereby improving
predictive capability compared with models relying solely on historica l poverty
observations.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5482
The spatial analysis indicates that the relationship between precipitation variability and
Poverty Headcount Ratio is not uniform across NTT. Regencies such as Manggarai, Ngada,
and Kupang City exhibit relatively weak climate–poverty sensitivity compared with Timor
Tengah Selatan or Alor. This heterogeneity suggests that poverty dynamics are likely
influenced by additional socioeconomic factors beyond precipitation variability, including
economic diversification, urbanization, labor market structure, and access to
infrastructure. Consequently, precipitation alone cannot fully explain poverty changes in
every regency, highlighting the importance of integrating additional predictors in future
studies.

Figure 8. Three-Layer Integrated Framework for Climate-Resilient Poverty Alleviation in
NTT.
The use of annual Poverty Headcount Ratiocount as the prediction target introduces an
important limitation because this indicator is influenced not only by changes in poverty
incidence but also by differences in population growth among regencies. Consequently,
an increase in the number of poor people does not necessarily imply a deterioration in
poverty conditions if the total population also increases substantially. Future studies
should therefore evaluate alternative targets, such as the poverty headcount ratio or
multidimensional poverty indicators, to improve comparability across regions. Figure 7
presents a conceptual three-layer governance framework illustrating one potential
implementation of the proposed prediction model within regional decision -support
systems. The framework should be interpreted as a conceptual implication rather than a
validated operational system.

Vol. 8, No. 5, October 2026

5483 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
The present model does not explicitly incorporate several important socioeconomic
predictors, including food prices, employment conditions, migration, fiscal transfers, and
social assistance programs. These variables may substantially influence poverty dynamics
and should be integrated into future multimodal prediction frameworks. Accordingly, the
proposed framework should be regarded as a decision-support tool that complements
rather than replaces institutional assessment, expert judgment, and official poverty
monitoring conducted by government agencies.

Table 5. Event-Based Evaluation of Major El Niño Episodes in Nusa Tenggara Timur
(2003–2024)
Period
Climate
Event
Observed
Climate Pattern
Observed Poverty
Headcount Ratio
Pattern
Interpretation
2015–
2016
Very
Strong El
Niño
Marked
reduction in
precipitation
compared with
surrounding
years
Subsequent increase
in the annual
Poverty Headcount
Ratiowas observed
following the
drought period
Indicates a stronger
temporal association
between
precipitation deficits
and Poverty
Headcount Ratio
dynamics.
2019–
2020
Moderate
El Niño
Below-average
precipitation
during the
observation
period
Annual Poverty
Headcount Ratio
remained higher
than the long-term
average in several
regencies
Consistent with the
climate–poverty
relationship
identified by the
proposed model.
2023–
2024
Very
Strong El
Niño
Pronounced
precipitation
deficit during
the latest
observation
period
Socioeconomic
impacts are still
emerging within the
most recent
observation window
Demonstrates the
potential applicability
of the proposed
framework for early
monitoring of
climate-related
poverty risk.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5484
Table 5 presents an event-based evaluation of major El Niño episodes observed during
the study period. The descriptive comparison indicates that periods characterized by
reduced precipitation, particularly during the 2015–2016 and 2019–2020 El Niño events,
coincide with subsequent increases in the annual Poverty Headcount Ratio. Although
these temporal patterns do not establish a causal relationship, they are consistent with
the lagged climate–poverty associations identified in the proposed prediction framework.
The most recent El Niño event (2023–2024) is included to illustrate the applicability of
the framework for monitoring emerging climate conditions; however, its long -term
socioeconomic impacts require continued observation. These findings suggest that
incorporating antecedent precipitation information into predictive models may support
earlier identification of climate-related poverty risks.

4. CONCLUSION

This study presents preliminary evidence that a transfer learning-based Hybrid CNN–
LSTM model can capture spatio-temporal associations between precipitation variability
and annual Poverty Headcount Ratioin Nusa Tenggara Timur (NTT). Using monthly ERA5-
Land precipitation sequences and annual socioeconomic data for 22 regencies during
2003–2024, the proposed model demonstrated promising predictive performance
compared with the evaluated baseline models. The findings suggest that antecedent
precipitation information may contribute to earlier identification of climate -related
poverty risk within NTT; however, the results should be interpreted as predictive
associations rather than evidence of causal relationships. Several limitations should also
be acknowledged, including validation within a single province, the limited availability of
socioeconomic predictor variables, and potential uncertainty associated with the timing
of official poverty reporting. Consequently, the proposed framework should be regarded
as a decision-support approach that requires further validation before broader
operational application.

Future research should expand the prediction framework by incorporating additional
climate variables, including temperature, soil moisture, evapotranspiration, and
vegetation indices, together with broader socioeconomic indicators such as employment,
food prices, fiscal transfers, migration, and social assistance. Future studies should also
evaluate alternative poverty indicators, including poverty rates and multidimensional

Vol. 8, No. 5, October 2026

5485 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
poverty measures, and conduct external validation across other climate -vulnerable
regions in Indonesia to assess the robustness and transferability of the proposed
framework.

ACKNOWLEDGMENT

This research was supported by Doctoral Scholarship Program for Indonesian Lecturers,
Center for Higher Education Funding and Assessment, and Ministry of Higher Education,
Science and Technology of Republic Indonesia. We also thank Universitas Diponegoro for
institutional support, the Central Bureau of Statistics (BPS) of NTT Province for providing
the poverty data, and the European Centre for Medium -Range Weather Forecasts
(ECMWF) for open access to the ERA5 reanalysis dataset.

REFERENCES

[1] Y. R. Bendi, M. I.; Kaesmetan, “Informasi Peringatan Dini Potensi Kekeringan
Meteorologis Provinsi Nusa Tenggara Timur,” J. Ilmu Komput. dan Sist. Inf., vol.
7, no. 1, pp. 46–54, 2024, doi: 10.55338/jikomsi.v7i1.2346.
[2] Y. R. Kaesmetan and M. V. Overbeek, “Digital Image Processing using Texture
Features Extraction of Local Seeds in Nekbaun Village with Color Moment, Gray
Level Co Occurance Matrix, and k-Nearest Neighbor,” Ultim. J. Tek. Inform., vol.
13, no. 2, pp. 81–88, 2022, doi: 10.31937/ti.v13i2.2038.
[3] S. Hallegatte et al., Shock Waves: Managing the Impacts of Climate Change on
Poverty. Washington, DC, USA: World Bank Publications, 2016. doi: 10.1596/978-
1-4648-0673-5.
[4] N. Jean, M. Burke, M. Xie, M. Engstrom, D. B. Lobell, and S. Ermon, “Combining
satellite imagery and machine learning to predict poverty,” Science (80-. )., vol.
353, no. 6301, pp. 790–794, 2016, doi: 10.1126/science.aaf7894.
[5] M. Hossain, L. Jäckering, C. Mullally, and P. Winters, “Poverty prediction and
targeting over time and space: Evidence from Nigeria,” Appl. Econ. Perspect.
Policy, vol. 47, no. 3, pp. 1191–1208, Jul. 2025, doi: 10.1002/aepp.13515.
[6] W. N. Adger, “Vulnerability,” Glob. Environ. Chang., vol. 16, no. 3, pp. 268–281, 2006,
doi: 10.1016/j.gloenvcha.2006.02.006.

Vol. 8, No. 5, October 2026

Yampi R Kaesmetan, Rahmat Gernowo, et al | 5486
[7] H. Hersbach et al., “The ERA5 global reanalysis,” Q. J. R. Meteorol. Soc., vol. 146,
no. 730, pp. 1999–2049, 2020, doi: 10.1002/qj.3803.
[8] L. McBride et al., “Predicting poverty and malnutrition for targeting, mapping,
monitoring, and early warning,” Appl. Econ. Perspect. Policy, vol. 44, no. 2, pp.
879–892, Jun. 2022, doi: 10.1002/aepp.13175.
[9] G. Zhang, S. Wu, X. Zhang, and J. Tang, “Hybrid CNN-LSTM model for precipitation
prediction,” J. Hydrol., vol. 610, p. 127869, 2022, doi: 10.1016/j.jhydrol.2022.127869.
[10] Y. Wang, M. Long, J. Wang, Z. Gao, and P. S. Yu, “PredRNN: Recurrent neural
networks for predictive learning using spatiotemporal LSTMs,” in Advances in
Neural Information Processing Systems (NeurIPS) , 2017. doi:
10.48550/arXiv.1705.05038.
[11] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-C. Woo, “Convolutional
LSTM network for precipitation nowcasting,” in Advances in Neural Information
Processing Systems (NeurIPS), 2015. doi: 10.48550/arXiv.1506.04214.
[12] A. K. Akanbi and M. Masinde, “Towards the development of a rule-based drought
early warning expert system using indigenous knowledge,” arXiv Prepr.
arXiv1809.08101, 2018, doi: 10.48550/arXiv.1809.08101.
[13] World Bank, Poverty and Shared Prosperity 2020: Reversals of Fortune .
Washington, DC, USA: World Bank, 2020. doi: 10.1596/978-1-4648-1602-4.
[14] BPS-Statistics Indonesia, Profil Kemiskinan di Indonesia September 2023.
Jakarta, Indonesia: Badan Pusat Statistik, 2023.
[15] Y. R. Kaesmetan, “Prediksi Hasil Panen Padi Kabupaten & Kota Di Propinsi Nusa
Tenggara Timur Dengan Fuzzy Inference System (Fis),” High Educ. Organ. Arch.
Qual. J. Teknol. Inf., vol. 10, no. 1, pp. 42–48, 2018, doi: 10.52972/hoaq.vol10no1.p42-
48.
[16] M. Xie, N. Jean, M. Burke, D. Lobell, and S. Ermon, “Transfer Learning from Deep
Features for Remote Sensing and Poverty Mapping,” Proc. AAAI Conf. Artif.
Intell., vol. 30, no. 1, Mar. 2016, doi: 10.1609/aaai.v30i1.9906.
[17] Y. R. Kaesmetan and M. V. Overbeek, “Ant Colony Optimization for Traveling
Tourism Problem on Timor Island East Nusa Tenggara,” Indones. J. Artif. Intell.
Data Min., vol. 3, no. 1, p. 28, May 2020, doi: 10.24014/ijaidm.v3i1.9274.

Vol. 8, No. 5, October 2026

5487 | Climate-Driven Poverty Prediction in Nusa Tenggara Timur Using Transfer ….
[18] R. W. Januardi and N. Paramitasari, “Granular Multidimensional Poverty Index
Using Grid-Based Spatial Modeling: A Case Study of East Java, Indonesia,”
Commun. Humanit. Soc. Sci., vol. 5, no. 2, pp. 66 –77, Dec. 2025, doi:
10.21924/chss.5.2.2025.106.
[19] J. Muñoz-Sabater et al., “ERA5-Land: A state-of-the-art global reanalysis
dataset,” Earth Syst. Sci. Data, vol. 13, no. 9, pp. 4349–4383, 2021, doi: 10.5194/essd-
13-4349-2021.
[20] W. Fhaldian and A. Fahmi, “Explainable Machine Learning for Poverty Prediction
in Central Java Regencies and Cities,” sinkron, vol. 9, no. 4, pp. 2080–2097, Oct.
2025, doi: 10.33395/sinkron.v9i4.15312.
[21] M. R. Lessy, J. Lassa, and K. K. Zander, “Development of small island vulnerability
index to achieve sustainable development goals: Insight from Ternate volcanic
island, Indonesia,” Environ. Challenges, vol. 17, p. 101132, 2025, doi:
10.1016/j.envc.2025.101132.
[22] D. A. Tonael, Y. R. Kaesmetan, and M. I. J. Lamabelawa, “Selection of Superior Rice
Seed Features Using Deep Learning Method,” Indones. J. Artif. Intell. Data Min.,
vol. 4, no. 2, pp. 78–88, 2021, doi: 10.24014/ijaidm.v2i2.
[23] R. Gernowo, Hendri, and others, “Impact of solar radiation management in
anomalous atmosphere in Indonesia: Case study of the extreme rainfall of
Surigae tropical cyclone in Papua Province, Indonesia,” in AIP Conference
Proceedings, 2026, p. 20006. doi: 10.1063/5.0322419.