افزایش دقت داده‌های بارش ماهواره‌ای: روش ریزمقیاس سازی داده‌های PERSIANN با استفاده از NDVI، LST و DEM

نوع مقاله : مقاله پژوهشی

نویسنده

• زهرا شیرمحمدی علی اکبرخانی

استادیار گروه مهندسی آب، مجتمع آموزش عالی کشاورزی و دامپروری تربت‌جام، ایران

چکیده

کلیدواژه‌ها

• افزایش وضوع مکانی
• بارش
• داده‌های ماهواره پرشین
• یادگیری ماشین

موضوعات

• هیدرولوژی

عنوان مقاله [English]

Improving Precipitation Accuracy: A Rescaling Method for PERSIANN Using NDVI, LST, and DEM Data

نویسنده [English]

• Zahra Shirmohammadi -Aliakbarkhani

کلیدواژه‌ها [English]

• Machine Learning
• PERSIANN-CDR
• Precipitation
• Spatial Resolution Enhancement

Introduction

Precipitation plays a critical role in the global water cycle and in processes involving material and energy exchanges. High-resolution precipitation data are essential for accurate hydrological, meteorological, and ecological studies, especially at regional scales. Traditionally, rain gauges provide precise point data, but their irregular distribution and the high spatial-temporal variability of precipitation make them insufficient for generating fine-resolution datasets. Satellite observations offer valuable insights into water and energy exchanges between land and atmosphere. They are particularly effective in estimating precipitation across vast areas, including mountainous and sparsely gauged regions. For example, remotely sensed data such as the PERSIANN-CDR product are widely used in hydrological and meteorological studies. However, the 27 km spatial resolution of PERSIANN data often limits its applicability in detailed hydrological simulations and environmental assessments at local scales. To address this limitation, this study introduces a novel approach by combining land surface temperature (LST), normalized difference vegetation index (NDVI), and digital elevation model (DEM) data with the Random Forest (RF) algorithm to downscale annual PERSIANN precipitation data. The study focuses on southeastern Khorasan Razavi from 2000 to 2023—a region with limited ground-based observations and complex topography. To the best of our knowledge, this combination of variables and techniques has not been applied in this region before. The outcomes of this research have significant practical implications. High-resolution precipitation maps generated by this method can improve water resource management, enhance flood prediction accuracy, and support sustainable development in arid and semi-arid regions. By addressing the limitations of traditional precipitation datasets, this study provides a foundation for more reliable hydrological and environmental analyses.

Method

Data Collection

 This study utilizes data from the Torbat Jam rain gauge stations over a 23-year period (2001–2023). The time period of 2001-2023 was chosen due to the availability of continuous data and to capture long-term climatic trends. While some data gaps and noise were present, they were addressed through preprocessing techniques to ensure the robustness of the analysis. The PERSIANN dataset, accessible on Google Earth Engine as ee. ImageCollection("NOAA/PERSIANN-CDR"), provided precipitation data for the period from 2001 to 2023. To obtain NDVI data, MOD13Q1 data (16-day NDVI at 250-meter resolution) was accessed through Google Earth Engine under the code ee. ImageCollection("MODIS/061/MOD13Q1"). For land surface temperature, the study used MOD11A2 data, which includes 8-day LST at a 1 km resolution, available since 2000 as ee. ImageCollection("MODIS/061/MOD11A2"). Elevation data was sourced from the SRTM version 3 (SRTM Plus) product, with an approximate 30-meter accuracy, accessible via ee. Image("USGS/SRTMGL1_003").

Preprocessing

 The study focuses on downscaling PERSIANN satellite precipitation data, originally at a 27 km resolution, to a finer 1 km resolution. This scaling was achieved using NDVI (Normalized Difference Vegetation Index), LST (Land Surface Temperature), and DEM (Digital Elevation Model) data, applying the Random Forest (RF) machine learning algorithm.

Modeling

The Random Forest algorithm implemented in Google Earth Engine was chosen for its robustness and ability to handle complex datasets. This platform provides optimized parameters for the algorithm, making it efficient for large-scale environmental data processing.

Residual Correction

To correct the residuals, the difference between the actual precipitation values (ground data) and the predicted values from the microscaling model (using NDVI, LST, and DEM data) was first calculated. Next, a machine learning method called Random Forest was applied to model these residuals, as it is capable of identifying and modeling complex patterns in the residual data. Finally, the residuals predicted by the machine learning models were added to the initial predicted values to obtain the adjusted predictions. The residual correction technique offers significant advantages:

Enhanced Model Accuracy

 The residual correction technique offers significant advantages: by addressing the differences (residuals) between observed and predicted values, the model's accuracy is significantly improved. This method reduces both systematic and random errors in the initial predictions, leading to more reliable results. Additionally, the Random Forest algorithm, used for residual correction, is capable of capturing complex relationships within the residuals that simpler models might miss.

Evaluation

The model’s performance, before and after the modification, was evaluated using root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC) metrics. The residual correction technique significantly improved model performance, highlighting its utility in enhancing prediction accuracy. Specifically, it led to reductions in RMSE and MAE, and increases in CC, demonstrating the method's effectiveness in refining precipitation predictions.

Results

Statistical analyses indicate that precipitation downscaling yielded favorable results at Firouzkouh and Robat Samanegan stations on monthly and annual scales, and at Torbat-e Jam station on an annual scale. The limited improvement in monthly downscaling at Torbat Jam can be attributed to the high spatial-temporal variability of precipitation in this region, which makes accurate downscaling more challenging. Similarly, studies by Cho et al. (2013) on the Korean Peninsula achieved improved local precipitation accuracy by downscaling TRMM precipitation data to a resolution of 1 km, while Zhang et al. (2018) demonstrated that 1 km maps in mountainous areas offer higher precision. Research by Ghorbanpour et al. (2021) also found that finer resolutions, such as 1 km, are beneficial in arid regions like Lake Urmia, especially where ground stations are sparse. Noor et al. (2023) confirmed that using 1 km downscaling improved precipitation accuracy in downstream regions of the Indus River. In Torbat Jam, the residual correction method was applied due to the lack of significant improvement in monthly downscaling, leading to enhanced model accuracy across all stations and time scales. Overall, residual correction proved highly effective in improving downscaling models’ accuracy, especially in areas with limited data or spatial anomalies. This technique enables models to correct for initial downscaling errors, resulting in better rainfall prediction by incorporating real data. Several studies support the efficacy of residual correction with machine learning methods like Random Forest (RF) and Kriging. For example, Zhan et al. (2018) demonstrated enhanced accuracy in semi-arid regions, and Chen et al. (2020) highlighted RF’s ability to reduce errors in arid environments. Zhao (2021) showed that residual correction led to a significant reduction in MAE and RMSE errors by 19% and 21%, respectively, in mountainous areas. In this study, applying the residual correction method significantly improved prediction accuracy across all studied stations and time scales. At Torbat Jam station, monthly correlation increased by 22% and annual correlation by 34%, with RMSE and MAE reductions of 64% and 66% on a monthly scale, and 71% and 69% on an annual scale, respectively. At Firouzkouh station, monthly correlation improved by 29% and annual correlation by 35%, with RMSE reductions of 63% monthly and 69% annually, and MAE reductions of 68% monthly and 76% annually. At Robat Samanegan station, monthly correlation increased by 26% and annual correlation by 7%, with RMSE reductions of 61% monthly and 74% annually, and MAE reductions of 60% monthly and 74% annually. These results underscore the efficacy of residual correction in enhancing model accuracy and reducing prediction errors across varying temporal and spatial scales. Consequently, combining RF downscaling with residual correction provides a suitable method for achieving high-resolution precipitation data (1 km), especially when using auxiliary variables like NDVI, LST, and DEM data. Satellite data offers extensive spatial and temporal coverage, making it valuable for precipitation modeling and climate analysis. Downscaling and residual correction techniques improve model accuracy significantly, supporting more precise hydrological and climate studies.

Conclusions

High-resolution precipitation maps at 1 km provide more precise insights into precipitation distribution, allowing for the identification of local patterns, climate anomalies, and precipitation-sensitive areas. These finer maps are instrumental in water resource management, flood and drought mitigation, and agricultural and urban planning, while 27 km maps are more suitable for regional-scale analyses. Machine learning techniques, particularly when paired with residual correction, significantly enhance precipitation prediction accuracy in regions with sparse ground data or complex topographies. However, this study has certain limitations. The effectiveness of downscaling methods can be sensitive to the quality and availability of input data, such as auxiliary variables (e. g. , NDVI, LST, DEM). Additionally, computational challenges in training advanced machine learning models must be considered. Future research should focus on improving machine learning models, incorporating newer and higher-resolution satellite data, and testing these methods across diverse regions with varying climatic and topographical characteristics. Addressing these areas can further enhance the robustness and applicability of downscaling approaches for hydrological, ecological, and climate studies.

Author Contributions

The author contributed to the conceptualization of the article and writing of the original and subsequent drafts.

Data Availability Statement

Data available on request from the author.

Acknowledgements

We acknowledge the financial support received from the University of Torbat-e Jam.

Ethical considerations

The author avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.