Abstract
This study evaluated satellite-based precipitation in Bolivia using a combination framework with four equations: relative error, geometric, weighted linear, and sigmoidal. These were compared to enhance daily GSMaP_Gauge V8 estimates. Considering elevation variability, the interpolation method KED proved critical to obtain proper rainfall patterns. The conventional method OK showed weak performance indicators (R2 = 0.18, NSE = −4.19, PBias = 48.28%). In contrast, the KED approach restored spatial continuity and predictive capability nationwide (R2 = 0.71, NSE = 0.68, KGE = 0.79, PBias = 11.24%, RMSE = 1.3 mm/day). Therefore, it provided a more consistent bias correction. On the other hand, sensitivity heat maps demonstrate that iterative algorithms require strict bounding. Specifically, iterating beyond optimal thresholds (N > 6 for relative error; N > 5 for geometric; N > 2 for weighted linear and sigmoidal) triggers exponential residual amplification. The optimized framework successfully bridges the gap between rain-gauge accuracy and gridded estimates in ungauged zones. The relative error showed best performance in the Amazon region, with geometric on the Altiplano plateau, and weighted linear in La Plata basin. These products yield reliable daily precipitation estimates for hydrological modeling, water-resource management, and climate change assessments in Bolivia. Finally, this approach can also be applied in other countries, identifying proper combination methods per hydrologic units.
1. Introduction
Precipitation is a key driver in climatological and hydrological analyses. It plays a fundamental role in water-balance modeling, climate risk evaluations, and long-term trend assessments under climate change dynamics [1,2,3,4]. Consequently, capturing spatial and temporal precipitation patterns requires highly reliable observation networks [5]. Traditionally, precipitation monitoring relies on direct methods (such as rain gauges) and indirect approaches using ground-based radar and Satellite-Based Precipitation Products (SBPs). Rain gauges provide direct physical measurements at ground level. However, their point-based nature, high maintenance requirements, and sparse spatial distribution—especially in Andean terrains—limit their regional representativeness [5]. Conversely, SBPs offer continuous spatial coverage. Nevertheless, their estimations involve inherent uncertainties, warm-cloud convective biases, and severe topographically induced errors [6,7].
To mitigate SBP limitations, various bias correction and data fusion techniques have been developed. These range from non-parametric statistical methods, such as Quantile Delta Mapping (QDM) [8], to advanced machine learning algorithms. Specifically, algorithms including Random Forest (RF) and K-Nearest Neighbors (KNN) have shown success in complex environments such as the Malay Peninsula [9,10]. Additionally, integrating gauge data into spatial interpolation frameworks serves as a primary correction strategy. Traditional techniques like Ordinary Kriging (OK) rely solely on spatial autocorrelation [11,12]. However, incorporating spatial covariates such as Digital Elevation Models (DEMs) through Kriging with External Drift (KED) significantly improves spatial interpolation accuracy. In doing so, KED explicitly accounts for orographic precipitation forcing across mountain barriers [13,14].
Bolivia exhibits extreme topographic and climatic gradients. Specifically, the region transitions rapidly from the high-altitude Andean Altiplano (>4000 m a.s.l.) to tropical lowlands. These low-lying areas span the Amazon and La Plata basins (<200 m a.s.l.). Consequently, such severe orographic constraints make the region a demanding natural laboratory for evaluating and refining satellite precipitation algorithms.
In Bolivia, previous data-merging efforts adopted an iterative combination methodology based on at the sub-basin scale. Specifically, this framework merged interpolated rain-gauge observations with satellite products. Initially developed for prioritized basins on a monthly scale, the approach was later extended to daily- [15] and hourly-reference precipitation products [16]. However, despite its operational adoption, this methodology presents key constraints. First, it displays a systematic tendency toward precipitation underestimation compared to ground gauges. Second, its performance is highly sensitive to sub-basin delineation and station density. For instance, using overly coarse sub-basin configurations—such as nine sub-basins in the Choqueyapu basin—degrades the spatial quality of the resulting precipitation fields [16].
Beyond sub-basin geometry, a critical limitation of reported iterative combination frameworks is the lack of controlled stopping criteria. Specifically, unbounded or excessive iterations trigger mathematical divergence, spatial overfitting, and severe noise amplification. This instability is particularly severe over semi-arid, high-elevation domains (e.g., the Altiplano) and highly convective tropical lowlands (e.g., the Amazon basin). Furthermore, systematic comparative research evaluating different mathematical formulations remains lacking. In particular, limited work assesses how relative error, geometric, weighted linear, and sigmoidal formulations [17,18,19] perform under strict iteration thresholds across contrasting physiographic regimes.
To address these challenges, this study proposes an iteration method incorporating distinct mathematical equations. Specifically, this approach aims to optimize daily satellite-based precipitation estimates across Bolivia. Accordingly, this research pursues three specific objectives:
- Evaluate interpolation methods: Ordinary Kriging (OK) and Kriging with External Drift (KED), considering the topographic relief of Bolivia.
- Determine the optimal blending formulation among equations: relative error, geometric, weighted linear, and sigmoidal at different scales.
- Analyze iteration thresholds and parameter settings required to prevent error propagation across Bolivia
2. Materials and Methods
2.1. Study Zone
Bolivia covers approximately 1,098,006 km2 in central South America, with elevations ranging from 200 to 5500 m above sea level (m a.s.l.). Consequently, this broad altitudinal gradient creates three distinct ecological zones: the highlands, valleys, and lowlands. Driven by complex mountain systems, national hydrology divides into three major macro-basins: the Altiplano, Amazon, and La Plata (Figure 1).
Figure 1.
Hydrological and orographic map of Bolivia.
The Altiplano basin occupies southwestern Bolivia, spanning 66°05′ W–69°44′ W longitude and 14°35′ S–22°57′ S latitude. Covering an area of 151,722 km2, this endorheic system features elevations between 3500 and 5000 m a.s.l. Climatically, cold conditions prevail across the basin, with mean temperatures ranging from −4 to 5 °C. Geomorphologically, this high-altitude domain encompasses several major salt flats.
The Amazon basin extends across central, northern, and northeastern Bolivia, spanning 59°39′ W–69°40′ W longitude and 9°36′ S–20°32′ S latitude. Covering 720,792 km2, this vast system encompasses all three ecological zones, with elevations reaching up to 4200 m a.s.l. Hydrologically, the basin drains into the Atlantic Ocean through Brazil. Climatically, average temperatures across the region range from 4 to 28 °C.
The La Plata basin encompasses southern and southeastern Bolivia, spanning 57°23′ W–67°02′ W longitude and 16°10′ S–22°55′ S latitude. Covering an area of 225,492 km2, this system spans all three ecological zones, with elevations reaching up to 4000 m a.s.l. Hydrologically, the basin drains into the Atlantic Ocean through Argentina. Climatically, average temperatures range from −4 to 25 °C.
2.2. Precipitation Data
2.2.1. Rain Gauge
In this study, the precipitation database from the Bolivian Water Balance was used for the period 1980–2020. This dataset includes records from 171 stations. These records were gap-filled and quality-controlled for implementation in the water-balance analysis [20]. Figure 2 shows the spatial distribution of stations across Bolivia. Notably, a large proportion of stations are concentrated around the country’s main cities. In detail, 34 stations are in the Altiplano basin, 72 in the Amazon basin, and 65 in the La Plata basin.
Figure 2.
Location of available rain gauges in Bolivia.
The basin area and number of stations were analyzed to determine spatial monitoring density. Specifically, coverage values equal 4462.4 km2 per station in the Altiplano basin, 9740.4 km2 per station in the Amazon basin, and 3469.1 km2 per station in the La Plata basin. Consequently, none of the basins meet the density recommendations established by the World Meteorological Organization (WMO) [21].
2.2.2. Satellite Precipitation Product: GSMaP
Spatial and temporal rainfall patterns across the study domain were characterized using Global Satellite Mapping of Precipitation (GSMaP) estimates. These products were developed by the Japan Aerospace Exploration Agency (JAXA) as a core contribution to the Global Precipitation Measurement (GPM) mission initiative. The GSMaP algorithm synthesizes observational data from a heterogeneous constellation of satellite sensors to balance physical accuracy and temporal continuity [22]:
- Passive Microwave (PMW) Radiometers: carried by Low Earth Orbit (LEO) satellites (e.g., GPM Core Observatory, GMI, AMSR2, SSMIS, ATMS), PMW sensors measure thermal microwave radiation emitted and scattered by cloud hydrometeors. While PMW provides direct, physically based rain-intensity estimates, its temporal coverage remains constrained by satellite orbital passes.
- Infrared Radiometers (IRs): mounted on Geostationary Earth Orbit (GEO) satellites (e.g., GOES, Himawari, Meteosat), IR sensors measure cloud-top brightness temperatures at high spatial and temporal frequencies (every 10–15 min). Although IR has an indirect relationship to surface precipitation, it effectively fills temporal gaps between LEO PMW passes.
To merge these data sources, GSMaP employs atmospheric motion vectors from IR imagery to kinematically propagate PMW-derived precipitation estimates. Subsequently, a Kalman filter optimizes the trade-off between PMW physical accuracy and IR temporal frequency.
The GSMaP processing system categorizes its estimates into three primary operational streams, based on latency and post-processing complexity:
- GSMaP_NOW: A real-time product provided with near-zero latency, suitable for immediate operational hazard monitoring.
- GSMaP_NRT: A near-real-time product released approximately 4 h post-observation, applying basic orbital and sensor corrections.
- GSMaP_Standard: a delayed product available within 1 to 3 days post-observation. This stream applies rigorous reanalysis algorithms, orbital overlap integrations, and multi-satellite calibrations to achieve research-grade accuracy.
The standard operational stream has evolved through successive algorithm updates from version 5 to version 8. Currently, it includes three distinct sub-products based on their post-processing level:
- MVK (Moving Vector with Kalman filter): The core satellite-only product, combining PMW and IR data via propagation and filtering schemes.
- Gauge: A gauge-adjusted product applying post-processing bias-correction algorithms to MVK estimates using global daily gauge networks.
- GRev: A gauge-re-evaluated product developed specifically as a re-analysis iteration within version 6.
In this study, the GSMaP version 8 Gauge product (GSMaP_Gauge V8) under the standard stream was selected. Available from January 1998, this product provides a spatial resolution of 0.1° and hourly temporal sampling. To ensure temporal alignment with ground-based observations, hourly GSMaP data were accumulated into daily totals. Specifically, a -4 h temporal shift (UTC-4) was applied before accumulation. Consequently, this adjustment aligns the satellite retrieval window with Bolivia’s local time zone and surface station schedules, preventing the temporal misallocation of precipitation events.
2.3. Methodology
The scheme considers the temporal evaluation windows, the spatial distribution of rain-gauge precipitation data, and set spatial control units. Moreover, performance equations are proposed to ensure convergence.
To maximize data availability, the study defined two distinct temporal windows. A 41-year baseline (1980–2020) supported ground-based spatial interpolation. Meanwhile, a 23-year overlapping period (1998–2020) enabled satellite product evaluation and non-linear adjustment.
2.3.1. Rain-Gauge Interpolation
Spatial interpolation of rain-gauge observations was conducted over the 41-year period from 1 January 1980 to 31 December 2020 (14,976 daily time steps). This extended timeline maximized the historical baseline and captured long-term spatial climatology. While previous studies adopted Ordinary Kriging (OK) as the primary interpolation approach [11,12], OK does not explicitly account for Bolivia’s pronounced elevation gradients. To address topography-driven spatial variability, Kriging with External Drift (KED) was evaluated as a robust alternative [13,14]. In this approach, the ASTER Global Digital Elevation Model (ASTER GDEM) was utilized as an auxiliary spatial covariate across mountain terrains. Specifically, the DEM was resampled from its original 30 m spatial resolution to 5 km, to optimize computational efficiency during interpolation.
To evaluate spatial performance across this 41-year period, OK and KED were compared using a spatial cross-validation scheme standard for Andean topography [12,14]. Specifically, the rain-gauge network was randomly split into training (80%) and validation (20%) datasets. In the Figure 3, the validation set comprised 34 stations distributed across the country’s major basins: 14 in the Amazon, 7 in the Altiplano, and 13 in the La Plata basin.
Figure 3.
Location of training and validation gauges in Bolivia.
Furthermore, the interpolation workflow required pre-processing to prevent numerical errors on completely dry days (i.e., when all stations record zero precipitation). In multiplicative and iterative satellite–gauge blending frameworks, zero-precipitation spatial fields induce numerical instability, indeterminate divisions, or severe overestimation artifacts [12,23]. To resolve this issue, dry days were assigned a baseline representative minimum value of 0.1 mm/day. This adjustment aligns with established precipitation detection thresholds in multi-source merging algorithms [21,24].
Finally, due to the unbounded mathematical formulation of geostatistical interpolation, negative estimates may occasionally occur in regions with sparse networks or steep topographic gradients. Because negative values are physically inconsistent with rainfall, a post-processing threshold was applied. Consequently, any negative value was constrained to the minimum physical threshold of 0.1 mm/day [24].
2.3.2. Iterative Combination Methodology
Combined precipitation fields were generated using an iterative combination method over a 23-year evaluation period from 1 January 1998 to 31 December 2020. In total, this timeframe spans 8401 daily time steps per monitoring station. Specifically, this method compares ground observations with satellite precipitation datasets to minimize systemic discrepancies. As proposed by [11], the methodology uses sub-basins as spatial control units to facilitate input preparation for semi-distributed hydrological modeling. Within this framework, relative-error correction previously yielded satisfactory results in Bolivia across monthly, daily [15] and hourly [16] temporal resolutions.
To enhance SBP bias correction at the daily scale, four iterative combination equations were evaluated: relative error, weighted linear, geometric, and sigmoidal. These equations progressively reduce discrepancies between interpolated rain-gauge precipitation and satellite-derived estimates, through successive correction steps. Here, the key variables are defined as the following:
- Sim: represents the precipitation obtained from the satellite product.
- Obs: represents the interpolated rain-gauge precipitation.
- SimCorr: represents the corrected product generated at each iteration.
Sequentially, the iterative scheme updates the spatial field at each step. Consequently, the corrected product generated in one iteration serves as the direct input for the subsequent iteration.
2.3.3. Relative-Error Equation
The relative error follows the framework proposed by [11]. Specifically, the correction factor is derived from the proportional difference between observed and satellite precipitation (Equation (1)).
Subsequently, this factor is applied multiplicatively to the satellite product (Equation (2)).
In so doing, this approach progressively minimizes proportional discrepancies between both datasets across iterations.
2.3.4. Geometric Equation
The geometric method applies a multiplicative correction based on a power-law relationship between simulated and observed precipitation (Equation (3)).
here, the primary parameter is defined as
- k: represents the correction coefficient, with values ranging between 0 and 1.
Consequently, this formulation preserves proportional precipitation relationships while mitigating the influence of extreme discrepancies between datasets [18,25,26].
2.3.5. Weighted Linear Equation
The weighted linear method generates corrected precipitation through a linear combination of satellite and observed data (Equation (4)).
here, the primary parameter is defined as
- α: represents the weighting coefficient, ranging between 0 and 1.
Specifically, this parameter dictates the degree of adjustment toward ground measurements. Higher α values increase rain-gauge weight, whereas lower values preserve a greater proportion of original satellite information [17,27].
2.3.6. Sigmoidal Equation
The sigmoidal method employs an exponential relaxation function to progressively adjust satellite estimates toward ground observations (Equation (5)).
here, the primary parameter is defined as
- λ: represents the relaxation coefficient controlling the correction rate.
Specifically, higher λ values drive faster convergence toward observed data. Conversely, lower values maintain a stronger influence from the original satellite product [19,28].
2.3.7. Statistical Indicators
A comprehensive evaluation framework was implemented to quantitatively evaluate original satellite products, spatial interpolation fields, and blended satellite–gauge datasets against ground observations [29,30]. Specifically, all evaluations were conducted at a daily time step using a point-to-pixel extraction approach paired with cross-validation. Performance was assessed using continuous statistical metrics and categorical detection indicators based on a 0.1 mm/day rainfall threshold [30,31].
- Continuous Metrics: continuous performance was evaluated using standard goodness-of-fit metrics, detailed in Table 1. Specifically, linear association and variance explanation were quantified using the Correlation Coefficient and Coefficient of Determination [29,32]. Error magnitudes were captured using Mean Absolute Error and Root Mean Square Error, the latter penalizing larger deviations [33]. Systematic over- or underestimation was evaluated using Percent Bias [33]. Finally, overall model agreement was measured via Nash–Sutcliffe Efficiency [34] and Kling–Gupta Efficiency, which integrates correlation, bias, and variability components [35].
- Categorical Metrics: rainfall detection capability was evaluated using 2 × 2 contingency matrix based on Hits (H), False Alarms (F), and Misses (M). Specifically, detection efficiency was measured using Probability of Detection, False-Alarm Ratio, and Critical Success Index [30,31] (Table 1).
Table 1.
Summary of continuous and categorical statistical metrics.
3. Results
This section presents the main findings of the study. First, spatial interpolation of ground-based rain-gauge observations was evaluated using OK and KED. Next, the performance and convergence behavior of the iterative error-reduction algorithms were analyzed. Finally, a comparative evaluation assessed the spatial accuracy and efficiency of optimized precipitation estimates across Bolivia’s major hydroclimatic regions. Specifically, these domains encompass the Amazon, La Plata, and Altiplano basins.
3.1. Rain-Gauge Interpolation
3.1.1. Statistical Analysis for Interpolation
To establish a robust baseline for precipitation interpolation, OK and KED were evaluated using cross-validation. Specifically, an 80% training setup was tested against an independent 20% validation-station network.
As detailed in Table 2, cross-validation revealed a severe vulnerability to spatial overfitting in the OK approach. When evaluated against independent stations, OK metrics collapsed across Bolivia’s complex topographical relief. Specifically, OK yielded a nationwide R2 of only 0.18, an R of 0.43, and an unacceptable PBias, jumping to 48.28%. Furthermore, OK efficiency collapsed nationwide. This drop was reflected by a negative NSE and a KGE near zero, driven by severe error variance (RMSE of 5.1 mm/day). Consequently, these results demonstrate that OK depends heavily on station density. Moreover, it fails to generalize across Bolivia’s physiographic gradients.
Table 2.
Statistical Indicators for OK and KED.
In contrast, KED demonstrated high robustness and spatial predictability. By integrating topographic elevation as an external drift, KED successfully constrained regional errors. In doing so, it restored the predictive capability of the gauge network. Countrywide, KED elevated R2 to 0.71, boosted R to 0.84, and suppressed RMSE to 1.3 mm/day, while lowering MAE to 0.7 mm/day. Additionally, bias was curtailed from 48.28% to 11.24%, lifting NSE to 0.68 and KGE to 0.79, nationwide.
To test whether the performance superiority of KED over OK was statistically significant, a paired non-parametric Wilcoxon signed-rank test was applied to cross-validation residuals (Pobs-Pest) and metrics [36]. Notably, the test confirmed that KED yields a statistically significant improvement over OK across all study domains (p < 0.05). Consequently, this marked improvement stems directly from incorporating elevation. This covariate enables KED to model orographic precipitation gradients that isotropic OK systematically fails to capture.
At the regional scale, incorporating topographic dynamics through KED proved particularly crucial across macro-basins:
- Amazon Basin: Severe convective rainfall error variances under OK (RMSE = 7.1 mm/day, NSE = −2.87) were substantially mitigated by KED, which achieved an R2 of 0.63, an NSE of 0.60, and a KGE of 0.77.
- Altiplano Basin: OK suffered a catastrophic failure (R2 = 0.09, NSE = −6.68, KGE = −0.23) due to sparse gauge density across high-elevation plateaus. KED effectively reversed these limitations, attaining R2 of 0.57, an NSE of 0.53, and a KGE of 0.63, while reducing the RMSE from 5.7 to 1.4 mm/day.
- La Plata Basin: KED reduced the percentage bias from 57.28% to 14.45%, elevating the NSE from −1.40 to 0.42 and the KGE from 0.24 to 0.65.
According to categorical occurrence detection, both interpolation schemes exhibited remarkably high detection capability, with values exceeding 0.96 nationwide. Although KED slightly increased nationwide FAR, from 0.21 to 0.27, bringing a minor CSI decline from 0.78 to 0.73, this trade-off is strongly outweighed by substantial gains in continuous statistical accuracy and volumetric precision.
Overall, these findings justify selecting KED over OK as the baseline spatial interpolation framework. Specifically, KED supports satellite precipitation bias correction in data-scarce, topographically complex environments like the Bolivian Andes.
3.1.2. Spatial Analysis for Interpolation
The spatial distributions of mean daily precipitation generated by OK (Figure 4a) and KED (Figure 4b) highlight the structural impact of incorporating elevation as an auxiliary variable:
Figure 4.
Daily mean precipitation map of Bolivia for (a) OK Interpolation and (b) KED Interpolation.
- The OK surface exhibits severe spatial over smoothing. Specifically, it displays prominent circular interpolation artifacts (“bullseye” patterns) centered strictly around isolated station clusters. Because OK relies purely on isotropic spatial distance, it fails to reproduce the sharp physical boundaries imposed by the eastern Andean slope. Consequently, it projects unrealistically homogeneous high-precipitation zones across topographically heterogeneous terrain.
- By integrating topographic relief, the KED formulation effectively aligns spatial rainfall gradients with underlying physiography. Specifically, the resulting surface clearly delineates the high-precipitation belt along the humid sub-Andean front (oriental Andes). Simultaneously, it accurately captures the rain-shadow effect over the arid Altiplano plateau and dry valleys.
3.2. GSMaP Analysis
3.2.1. Statistical Analysis for GSMaP
The baseline statistical evaluation of the uncorrected GSMaP_Gauge V8 product against independent ground observations reveals moderate temporal correlation alongside high event-occurrence detection capability across Bolivia (Table 3). Specifically, this baseline is established prior to local iterative optimization. Nationally, GSMaP exhibits moderate linear correspondence with rain gauges, achieving an R2 of 0.6292 and an R of 0.7932. Systematically, the product exhibits a modest national volume deficit, with a PBias of −5.09%.
Table 3.
Statistical Indicators for GSMaP_Gauge V8 Product for Main Basins and Bolivia.
At the national scale, the baseline product demonstrates moderate performance, yielding an NSE of 0.5715 and a KGE of 0.7704. The higher KGE score reflects the metric’s explicit decomposition into correlation, variability, and bias ratios. This formulation preserves credit for GSMaP’s temporal alignment (R = 0.7932), despite daily residual dispersion (MAE = 0.8 mm/day, RMSE = 1.4 mm/day).
Across individual macro-basins, baseline continuous performance is more constrained, reflecting regional hydroclimatic and topographic controls:
- Amazon Basin: characterized by high precipitation volumes and convective activity, the Amazon displays moderate temporal correlation (R2 = 0.6891, KGE = 0.6608). However, it exhibits a noticeable volumetric underestimation (PBias = −11.74%), leading to a limited daily efficiency (NSE of 0.4043, MAE of 1.4 mm/day, RMSE of 2.7 mm/day).
- Altiplano Basin: strongly influenced by high-elevation topography and semi-arid regimes, the Altiplano exhibits the lowest predictive skill (R = 0.6135, NSE = 0.2311, KGE = 0.6042). Although volumetric deviation remains low (PBias = −6.20%) and error magnitudes are small (MAE = 0.8 mm/day; RMSE = 1.8 mm/day) due to low regional rainfall, daily temporal variance is poorly captured.
- La Plata Basin: demonstrates consistent correlation (R = 0.6926, KGE = 0.6231) alongside low efficiency (NSE = 0.3046). In contrast to the negative volumetric bias observed in the Amazon (−11.74%) and Altiplano (−6.20%), La Plata exhibits a slightly positive bias (6.36%). This distinct positive tendency is attributed to validation-station spatial sampling. In La Plata, gauges captured sub-regions dominated by satellite overestimation, whereas Amazon validation gauges were in foothills dominated by warm-rain underestimation.
In contrast to continuous quantitative metrics, GSMaP_Gauge V8 shows solid categorical skill in detecting precipitation occurrence:
- POD: High overall detection skill nationwide of 0.8797, peaking at 0.8622 in the Amazon and 0.8570 in La Plata, while maintaining 0.7982 in the Altiplano.
- FAR: Consistently low false-alarm rates nationwide of 0.0825, with regional values of 0.1016 in the Amazon, 0.1522 in La Plata, and 0.1931 in the Altiplano.
- CSI: Reaches an exceptional 0.8153 countrywide, maintained at 0.7856 in the Amazon, 0.7427 in La Plata, and 0.6702 in the Altiplano.
These baseline findings demonstrate that while GSMaP_Gauge V8 reliably captures precipitation timing and occurrence boundaries, its continuous daily quantitative skill is constrained—particularly in macro-basins like the Altiplano (NSE = 0.2311) and Amazon (NSE = 0.4043). These regional limitations confirm that raw baseline estimates cannot be considered fully reliable for direct modeling without spatial adjustment, thereby reinforcing the necessity of local optimization.
3.2.2. Spatial Analysis for GSMaP
Spatial comparisons between the interpolated surface and the satellite product illustrate structural differences in rainfall representation (Figure 5). Specifically, KED (Figure 5a) captures topographically driven regional-precipitation amounts. In contrast, GSMaP (Figure 5b) preserves continuous spatial details and convective features across ungauged areas. However, GSMaP yields noticeably lower overall precipitation magnitudes, due to widespread underestimation.
Figure 5.
Daily mean precipitation map of Bolivia for: (a) Rain-Gauge Interpolation KED and (b) GSMaP_Gauge V8 Product.
To further evaluate spatial error patterns, continuous metric maps were generated across Bolivia (Figure 6):
Figure 6.
Spatial statistical indicator maps for GSMaP_Gauge V8 Product: (a) RMSE, (b) PBias and (c) KGE. Black triangles (▲) and white circles (○) represent training gauges and validation gauges, respectively.
- MAE (Figure 6a): values span from 0 to 6 mm/day, with peak localized errors reaching 6 mm/day across high-relief and convective Amazonian zones. High error levels (>3 mm/day) are concentrated predominantly over the eastern Andean slopes and the lowland Amazon basin. Conversely, the western Altiplano maintains lower, bounded error ranges (<1.5 mm/day).
- PBias (Figure 6b): widespread underestimation (blue gradients, <−40%) dominates the northern, central, and highland regions. However, isolated pockets of positive PBias (red tones) emerge in the far northern tropical lowlands along the international borders with Peru and Brazil. These overestimations stem from sparse gauge coverage and satellite convective overestimation. Additionally, localized overestimations occur along mountain transition zones near southern borders.
- KGE (Figure 6c): despite the widespread volumetric bias, KGE values remain consistently high (green tones, 0.60–0.75) over most of the Amazon lowlands and eastern plains. In contrast, moderate drops in KGE (<0.5) occur along high-elevation mountain gradients in the western Andes, where topographical relief impairs passive microwave algorithms.
3.3. Sensitivity Analysis and Convergence Evaluation
To evaluate stability, convergence rate, and overall performance of the four iterative combination equations, a comprehensive sensitivity analysis was conducted. Specifically, this evaluation spanned Bolivia’s three main macro-basins and included the national average. Each method was systematically tested across a range of iteration steps and weight/scaling parameters. In doing so, performance was assessed using ten statistical and categorical metrics. In total, 120 sensitivity response figures detailing metric variations across parameter spaces and basins are provided in Figures S1–S120 of the Supplementary Materials. Consequently, the following subsections detail the performance behavior of each specific formulation.
3.3.1. Relative-Error Equation Analysis
The relative-error (RE) Equation adjusts GSMaP satellite estimates by calculating local multiplicative bias factors derived from ground observations across iterative steps. Specifically, Table 4 details national-level performance evolution over 10 iterations. Additionally, Tables S1, S2, and S3 in the Supplementary Materials provide detailed results for the Amazon, Altiplano, and La Plata basins, respectively.
Table 4.
Statistical Indicators for Relative-Error Iterations in Bolivia.
The sensitivity analysis for the RE equation successfully captured the delicate trade-off between bias reduction and error amplification across iterative steps. As shown in Table 4, the method exhibits rapid initial convergence during steps 1 to 5, where continuous metrics (RMSE, PBias and KGE) experience steep improvements. At Iteration 6, national independent validation metrics reached their optimal performance. KGE peaked at 0.8665, RMSE was minimized to 1 mm/day, NSE reached 0.8005, and systematic bias was virtually eliminated.
At Iteration 6, the RE equation demonstrates high spatial robustness, maintaining stable KGE values across Bolivia’s basins:
- Amazon Basin: it is characterized by high convective precipitation volumes, and convergence is smooth and stable, achieving KGE = 0.8049 and NSE = 0.6773. The initial volumetric underestimation is effectively scaled down (PBias reduced to −6.11%) while preserving regional spatial gradients.
- Altiplano Basin: it is marked by high-elevation topography and semi-arid conditions; this basin exhibits the highest initial satellite bias. By Iteration 6, it achieves significant relative gains (NSE = 0.6500, KGE = 0.6843), successfully mitigating systematic underestimation without generating extreme spatial artifacts.
- La Plata Basin: it demonstrates adjustment (KGE = 0.7374, NSE = 0.5691), striking an optimal balance between local error minimization (PBias = 4.49%) and spatial continuity.
Beyond Iteration 6, the analysis revealed clear mathematical divergence and spatial overfitting. As iterations progressed toward N = 10, percent bias shifted into positive territory, and escalated consistently. PBias reached +12.07% nationwide, +20.54% in the Altiplano and +17.31% in La Plata at Iteration 10. Concurrently, this bias overshoot degraded continuous error metrics. RMSE increased back to 1.1 mm/day and MAE to 0.6 mm/day, and national KGE dropped to 0.7947. Consequently, this pattern confirms the vulnerability of ratio-based corrections, which amplify local mathematical residuals if allowed to iterate beyond the convergence threshold.
Regarding categorical occurrence, the iterative process successfully boosts POD from 0.8797 in raw GSMaP to 0.9745 at Iteration 1, stabilizing to 0.9992 by Iteration 6. Although this improvement coincides with an increase in the FAR from baseline values (0.0825) to 0.2717 at Iteration 1 and stabilizing at 0.2852 by Iteration 6, the CSI remains high (0.7144). Therefore, the substantial reduction in systematic bias strongly justifies this minor categorical trade-off.
3.3.2. Geometric Equation Analysis
The geometric equation refines satellite precipitation estimates by incorporating a geometric adjustment factor. This factor is governed by the scaling hyperparameter (exponent) and the iteration step N. Figure 7 illustrates two-dimensional sensitivity heat maps of performance metrics across the (k, N) parameter space nationally. Additionally, Figures S1–S40 in the Supplementary Materials provide detailed basin-by-basin evaluations across all ten statistical and categorical metrics.
Figure 7.
Iteration heat map using geometric equation of Bolivia for KGE.
The two-dimensional sensitivity analysis for the geometric equation demonstrates an asymptotic convergence pattern. This behavior contrasts sharply with the volatility and error amplification observed in pure ratio-based approaches (e.g., relative error). Across all evaluated macro-basins, the highest error metrics (RMSE > 1.4 mm/day) are strictly confined to an underfitting zone. Specifically, this zone is characterized by low iteration counts (N ≤ 2) and conservative exponent values (k ≤ 0.3). At the national scale, optimal trade-offs between bias reduction, variance preservation, and error minimization occurred at Iteration 5 (N = 5) with an exponent value of k = 0.6. Within this specific parameter space:
- KGE stabilized at its peak national performance.
- RMSE and MAE were substantially reduced, while PBias was successfully minimized across Bolivia’s contrasting topographies.
- This configuration avoided localized over-correction and spatial distortion in areas with sparse station coverage.
- At Iteration 5 and k = 0.6, the macro-basin analysis confirms spatial stability:
- Amazon Basin (Figures S1–S10): exhibits stable behavior around k = 0.6 (KGE = 0.7422, NSE = 0.6610). The geometric scaling factor effectively scales down wet-season overestimation (PBias = −16.98%), while maintaining spatial smoothness across dense convective rainfall regions.
- Altiplano Basin (Figures S11–S20): shows higher sensitivity to parameter k, due to rugged high-elevation terrain and precipitation regimes. Setting k = 0.6 at Iteration 5 achieves KGE = 0.6967 and NSE = 0.6158. Consequently, this setup prevents extreme adjustments in semi-arid zones, while yielding significant bias reductions (PBias = −21.75%).
- La Plata Basin (Figures S21–S30): demonstrates balanced performance (KGE = 0.7162, NSE = 0.5608). It reaches stable metric plateaus by Iteration 5, striking an optimal balance between local error reduction (PBias = −14.14%) and spatial continuity.
Beyond Iteration 6 and for high exponent values (k ≥ 0.8), the geometric equation reaches an asymptotic plateau. Here, further iterations yield negligible efficiency gains (<0.01 gain in KGE). Regarding precipitation occurrence, the geometric equation demonstrated high detection precision (CSI = 0.8094, FAR = 0.0466, POD = 0.8427) at Iteration 5. Therefore, these characteristics confirm that N = 5 with k = 0.6 constitutes a robust, operationally stable standard for geometric merging.
3.3.3. Weighted Linear Equation Analysis
The weighted linear equation adjusts satellite estimates by blending the background field with observation-based interpolation of residuals. Specifically, this process is controlled by weighting parameter α and iteration step N. Figure 8 presents the national-scale sensitivity heatmap across the (α, N) parameter space for KGE. Additionally, Figures S41–S80 in the Supplementary Materials provide detailed basin-specific evaluations across all ten performance metrics.
Figure 8.
Iteration heat map using weighted linear equation of Bolivia for KGE.
The two-dimensional sensitivity analysis for the weighted linear method revealed a highly sensitive operational space (Figure 8). It displayed a distinct diagonal degradation pattern. Unlike the asymptotic stability observed in the geometric approach, the linear formulation proved prone to rapid spatial overfitting as iterations and weights increased. The bottom-right domain of the heatmap (N > 4) combined with aggressive blending weights (α ≥ 0.5) exhibited severe inflation of error metrics (RMSE > 1.3 mm/day) and a systematic collapse of efficiency.
To ensure cross-regional stability and avoid mathematical residual amplification, optimization restricted the operational threshold to Iteration 2, with a conservative weight of α = 0.2. At this specific intersection, national performance optimizes rapidly:
- National KGE reaches its maximum value of 0.8536, with NSE stabilizing at 0.7270, RMSE minimized to 1.2 mm/day, and MAE at 0.7 mm/day.
- Systematic volumetric bias is successfully neutralized. PBias shifts from a mild underestimation (−1.42% at N = 1) to an optimal near-zero value of 1.52% at N = 2. Consequently, this preserves the foundational spatial structure of the GSMaP sensor without mathematical residual breakdown.
Basin-scale evaluations under N = 2 and α = 0.2 highlight distinct spatial behaviors across topography:
- Amazon Basin (Figures S41–S50): show high stability (KGE = 0.7728, NSE = 0.6188, RMSE = 2.2 mm/day). By Iteration 2, satellite overestimation during convective rain events is successfully scaled back (PBias = −5.23%), while preserving continuous regional rainfall patterns.
- Altiplano Basin (Figures S51–S60): marked by sparse station networks and high elevation, this region benefits from the rapid adjustment (KGE = 0.6878, NSE = 0.5046, RMSE = 1.4 mm/day). Setting α = 0.2 effectively mitigates severe systematic bias (PBias = 3.62%) without causing localized distortions or artificial peaks around isolated rain gauges.
- La Plata Basin (Figures S61–S70): showed the highest vulnerability to parameter selection, due to linear amplification of interpolation residuals over transition terrain. Restricting parameters to α = 0.2 and Iteration 2 proved critical, yielding KGE = 0.7173, NSE = 0.5038, and constraining error inflation (RMSE = 1.8 mm/day, PBias = 10.82%).
Regarding categorical detection at α = 0.2 and Iteration 2, the formulation achieves high detection coverage (POD = 0.9685, CSI = 0.8265). However, this high sensitivity entails a moderate False-Alarm Ratio (FAR = 0.1507 nationally). Regional FAR peaks at 0.3665 in the Altiplano and 0.2625 in La Plata.
3.3.4. Sigmoidal Equation Analysis
The sigmoidal equation employs a non-linear S-shaped adjustment function to refine satellite precipitation estimates. Specifically, it provides smooth thresholding that prevents extreme over-corrections at the tails of the rainfall distribution. Here, the non-linear response is governed by the shape hyperparameter λ and the iteration step N. Figure 9 illustrates the national-scale sensitivity heatmap across the (λ, N) parameter space for KGE. Additionally, Figures S81–S120 in the Supplementary Materials provide detailed basin-specific evaluations for all ten continuous and categorical performance metrics.
Figure 9.
Iteration heat map using sigmoidal equation of Bolivia for KGE.
The parameter λ controls the steepness and transition boundary of the non-linear transfer function. Lower values produce a nearly linear response, whereas higher values create an abrupt threshold. Across the evaluation space, an optimal performance window emerged at Iteration 2. Within this operational window, setting λ = 0.7 at Iteration 2 offered the best regional compromise for systematic bias reduction without inducing localized spatial noise. At this operational intersection:
- National KGE reaches 0.8189, with NSE stabilizing at 0.7298, RMSE minimized to 1.2 mm/day, and MAE at 0.6 mm/day.
- Systematic precipitation deviation transitions to a controlled positive volume bias (PBias = 8.62%). In doing so, it damps underestimation during low-intensity drizzle while smoothly scaling down heavy convective rainfall peaks.
Beyond Iteration 3, the sigmoidal equation triggered a uniform overfitting cascade, as indicated by the widespread shift toward suboptimal error metrics across the validation matrix. Performance metrics rapidly collapsed into an asymptotic lower plateau (KGE = 0.7746, RMSE = 1.3 mm/day, and PBias = 13.27% across most λ values for N ≥ 6). Consequently, accumulating non-linear adjustments beyond two steps leads to mathematical saturation and spatial degradation.
In the case of basin analysis at Iteration 2 with λ = 0.7, macro-basin evaluations confirm distinct spatial responses:
- Amazon Basin (Figures S81–S90): benefits significantly from sigmoidal thresholding (KGE = 0.7884, NSE = 0.661, RMSE = 2.1 mm/day). Non-linear scaling achieves an optimal volume balance (PBias = 1.75%), effectively scaling down wet-season overestimation, while preserving continuous convective rainfall patterns.
- Altiplano Basin (Figures S91–S100): marked by high-elevation topography and sparse station coverage, this basin reaches its global optimum for error minimization (RMSE = 1.3 mm/day) and variance preservation (NSE = 0.6064, KGE = 0.6064). Consequently, this configuration acts as a protective buffer against gauge noise, despite yielding a higher volume bias (PBias = 14.17%).
- La Plata Basin (Figures S101–S110): reaches KGE = 0.6752 and NSE = 0.4767 (RMSE = 1.9 mm/day). However, it exhibits vulnerability to linear residual amplification in transition terrain, resulting in PBias = 15.61%.
Across the tested parameters domain, categorical detection metrics show high resilience around λ = 0.7 and Iteration 2. The method maintains a high POD (0.9940 nationally and >0.9900 across all basins) and a national CSI (0.7628). However, this high sensitivity leads to an elevated FAR nationally (0.2337), peaking at 0.4750 in the Altiplano and 0.3625 in La Plata. Therefore, while the sigmoidal function selectively scales intensity magnitudes, restricting parameters to Iteration 2 with λ = 0.7 is critical to prevent the unconstrained spatial expansion of false-precipitation areas.
3.4. Selected Combined-Product Analysis
Following the parameter optimization conducted in Section 3.3, the four combined daily precipitation products—relative error (Iteration 6), Geometric (k = 0.6, Iteration 5), weighted linear (α = 0.2, Iteration 2), and sigmoidal (λ = 0.7, Iteration 2)—were evaluated against ground-based observations to assess their performance across Bolivia’s diverse physiographic regions.
3.4.1. Statistical Analysis for Combined-Product
In the high-volume convective regime of the Amazon basin, the baseline GSMaP_Gauge V8 exhibits moderate performance alongside notable volumetric underestimation (Table 5). Applying iterative combination methodologies yields distinct performance improvements across all dimensions:
Table 5.
Statistical indicators for combination equation in Amazon basin.
- Relative error emerges as the superior choice for continuous performance. It leads across nearly all statistical indicators, achieving the highest explained variance, linear correlation, and overall efficiency (NSE and KGE), while minimizing random daily errors.
- Geometric attains the lowest absolute daily error and superior false-alarm suppression. However, it tends to over-suppress convective rainfall peaks, resulting in a pronounced negative volumetric bias.
- Weighted linear delivers balanced overall performance, while achieving the highest categorical detection success rate (CSI).
- Sigmoidal provides the most effective volumetric bias adjustment—virtually eliminating the baseline underestimation—while maintaining high efficiency, though it slightly trails the RE equation in overall error reduction.
Ultimately, for humid convective regimes where accurate peak-magnitude estimation and water balance are paramount, the relative-error equation proves most robust. Specifically, it markedly enhances efficiency (improving NSE from 0.40 to 0.68 and KGE from 0.66 to 0.80), while cutting random error variance by over 26% relative to raw GSMaP.
In the high-elevation, semi-arid environment of the Altiplano basin, topographical relief and sparse rain-gauge coverage severely constrain the baseline performance of GSMaP_Gauge V8 (Table 6). Evaluating the iterative blending approaches reveals clear regional trade-offs:
Table 6.
Statistical Indicators for Combination equation in Altiplano Basin.
- Relative error delivers the highest linear correlation and NSE. However, severe multiplicative residual amplification over dry soil triggers a dramatic inflation of false alarms (FAR exceeding 0.55), limiting its reliability for spatial mapping.
- Geometric outperforms all equations in overall efficiency (KGE) and spatial noise suppression. It achieves the lowest MAE, and excels at controlling false rain detections, making it particularly suitable for high-elevation plateaus.
- Weighted linear delivers the best volumetric mass balance, though it exhibits lower overall skill in capturing daily variance (R2 and NSE).
- Sigmoidal displays strong linear correlation but suffers from positive volumetric overestimation and elevated false-alarm rates.
In high-elevation, semi-arid watersheds, suppressing artificial rainfall artifacts around isolated gauge stations is paramount. Although the relative-error method achieves slightly higher linear variance, its excessive false-alarm rate makes it unreliable for continuous spatial mapping. Consequently, the geometric approach emerges as the most dependable framework for the Altiplano. It provides superior spatial structure, highest overall efficiency (KGE rising from 0.60 to 0.70), lowest absolute error (MAE reduced to 0.60 mm/day), and robust false-alarm suppression (FAR controlled at 0.14).
In La Plata basin, the baseline GSMaP_Gauge V8 product exhibits moderate temporal correlation but lower overall efficiency (NSE = 0.3046), accompanied by a slight positive volumetric bias (Table 7). The iterative combination equations yield distinct operational trade-offs across continuous and categorical performance:
Table 7.
Statistical Indicators for Combination Equation in La Plata Basin.
- Relative error leads to continuous performance, achieving the highest efficiency indicators (NSE and KGE), lowest random daily error (RMSE), and near-optimal volume balance. However, ratio-based scaling over low-rainfall thresholds leads to an increased false-alarm rate.
- Geometric performs better than linear correlation (R2), absolute error minimization (MAE), and categorical event detection (CSI). However, its primary limitation is a conservative negative volumetric bias (PBias = −14.14%).
- Weighted linear offers balanced continuous performance, though it introduces moderate positive-volume overestimation.
- Sigmoidal exhibits the lowest relative skill among the optimized models, hampered by elevated positive bias and lower efficiency.
In this basin, selection depends on primary operational goals:
- The relative-error method is optimal, due to its superior efficiency (NSE rising from 0.30 to 0.57; KGE reaching 0.74), lowest RMSE, and minimal volumetric bias (PBias = 4.48%).
- The geometric approach is preferred, as it successfully suppresses false-alarm inflation (FAR ≈ 0.12, compared to 0.43 under relative-error method) while maximizing categorical success (CSI ≈ 0.75).
At the national scale, the baseline GSMaP_Gauge V8 product provides a solid initial benchmark, characterized by moderate efficiency and a slight nationwide volume deficit (Table 8). Applying the iterative combination equations yields clear performance distinctions nationwide:
Table 8.
Statistical Indicators for Combination Equation in Bolivia.
- Relative error dominates virtually all continuous statistical indicators nationwide. It achieves the highest explained variance, linear correlation, and efficiency (NSE and KGE), while reducing random daily error below 1.0 mm/day and practically eliminating national volumetric bias (PBias = −0.24%).
- Geometric attains the lowest absolute daily error nationwide (MAE) and exceptional false-alarm suppression (FAR ≈ 0.05). However, its application is constrained by a conservative negative-volume bias.
- Weighted linear ranks second in overall efficiency while achieving the highest nationwide critical success index (CSI ≈ 0.83). It strikes an optimal operational balance between high event-detection skill (POD) and robust FAR.
- Sigmoidal demonstrates strong overall efficiency, though it introduces a slight positive overestimation of nationwide rainfall volume.
Overall, the RE equation emerges as the undisputed national benchmark leader for continuous streamflow driving and water-balance assessments. It boosts NSE from 0.57 to 0.80 and KGE from 0.77 to 0.87, bringing total volumetric deviation virtually to zero. However, where categorical event detection and false-alarm minimization are prioritized over strict volumetric mass balance, the weighted linear approach serves as the optimal operational alternative (CSI ≈ 0.83, KGE ≈ 0.85).
3.4.2. Spatial Analysis for Combined-Product
All four optimized products successfully reproduce Bolivia’s prominent hydroclimatic gradient. First, mean daily precipitation transitions smoothly from high-volume convective rainfall belts in northern and central Amazon lowlands (>4.5 to 6.0 mm/day) to semi-arid conditions across the high-elevation Altiplano plateau (<1.5 mm/day, pale blue).
Next, relative error (Figure 10a) and sigmoidal (Figure 10d) generate continuous precipitation fields that preserve regional macro-scale moisture corridors entering from the southwestern Amazon without introducing localized artificial steps. In contrast, geometric (Figure 10b) exhibits a noticeably drier rainfall field across eastern lowlands and southern Chaco regions. This visual attenuation reflects its overall conservative power-law scaling. Finally, weighted linear (Figure 10c) reproduces general regional patterns, but displays minor spatial disruptions in the sub-Andean transition zone.
Figure 10.
Daily mean precipitation map of Bolivia for (a) Relative-Error Equation, (b) Geometric Equation, (c) Weighted Linear Equation and (d) Sigmoidal Equation.
Figure 11a,b,d demonstrate exceptional daily error suppression across the vast majority of the national territory. MAE values remain uniformly low (<1.5 mm/day) throughout the Altiplano, La Plata basin, and lower Amazon lowlands. In contrast to the other three products, weighted linear (Figure 11c) exhibits a severe, localized error hotspot in the northern/northwestern Amazon basin, where daily MAE spikes to 3 to 6 mm/day (orange, red, and magenta shading). This spatial artifact highlights the mathematical vulnerability of linear combination methods when interpolating residuals over high-intensity convective rainfall regimes. Relative error (Figure 11a) and geometric (Figure 11b) maintain minimal random error magnitudes, even in the highest rainfall zones of the Amazon, proving that non-linear and ratio-based adjustments prevent localized error inflation.
Figure 11.
MAE Map of Bolivia for (a) Relative-Error Equation, (b) Geometric Equation, (c) Weighted Linear Equation and (d) Sigmoidal Equation.
Regarding volumetric accuracy, the relative error (Figure 12a) achieves an almost complete elimination of volumetric bias across the Amazon macro-basin and central lowlands, represented by near-zero white and pale-blue shading (PBias ≈ 0%). Slight negative bias (blue shading) is confined to the southwestern Altiplano margins. The geometric (Figure 12b) displays widespread negative bias ranging from −20% to −60% across all three macro-basins. This pattern confirms that while exponential scaling (k = 0.6) successfully controls extreme values, it systematically underestimates the total rainfall volumes in dry and transitional regions. Additionally, both weighted linear (Figure 12c) and sigmoidal (Figure 12d) display localized pockets of positive volumetric overestimation (pale red/pink tones, PBias > 30% to 50%) in southern Andean–La Plata transition zones and northwestern border areas with Peru.
Figure 12.
PBias Map of Bolivia for (a) Relative-Error Equation, (b) Geometric Equation, (c) Weighted Linear Equation and (d) Sigmoidal Equation.
High efficiency (KGE > 0.75, deep blue) covers the vast majority of the Amazon basin and eastern plains across all four post-processed products, demonstrating that satellite-gauge integration is highly effective in low-relief, gauge-dense lowlands. However, the critical spatial differentiator among products occurs along the high-elevation Andean cordillera and Altiplano plateau:
- Relative error (Figure 13a) maintains high efficiency (KGE > 0.75) across 85% of the national territory, but suffers a sharp efficiency drop (KGE < 0.25, yellow–green shading) along the narrow, extreme western high-elevation border with Peru and Chile.
- Geometric (Figure 13b) provides the most spatially continuous and stable efficiency field (KGE ≈ 0.50 to 0.75, green–blue shading) over the Altiplano plateau, avoiding sharp boundary drop-offs seen in other methods. This performance confirms its robustness in high-elevation, semi-arid environments
- Weighted linear (Figure 13c) displays the broadest zone of efficiency degradation (KGE < 0.50, bright green/yellow shading) extending from the western Altiplano into the central Andean valleys.
Figure 13.
KGE Map of Bolivia for (a) Relative-Error Equation, (b) Geometric Equation, (c) Weighted Linear Equation and (d) Sigmoidal Equation.
Sigmoidal (Figure 13d) achieves exceptional spatial uniformity over the eastern two-thirds of the country (KGE > 0.80), displaying a clear efficiency transition along the western mountainous barrier.
4. Discussion
Bias correction in satellite precipitation products cannot be treated as a simple, independent numerical adjustment at individual stations [8,10]. Correcting only temporal series at point locations introduces severe spatial discontinuities across the grid [17,24]. Therefore, to preserve hydrological cycle physics, temporal bias correction must be intrinsically coupled with continuous spatial rainfall-field reconstruction [26]. Actually, the spatiotemporal integration ensures that precipitation volume redistribution respects seasonal variability and basin-scale storm-front coherence.
It was shown that using KED with a Digital Elevation Model (DEM) overcomes OK rigidity and isotropy, removing spatial “bullseye” artifacts [12,13,14]. However, relying exclusively on a static elevation covariate explains spatial discrepancies along topographical relief boundaries. The KGE drop below 0.25 along the western high-elevation cordillera stems from extreme rain-shadow effects. At these high altitudes, precipitation is physically suppressed by dry, descending leeward airflows. Consequently, elevation applied as a positive linear drift causes model overestimation. Similarly, Yungas transition zones exhibit sharp precipitation gradients. Here, windward slopes experience intense convective orographic uplift, while adjacent interior valleys at identical elevations lie within sheltered rain shadows. Because elevation alone does not account for wind directionality or microphysical air movement, univariate external-drift models struggle to capture these directional contrast boundaries. This limitation highlights the need to evaluate dynamic wind-field covariates or orographic moisture fluxes in future work [37,38].
Discrepancies in initial GSMaP performance across hydroclimatic regions [6,7] respond directly to passive microwave radiometry physical principles [22]. In the Altiplano, cold soils, low salt-flat emissivity, and shallow precipitation with low liquid-water content cause scattering algorithms to misinterpret cold-ground thermal signatures as precipitating clouds [22,39]. Therefore, this misinterpretation may lead to systematic overestimation. On the other hand, in Amazon and La Plata lowlands dominated by deep convective systems with abundant atmosphere condensation and ice particles, radiometric signatures are captured with higher accuracy, confining errors primarily to extreme event underestimation.
For previous approaches that iterated relative-error algorithms unbounded at the basin scale [15,16]—causing divergence, noise amplification, and excessive positive bias—defining an explicit stopping criterion is critical. While minor local peak fluctuations exist across individual basins, standardizing the operational pipeline to a fixed iteration limit (Iteration 5 for geometric and Iteration 6 for relative error) yields an optimal trade-off. Specifically, this standardization results in negligible performance losses (<3% efficiency variation), while maximizing nationwide and and achieving minimal residual volumetric bias (). Operationally, enforcing uniform iteration limits avoids workflow fragmentation and guarantees spatial continuity across basin boundaries during grid generation.
However, the region-specific performance of each equation demonstrates that no universal bias-correction algorithm exists [11,20]. The geometric equation with exponent k = 0.6 acts as a non-linear low-pass filter. Specifically, it attenuates weak signals and suppresses background noise. This behavior makes it ideal in the semi-arid Altiplano for eliminating false rain detections () and enhancing detection skill (). In contrast, in hyper-humid convective Amazon and La Plata basins, the relative-error equation proves superior, due to its mass-conservation capacity. In doing so, it adjusts massive, accumulated volumes without distorting peak-precipitation variability.
Compared to global [1] or fixed-scale corrections applied across Andean and tropical basins [6,7,27], the proposed adaptive framework demonstrates that adjustment mathematical structure must respond to local climate physics [19,28]. Evaluating this dynamic framework against static models in the literature reveals substantial metric-stability improvements. This result validates the fact that combining modular orographic interpolation with bounded iterative optimization overcomes overfitting and divergence bottlenecks identified in previous regional studies [11,15].
Eliminating spatial artifacts and minimizing volumetric bias has direct implications for water-resource modeling and national water-balance assessments [4,20]. Moreover, in distributed hydrological modeling schemes [3,15], uncorrected satellite precipitation fields distort concentration-time estimates and overestimate peak flows, due to fictitious spatial-rainfall centers [29]. Conversely, the spatial homogenization achieved here ensures that flow routing across drainage networks reflects realistic rainfall volumes and distributions, thereby increasing certainty in peak-flow hydrographs and drought simulations.
An inherent limitation of this study is the uneven rain-gauge network density [21], which introduces representativeness uncertainties into cross-validation cells, particularly in high-mountain areas. To overcome gauge scarcity and the univariate constraint of KED, future research should explore integrating machine learning algorithms [9] coupled with dynamic wind-field covariates (e.g., ERA5 wind vectors), aspect-slope exposure indices, and orographic moisture-flux variables [37,38]. Incorporating these dynamic atmospheric variables will help resolve localized windward–leeward microclimates in mountain terrain where topographic elevation alone is insufficient.
5. Conclusions
This study evaluated four combination equations—relative error, geometric, weighted linear, and sigmoidal—to improve daily satellite precipitation estimates from GSMaP_Gauge V8 in Bolivia.
Integrating elevation as an external covariate via KED proved essential to obtain rainfall patterns in complex terrains. The OK method showed weak performance indicators (R2 = 0.18, NSE = −4.19, PBias = 48.28%). In contrast, the KED approach restored spatial continuity and predictive capability nationwide (R2 = 0.71, NSE = 0.68, KGE = 0.79, PBias = 11.24%, RMSE = 1.3 mm/day). Consequently, it provided a physically consistent baseline for satellite bias correction.
Overall, the relative-error equation (N = 6) combined with KED serves as the nationwide benchmark. It achieved the highest efficiency (NSE = 0.8005, KGE = 0.8665) and virtually eliminated volumetric bias (PBias = −0.24%). However, performance varies by hydro-climatic regime, warranting sub-basin tailored selections:
- The relative-error equation (N = 6) is optimal for the convective Amazon basin (KGE = 0.8049).
- The geometric equation (k = 0.6, N = 5) is recommended for the arid and high-altitude Altiplano, to suppress false alarms (FAR = 0.1360, KGE = 0.6967).
- The weighted linear equation (α = 0.2, N = 2) is preferred for transition zones in the La Plata basin, due to its high event-detection capability (CSI = 0.8265).
Sensitivity heatmaps demonstrated that iterative algorithms must be strictly bound. Iterating beyond optimal thresholds (N > 6 for relative error; N > 2 for weighted linear and sigmoidal) triggers exponential residual amplification. Consequently, this over-iteration causes positive bias overshooting, higher random errors, and degraded spatial precision.
In short, the optimized iterative combination framework successfully bridges the gap between point-scale rain-gauge accuracy and continuous satellite coverage across topographically relief, data-scarce Andean–Amazonian regions. Crucially, the resulting merged products provide reliable daily precipitation estimates to advance hydrological modeling, water-resource management, and climate risk assessments in Bolivia. For future research, efforts should focus on evaluating sub-daily time steps and assessing blended product performance in driving distributed hydrological models. Furthermore, this methodology can be applied to other data-scarce countries, applying tailored iterative equations across major basins to maximize regional accuracy.
Supplementary Materials
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/meteorology5040033/s1, Tables S1–S3: Statistical indicators for relative error equation iteration for Amazon, Altiplano and La Plata basins; Figures S1–S10: Iteration Heat Map using geometric equation in Amazon basin for statistical indicators; Figures S11–S20: Iteration Heat Map using geometric equation in Altiplano basin for statistical indicators; Figures S21–S30: Iteration Heat Map using geometric equation in La Plata basin for statistical indicators; Figures S31–S40: Iteration Heat Map using geometric equation in Bolivia for statistical indicators; Figures S41–S50: Iteration Heat Map using weight linear equation in Amazon basin for statistical indicators; Figures S51–S60: Iteration Heat Map using weight linear equation in Altiplano basin for statistical indicators; Figures S61–S70: Iteration Heat Map using weight linear equation in La Plata basin for statistical indicators; Figures S71–S80: Iteration Heat Map using weight linear equation in Bolivia for statistical indicators; Figures S81–S90: Iteration Heat Map using sigmoidal equation in Amazon basin for statistical indicators; Figures S91–S100: Iteration Heat Map using sigmoidal equation in Altiplano basin for statistical indicators; Figures S101–S110: Iteration Heat Map using sigmoidal equation in La Plata basin for statistical indicators; Figures S111–S120: Iteration Heat Map using sigmoidal equation in Bolivia for statistical indicators.
Author Contributions
Conceptualization, J.U. and O.S.; methodology, J.U. and O.S.; software, J.U.; validation, J.U., O.S. and T.U.; formal analysis, J.U. and O.S.; investigation, J.U. and O.S.; resources, J.U.; data curation, J.U.; writing—original draft preparation, J.U.; writing—review and editing, O.S. and T.U.; visualization, J.U.; supervision, O.S. and T.U.; project administration, O.S. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
Publicly available data sets were analyzed in this study. The combined product data can be found here: https://zenodo.org/records/22088747 (accessed 29 September 2026).
Acknowledgments
This research has been supported by the 4th Research Announcement on the Earth Observations of the Japan Aerospace Exploration Agency (JAXA) for data sharing of GSMaP_Gauge version 8. During the preparation of this manuscript, the authors used Gemini AI (version 3.6 Thinking) to assist with the analysis of results and language editing. The authors have reviewed and edited the output and take full responsibility for the content of this publication.
Conflicts of Interest
The authors declare no conflicts of interest.
References
- Huffman, G.J.; Adler, R.F.; Behrangi, A.; Bolvin, D.T.; Nelkin, E.J.; Gu, G.; Ehsani, M.R. The New Version 3.2 Global Precipitation Climatology Project (GPCP) Monthly and Daily Precipitation Products. J. Clim. 2023, 36, 7635–7655. [Google Scholar] [CrossRef] [Scilit]
- Teston, A.; Scolaro, T.P.; Maykot, J.K.; Ghisi, E. Comprehensive Environmental Assessment of Rainwater Harvesting Systems: A Literature Review. Water 2022, 14, 2716. [Google Scholar] [CrossRef] [Scilit]
- Kundzewicz, Z.W.; Pińskwar, I. Are Pluvial and Fluvial Floods on the Rise? Water 2022, 14, 2612. [Google Scholar] [CrossRef] [Scilit]
- Chen, H.; Huang, S.; Xu, Y.-P.; Teegavarapu, R.S.V.; Guo, Y.; Xie, J.; Nie, H. Quantitative Assessment of Impact of Climate Change and Human Activities on Streamflow Changes Using an Improved Three-Parameter Monthly Water Balance Model. Remote Sens. 2022, 14, 4411. [Google Scholar] [CrossRef] [Scilit]
- Ryu, S.; Song, J.J.; Lee, G. Radar–Rain Gauge Merging for High-Spatiotemporal-Resolution Rainfall Estimation Using Radial Basis Function Interpolation. Remote Sens. 2025, 17, 530. [Google Scholar] [CrossRef] [Scilit]
- Blacutt, L.A.; Herdies, D.L.; De Gonçalves, L.G.G.; Vila, D.A.; Andrade, M. Precipitation Comparison for the CFSR, MERRA, TRMM3B42 and Combined Scheme Datasets in Bolivia. Atmos. Res. 2015, 163, 117–131. [Google Scholar] [CrossRef] [Scilit]
- Salazar, Á.; Larrea-Alcázar, D.M.; Bertin, A.; Gouin, N.; Pareja, A.; Morales, L.; Maillard, O.; Ocampo-Melgar, D.; Squeo, F.A. Spatio-Temporal Evaluation of MSWEP, CHIRPS and ERA5-Land Reveals Regional-Specific Responses Across Complex Topography in Bolivia. Atmosphere 2025, 16, 1281. [Google Scholar] [CrossRef] [Scilit]
- Cannon, A.J.; Sobie, S.R.; Murdock, T.Q. Bias Correction of GCM Precipitation by Quantile Mapping: How Well Do Methods Preserve Changes in Quantiles and Extremes? J. Clim. 2015, 28, 6938–6959. [Google Scholar] [CrossRef] [Scilit]
- Li, H.; Zhang, Y.; Lei, H.; Hao, X. Machine Learning-Based Bias Correction of Precipitation Measurements at High Altitude. Remote Sens. 2023, 15, 2180. [Google Scholar] [CrossRef] [Scilit]
- Iqbal, Z.; Shahid, S.; Ahmed, K.; Wang, X.; Ismail, T.; Gabriel, H.F. Bias Correction Method of High-Resolution Satellite-Based Precipitation Product for Peninsular Malaysia. Theor. Appl. Climatol. 2022, 148, 1429–1446. [Google Scholar] [CrossRef] [Scilit]
- Saavedra, O.; Ureña, J. Generation of Combined Daily Satellite-Based Precipitation Products over Bolivia. Remote Sens. 2022, 14, 4195. [Google Scholar] [CrossRef] [Scilit]
- Caloiero, T.; Pellicone, G.; Modica, G.; Guagliardi, I. Comparative Analysis of Different Spatial Interpolation Methods Applied to Monthly Rainfall as Support for Landscape Management. Appl. Sci. 2021, 11, 9566. [Google Scholar] [CrossRef] [Scilit]
- Cantet, P. Mapping the Mean Monthly Precipitation of a Small Island Using Kriging with External Drifts. Theor. Appl. Climatol. 2017, 127, 31–44. [Google Scholar] [CrossRef] [Scilit]
- Buttafuoco, G.; Conforti, M. Improving Mean Annual Precipitation Prediction Incorporating Elevation and Taking into Account Support Size. Water 2021, 13, 830. [Google Scholar] [CrossRef] [Scilit]
- Saavedra, O.; Ureña, J.; Perales, M. Implementation of HydroBID Model with Satellite-Based Precipitation Products in Guadalquivir Basin, Bolivia. Water 2023, 15, 3250. [Google Scholar] [CrossRef] [Scilit]
- Alcala, F.; Ureña, J.; Saavedra, O. Generación de Productos Combinados de Precipitación En La Cuenca Choqueyapu, Bolivia. Rev. Geogr. Chile Terra Aust. 2025, 61, 58–68. [Google Scholar] [CrossRef] [Scilit]
- Vila, D.A.; De Goncalves, L.G.G.; Toll, D.L.; Rozante, J.R. Statistical Evaluation of Combined Daily Gauge Observations and Rainfall Satellite Estimates over Continental South America. J. Hydrometeorol. 2009, 10, 533–543. [Google Scholar] [CrossRef] [Scilit]
- Tang, L.; Tian, Y.; Yan, F.; Habib, E. An Improved Procedure for the Validation of Satellite-Based Precipitation Estimates. Atmos. Res. 2015, 163, 61–73. [Google Scholar] [CrossRef] [Scilit]
- Goly, A.; Teegavarapu, R.S.V. Optimization and Variants of Quantile-Based Methods for Bias Corrections of Statistically Downscaled Precipitation Data. J. Hydrol. Eng. 2020, 25, 04020027. [Google Scholar] [CrossRef] [Scilit]
- Ministerio de Medio Ambiete y Agua (MMAyA); Stockholm Environment Institute. Balance Hídrico de Bolivia 1980–2020; Ministerio de Medio Ambiente y Agua (MMAyA): La Paz, Bolivia, 2024.
- World Meteorological Organization. Guide to Hydrological Practices, 6th ed.; WMO, Ed.; World Meteorological Organization: Geneva, Switzerland, 2008. [Google Scholar]
- Kubota, T.; Aonashi, K.; Ushio, T.; Shige, S.; Takayabu, Y.N.; Kachi, M.; Arai, Y.; Tashima, T.; Masaki, T.; Kawamoto, N.; et al. Global Satellite Mapping of Precipitation (GSMaP) Products in the GPM Era. In Satellite Precipitation Measurement; Levizzani, V., Kidd, C., Kirschbaum, D.B., Kummerow, C.D., Nakamura, K., Turk, F.J., Eds.; Advances in Global Change Research; Springer International Publishing: Cham, Switzerland, 2020; Volume 67, pp. 355–373. [Google Scholar]
- Agou, V.D.; Pavlides, A.; Hristopulos, D.T. Spatial Modeling of Precipitation Based on Data-Driven Warping of Gaussian Processes. Entropy 2022, 24, 321. [Google Scholar] [CrossRef] [Scilit]
- Le, X.-H.; Koyama, N.; Kikuchi, K.; Yamanouchi, Y.; Fukaya, A.; Yamada, T. Evaluating Geostatistical and Statistical Merging Methods for Radar–Gauge Rainfall Integration: A Multi-Method Comparative Study. Remote Sens. 2025, 17, 2622. [Google Scholar] [CrossRef] [Scilit]
- Leander, R.; Buishand, T.A. Resampling of Regional Climate Model Output for the Simulation of Extreme River Flows. J. Hydrol. 2007, 332, 487–496. [Google Scholar] [CrossRef] [Scilit]
- Elsebaie, I.H.; Kawara, A.Q.; Alharbi, R.; Alnahit, A.O. Bias Correction Methods Applied to Satellite Rainfall Products over the Western Part of Saudi Arabia. Atmosphere 2025, 16, 772. [Google Scholar] [CrossRef] [Scilit]
- Funk, C.; Peterson, P.; Landsfeld, M.; Pedreros, D.; Verdin, J.; Shukla, S.; Husak, G.; Rowland, J.; Harrison, L.; Hoell, A.; et al. The Climate Hazards Infrared Precipitation with Stations—A New Environmental Record for Monitoring Extremes. Sci. Data 2015, 2, 150066. [Google Scholar] [CrossRef] [Scilit]
- Alizadeh, Z.; Yazdi, J.; Najafi, M.S. Improving the Outputs of Regional Heavy Rainfall Forecasting Models Using an Adaptive Real-Time Approach. Hydrol. Sci. J. 2022, 67, 550–563. [Google Scholar] [CrossRef] [Scilit]
- Moriasi, D.N.; Arnold, J.G.; Van Liew, M.W.; Bingner, R.L.; Harmel, R.D.; Veith, T.L. Veith Model Evaluation Guidelines for Systematic Quantification of Accuracy in Watershed Simulations. Trans. ASABE 2007, 50, 885–900. [Google Scholar] [CrossRef] [Scilit]
- Wilks, D.S. Forecast Verification. In Statistical Methods in the Atmospheric Sciences; Elsevier: Amsterdam, The Netherlands, 2011; pp. 301–394. [Google Scholar]
- Schaefer, J.T. The Critical Success Index as an Indicator of Warning Skill. Weather Forecast. 1990, 5, 570–575. [Google Scholar] [CrossRef] [Scilit]
- Legates, D.R.; McCabe, G.J. Evaluating the Use of “Goodness-of-Fit” Measures in Hydrologic and Hydroclimatic Model Validation. Water Resour. Res. 1999, 35, 233–241. [Google Scholar] [CrossRef] [Scilit]
- Willmott, C.J. Some Comments on the Evaluation of Model Performance. Bull. Am. Meteorol. Soc. 1982, 63, 1309–1313. [Google Scholar] [CrossRef] [Scilit]
- Nash, J.E.; Sutcliffe, J.V. River Flow Forecasting Through Conceptual Models Part I—A Discussion of Principles. J. Hydrol. 1970, 10, 282–290. [Google Scholar] [CrossRef] [Scilit]
- Gupta, H.V.; Kling, H.; Yilmaz, K.K.; Martinez, G.F. Decomposition of the Mean Squared Error and NSE Performance Criteria: Implications for Improving Hydrological Modelling. J. Hydrol. 2009, 377, 80–91. [Google Scholar] [CrossRef] [Scilit]
- Wilcoxon, F. Individual Comparisons by Ranking Methods. Biom. Bull. 1945, 1, 80. [Google Scholar] [CrossRef] [Scilit]
- Shikhovtsev, M.Y.; Shikhovtsev, A.Y.; Kovadlo, P.G.; Obolkin, V.A.; Molozhnikova, Y.V. Influence of Air Movement Structure on Microphysical Properties of the Atmosphere over Listvyanka. Atmos. Ocean. Opt. 2025, 38, 180–187. [Google Scholar] [CrossRef] [Scilit]
- Shikhovtsev, M.Y.; Shikhovtsev, A.Y.; Lezhenin, A.A.; Gradov, V.S.; Khaikin, V.B.; Kirichenko, K.E.; Kovadlo, P.G. WRF Model Estimation of Precipitable Water Vapor over the Sayan Solar Observatory and Khulugaysha Peak. Atmos. Ocean. Opt. 2026, 39, 355–361. [Google Scholar] [CrossRef] [Scilit]
- Tashima, T.; Kubota, T.; Mega, T.; Ushio, T.; Oki, R. Precipitation Extremes Monitoring Using the Near-Real-Time GSMaP Product. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 5640–5651. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.












