Document Type : Original Research Paper
Authors
Department of Surveying Engineering, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
Abstract
Background and Objectives: Monitoring and predicting vegetation cover changes play an important role in natural resource management, assessment of climate change impacts, and ecosystem monitoring. Among remote sensing indicators, the Normalized Difference Vegetation Index (NDVI) is considered one of the most widely used indices for evaluating vegetation conditions and their temporal variations. Due to the complex temporal behavior of vegetation dynamics and the increasing need for accurate forecasting, the application of intelligent approaches and deep learning models has gained significant attention. Therefore, the main objective of this study was to develop a deep neural network model based on lagged and statistical features for predicting the NDVI time series in a part of the Hamoun wetland region located in southeastern Iran.
Methods: The data used in this study consisted of Landsat 5, 7, 8, and 9 satellite images acquired during the period from 2010 to 2026 through the Google Earth Engine platform. Images with cloud coverage lower than 20% were selected, and in cases where suitable images were unavailable, the cloud threshold was increased to 30%. After calculating NDVI, the time series data were smoothed using the Savitzky–Golay filter. Subsequently, lagged and statistical features, including Lag1, Lag2, Lag3, temporal difference, and moving average, were extracted. A deep neural network model with three hidden layers was then developed for time-series modeling and prediction. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²).
Findings: The results of the independent model evaluation using data from 2025 and 2026 showed that the proposed model achieved satisfactory performance in predicting the NDVI time series. The mean absolute error (MAE) was 0.000844, the root mean square error (RMSE) was 0.000959, and the coefficient of determination (R²) was 0.9112, indicating the model’s satisfactory ability to learn the temporal patterns of the data. The prediction results for the period 2027–2030 also showed that NDVI values in the study area exhibited interannual fluctuations rather than following a consistent increasing trend. The mean NDVI across the three seasons decreased from 0.053956 in 2027 to 0.037751 in 2030, although a relative increase was observed in 2029 compared with 2028.
Conclusion: The findings showed that deep neural networks based on temporal and statistical features have satisfactory capability for predicting remote sensing time series and analyzing vegetation changes. This approach can be effectively applied to monitoring sensitive ecosystems, natural resource management, and the assessment of environmental changes. The results can also provide a basis for developing intelligent models for predicting environmental changes in arid and semi-arid regions. The predictions for 2027–2030 indicated interannual fluctuations in NDVI and a decrease in its mean value toward the end of the forecast period. However, these values should be interpreted as model-based estimates because actual observational data for the future years were not available at the time of the study. The main limitation of the study was the use of mean NDVI and the absence of direct climatic and hydrological variables in the model, which can be addressed in future studies.
Keywords
Main Subjects
COPYRIGHTS
© 2026 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
)