Unveiling the Accuracy of New-Generation Satellite Rainfall Estimates across Bolivia’s Complex Terrain

Abstract

This study evaluated the accuracy of two new generation satellite rainfall estimates (SREs): Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) and Integrated Multi-satellite Retrieval for GPM (IMERG) over Bolivia’s complex terrain. These SREs were compared against rainfall data from rain gauge measurements on a point-to-pixel basis for the period 2002–2020. The evaluation was performed across three regions with distinct topographical settings: Altiplano (Highland), Valles (Midland), and Llanos (Lowland). IMERG exhibited better accuracy in rainfall detection than CHIRPS, with the highest rainfall detection skills observed in the Highland region. However, IMERG’s higher rainfall detection skill was countered by its higher false alarm ratio. CHIRPS provided a more accurate estimation of rainfall amounts across the three regions, exhibiting low random errors and relative biases below 10%. IMERG tended to overestimate rainfall amounts, with marked overestimation by up to 75% in the Highland region. Bias decomposition revealed that IMERG’s high false rainfall bias contributed to its marked overestimation of rainfall. We showcase the utility of long-term CHIRPS data to investigate spatio-temporal rainfall patterns and meteorological drought occurrence in Bolivia. The findings of this study offer valuable insights for choosing appropriate SREs for informed decision-making, particularly in regions of complex topography lacking reliable gauge data.

1. Introduction

Rainfall is a key component of the hydrologic cycle, and its accurate measurement and monitoring of its spatio-temporal distribution are essential for various applications. This includes monitoring droughts/floods under changing climate [1,2,3,4], water resources management [3,5,6], assessing land degradation by water erosion [7,8,9,10,11], and proposing appropriate land management options [12,13,14]. Reliable rainfall data can be obtained from traditional rain gauges, which offer direct rainfall measurements. However, Bolivia still struggles with a lack of reliable rainfall data of sufficient quality and quantity for large parts of the country. This problem is particularly critical since large parts of Bolivia are characterized by complex topography ranging from the high Andean peaks to Lowland Amazonian regions, where rainfall is highly variable across space and time. Traditional gauge-based methods often face limitations in capturing the spatio-temporal variability of rainfall in such heterogeneous terrain, because of the limited number of stations, their uneven distribution, and the poor data record [15,16,17].

Satellite-based rainfall estimates (SREs) offer a valuable alternative to complement data gaps in ground-based measurements, provide comprehensive coverage across different timescales and topographic settings [18,19,20]. However, it is important to acknowledge that SREs come with their own set of uncertainties, such as those stemming from retrieval algorithms and sampling frequency, which can impact the accuracy of rainfall estimates. Thus, it is crucial to assess the accuracy of SREs before integrating them into operational applications to inform decision-making [21,22]. The emergence of new generation SREs such as CHIRPS (Climate Hazards Groups InfraRed Precipitation with Station data [23] and IMERG (Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM) [24] has promised enhanced capabilities in capturing rainfall variability with higher resolution and improved accuracy. CHIRPS is a high-resolution rainfall dataset developed by the Climate Hazards Group at the University of California, Santa Barbara, which combines satellite infrared data with ground station observations [23]. IMERG, developed by NASA’s (National Aeronautics and Space Administration) GPM mission, integrates rainfall estimates from multiple satellite sources, including passive microwave sensors and infrared data [24,25]. These new generation SREs utilize advanced algorithms, incorporate data from multiple sources, offer global coverage, timely updates, and potential applications across various sectors, making them invaluable data sources.

Previous studies in Bolivia have evaluated the performance of different SREs by comparing them with gauge measurements. These evaluation studies mainly focused on Global Satellite Mapping of Precipitation (GSMaP) [17], Climate prediction center MORPHing (CMORPH) [17], and the Tropical Rainfall Measuring Mission (TRMM) [15,16,17]. Despite their widespread use, the accuracy of CHIRPS and IMERG rainfall estimates across different regions and rainfall regimes of Bolivia remains a subject of ongoing research and evaluation. For instance, in the Desaguadero system between Bolivia and Peru, Marti-Cardona et al. [26] reported an overall high accuracy by CHIRPS in estimating monthly rainfall. Similarly, Satgé et al. [27] compared IMERG and TRMM rainfall estimations with gauge measurements from three watersheds in Bolivia. The results revealed that the accuracy of the SREs in estimating rainfall varied depending on topography [27]. Even though those studies were realized mainly in the Altiplano (Highland) region, the geography and diversity of Bolivian terrain require more research that allows us to understand and delimitate the strengths and limitations of the use of SREs as an alternative or complement to the ground-based gauge measurements.

In this context, this study aims to evaluate the accuracy of CHIRPS and IMERG rainfall estimates at daily and monthly timescales by comparing them with gauge measurements across three regions of Bolivia with distinct topographical settings: Altiplano (Highland), Valles (Midland), and Llanos (Lowland). The CHIRPS and IMERG rainfall datasets were chosen due to their provision of long-term daily and monthly data at relatively high spatial resolutions of 0.05 degrees longitude by 0.05 degrees latitude (approximately 5 km × 5 km) and 0.1 degrees longitude by 0.1 degrees latitude (approximately 10 km × 10 km), respectively. By evaluating the accuracies of these SREs across three different topographies, ranging from mountainous regions to lowlands, we seek to elucidate their strengths and limitations in accurately quantifying rainfall variability across Bolivian complex terrain. The objectives of this study are to: (i) evaluate the rainfall detection capabilities and rainfall amount estimates of new-generation SREs (CHIRPS and IMERG) using categorical and continuous statistics over contrasting topographical regions of Bolivia, (ii) evaluate the accuracy of the SREs by the hit bias (HB), miss bias (MB), and false bias (FB) through bias decomposition, and (iii) demonstrate the utility of the SREs for characterizing spatio-temporal rainfall patterns and meteorological drought occurrence in Bolivia. The findings of this study hold significant implications for improving rainfall monitoring and its effective utilization in Bolivia, ultimately enhancing informed decision-making and supporting sustainable development initiatives across various sectors. This research can contribute to advancing our understanding of how SREs can be effectively utilized for water-related applications, particularly in regions with complex terrain like Bolivia where there is a sparse network of ground-based rain gauge stations.

2. Materials and Methods

2.1. Study Area

Bolivia is located in the central region of South America, between 57.45° and 69.58°W and 9.67° and 22.89°S with an area of about 1.1 million km². Elevation ranges between 72 and 6549 m following an increasing east-west pattern (Figure 1). The diverse geography often leads to the classification of its territory into three major physiographic regions [28]: the Altiplano (Highland), which encompasses high plateaus; the Valles (Midland), characterized by semitropical conditions; and the Llanos which consist of lowland areas. Each of these regions possesses specific characteristics in terms of altitude, air temperature, and water resources, contributing to Bolivia’s rich ecological diversity (

Table 1. Typical characteristics within three distinct zones of Bolivia.

Region	Mean Altitude (m)	Mean Air Temperature (°C)	Maximum Air Temperature (°C)	Minimum Air Temperature (°C)	% of Total Area
Altiplano	3770	15.0	19.2	−4.7	28
Valles	2405	21.8	34.9	6.6	13
Llanos	267	27.3	35.8	13.4	59

Source: Adapted from Evia et al. [28].

). Each region has distinct climate patterns that influence local weather conditions, ecosystems, and human activities. The Altiplano, a high plateau region covering about 28% of Bolivia, extends approximately 805 km in length and 130 km in width with a mean altitude of 3770 m above sea level (Figure 1). The Altiplano lies between the eastern and western ranges of the Andes Mountains. The Altiplano is home to the country’s highest peak, Sajama, and the famous Uyuni Salt Flats. The region is known for its rugged terrain, volcanic activity, and significant mining operations. The Valles region, characterized by semi-tropical valleys, lies to the east and northeast of the Altiplano, with a mean altitude of 2405 m above sea level (Figure 1). This transitional zone includes important agricultural areas, and it is known for its fertile soil and diverse ecosystems. The Llanos region, covering about 60% of the country, consists of lowland areas with a mean altitude of 267 m above sea level (Figure 1). The Llanos region is an extensive region which encompasses the Amazon Basin, the Gran Chaco, and the Pampas. It is characterized by vast tropical rainforests, wetlands, and savannas. Major rivers such as the Mamoré, Beni, and Iténez flow through the Llanos region, contributing to its rich biodiversity and economic activities like agriculture and cattle ranching. The climatic patterns in Bolivia are characterized by a distinct main rainy season extending from December to February and a main dry season (from June to August). We note that the rainfall received in October, November, and March also contributes to the annual total. However, we did not include these months as part of the main rainy season to maintain comparability with previous studies [15].

2.2. Data Sources

2.2.1. Rain Gauge Measurements

The ground-based rainfall data were collected from nine gauging stations strategically selected in the major cities of Bolivia with long historical records (2002–2020). Based on their altitudinal location, these nine gauging stations were categorized into three groups: the Altiplano stations (located at Highland elevations around 4000 m asl), the valley stations (located at Midland elevations approximately 2000 m asl), and the Llanos stations representing Lowland areas (at around 250 m asl). The nine rain gauge stations and the period from 2002 to 2020 were selected to ensure the availability of relatively high-quality, continuous, and independent daily rainfall data, which is essential for evaluating the accuracy of the SREs. The station data acquisition processes were executed via internet access to ensure a conventional scenario applicable to any user. At the rainfall stations shown in Figure 1, the mean annual rainfall during the period 2002–2020 is about 434 mm at the Highland stations, 554 mm at the Midland stations, and 1683 mm at the Lowland stations. The monthly rainfall follows a similar pattern at the Highland, Midland, and Lowland stations, but the amount of rainfall is markedly different with high rainfall amounts received in the Lowland stations (Figure 2).

2.2.2. CHIRPS

CHIRPS is a high-resolution precipitation dataset developed by the Climate Hazards Group at the University of California, Santa Barbara. This dataset provides gridded precipitation estimates on a global scale, incorporating both satellite imagery and ground-based station data. CHIRPS utilizes infrared imagery from geostationary and polar-orbiting satellites to generate precipitation estimates at a high spatial resolution of 0.05 degrees (about 5 km) [23]. The CHIRPS dataset aims to improve understanding and monitoring of precipitation patterns, particularly in regions where ground-based observations are sparse or unreliable. It is widely used in climate research [29,30], drought monitoring [31,32], agriculture [33], and water resource management [34,35]. The CHIRPS dataset is freely available for download and is regularly updated to provide near-real-time precipitation estimates. The CHIRPS version 2.0 daily and monthly datasets were downloaded from https://app.climateengine.org (accessed on 20 January 2023). For a more detailed description of the CHIRPS product, refer to Funk et al. [23].

2.2.3. IMERG

IMERG is a precipitation dataset developed by NASA Goddard Space Flight Center’s Precipitation Processing System. It provides global precipitation estimates at relatively high spatial and temporal resolutions by integrating data from multiple satellite sensors. IMERG combines passive microwave measurements from a constellation of satellites, including the GPM Core Observatory and partner satellites, with infrared data to produce precipitation estimates [36]. The dataset covers the entire globe from 60°N to 60°S latitude with a spatial resolution of approximately 0.1 degrees (about 10 km) and temporal resolutions ranging from 30 min to monthly. IMERG is designed to provide near-real-time precipitation data for a wide range of applications, including weather forecasting, climate monitoring, hydrological modeling, and disaster management. It is particularly valuable in regions with limited ground-based precipitation observations or where extreme weather events occur frequently. The IMERG dataset undergoes continuous updates and improvements through the integration of new satellite data and calibration techniques, ensuring its reliability and accuracy for various scientific and operational purposes [36]. The latest version (V07) daily and monthly IMERG data were downloaded from the GPM website (https://pmm.nasa.gov/data-access/downloads/gpm, accessed on 5 May 2023).

2.3. Evaluation Methods

Various analytical procedures were employed in validating the daily and monthly CHIRPS and IMERG rainfall products against gauge measurements. Categorical validation statistics, continuous validation statistics, and bias decomposition analyses were employed.

2.3.1. Categorical Validation Statistics

The categorical validation statistics were employed at a daily scale to assess the agreement between observed and modeled rainfall. This involved categorizing the data into distinct groups, such as rainy or non-rainy, and subsequently comparing these categories. Four statistical metrics: (i) the probability of detection (POD), (ii) false alarm ratio (FAR), (iii) frequency bias index (FBI), and (iv) Heidke skill score (HSS)) were employed to evaluate the agreement between gauge measurements and the CHIRPS and IMERG rainfall products (

Table 2. A suite of categorical metrics employed to evaluate the agreement between gauge-based rainfall measurements and the CHIRPS and IMERG rainfall products. A, B, C, and D correspond to hits, false alarms, misses, and correct negatives, respectively.

Statistics	Equation	Range	Best Value
POD	$\mathrm{P}\mathrm{O}\mathrm{D}=\frac{\mathrm{A}}{\mathrm{A}+\mathrm{C}}$	0 to 1	1
FAR	$\mathrm{F}\mathrm{A}\mathrm{R}=\frac{\mathrm{B}}{\mathrm{A}+\mathrm{B}}$	0 to 1	0
FBI	$\mathrm{F}\mathrm{B}\mathrm{I}=\frac{\mathrm{A}+\mathrm{B}}{\mathrm{A}+\mathrm{C}}$	0 to $\infty$	1
HSS	$\mathrm{H}\mathrm{S}\mathrm{S}=\frac{2(\mathrm{A}\mathrm{D}+\mathrm{B}\mathrm{C})}{\left(\mathrm{A}+\mathrm{C}\right)\left(\mathrm{C}+\mathrm{D}\right)+(\mathrm{A}+\mathrm{B})(\mathrm{B}+\mathrm{D})}$	$-\mathrm{\infty }$ to 1	1

). We used a rainfall threshold value of 1 mm day⁻¹ for the SREs to distinguish rainfall/no-rainfall events, consistent with previous studies [37,38]. In Table 2, the variables A, B, C, and D represent hits (both the SREs and rain gauge measurements detect rainfall), false alarms (rainfall is detected by the SREs but not by the rain gauge measurements), misses (a rainfall event was not detected even if rainfall was recorded by the rain gauge measurements), and correct negatives (both the SREs and rain gauge measurements detect no rainfall), respectively. POD measures the percentage of actual rainfall days that were correctly identified by the SREs. FAR represents the percentage of days the SREs predicted rain when there was no rainfall. Both POD and FAR range from 0 to 1, where 1 indicates a perfect POD and 0 indicates a perfect FAR. The FBI, which ranges from 0 to ∞, compares the frequency of rainfall-day detection by the SREs to that by rain gauge measurements. An FBI less than 1 indicates an underestimation of rainfall days, while an FBI greater than 1 indicates an overestimation. The HSS, ranging from −∞ to 1, measures the overall accuracy of rainfall-day estimates, accounting for random chance. An HSS below 0 suggests that random chance outperforms the SRE. An HSS of 0 indicates no skill in the SRE, and an HSS of 1 signifies a perfect estimation of rainfall days by the SRE.

2.3.2. Continuous Validation Statistics

The continuous validation statistics were computed both at a daily and monthly scale. This step involved evaluating the quantitative agreement between the amounts of rainfall estimates by CHIRPS and IMERG with gauge-based rainfall measurements. Error statistics such as mean error (ME), mean absolute error (MAE), Index of agreement (d), and bias were analyzed to assess the consistency and accuracy of the rainfall estimates (

Table 3. A suite of statistical metrics utilized in the analysis of the performance measures based on continuous validation statistics.

Statistics	Equation	Range	Best Value
ME	$\mathrm{M}\mathrm{E}=\frac{1}{\mathrm{N}}\sum (\mathrm{S}-\mathrm{G})$	−∞ to +∞	0
MAE	$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{\mathrm{N}}\sum \left|(\mathrm{S}-\mathrm{G})\right|$	0 to +∞	0
d	$\mathrm{d}=1-\frac{\sum {\left(\mathrm{S}-\mathrm{G}\right)}^2}{\sum {\left(\left|\mathrm{S}-\overline{\mathrm{G}}\right|+\left|\mathrm{G}-\overline{\mathrm{G}}\right|\right)}^2}$	0 to 1	1
Bias	$\mathrm{B}\mathrm{i}\mathrm{a}\mathrm{s}=\frac{\sum \mathrm{S}}{\sum \mathrm{G}}$	−∞ to +∞	0

).

Where, S denotes the rainfall total for a satellite product, G represents the rainfall total at a reference gauging station, $\overline{\mathrm{G}}$ is the mean observed rainfall total at a reference gauging station, and N represents the number of data pairs compared.

2.3.3. Bias Decomposition Assessment

Bias decomposition (

Table 4. A suite of statistical metrics utilized to evaluate the overestimated and underestimated errors through bias decomposition. S represents satellite rainfall and G denotes gauge rainfall measurements.

Statistics	Equation
Hit bias (HB)	$\mathrm{H}\mathrm{B}=\sum \left(\mathrm{S}-\mathrm{G}\right),\mathrm{S}>0 \mathrm{a}\mathrm{n}\mathrm{d} \mathrm{G}>0/\sum \mathrm{G}\times 100$
Miss bias (MB)	$\mathrm{M}\mathrm{B}=-\sum \mathrm{G},\mathrm{S}=0 \mathrm{a}\mathrm{n}\mathrm{d} \mathrm{G}>0/\sum \mathrm{G}\times 100$
False bias (FB)	$\mathrm{F}\mathrm{B}=\sum \mathrm{S},\mathrm{S}>0 \mathrm{a}\mathrm{n}\mathrm{d} \mathrm{G}=0/\sum \mathrm{G}\times 100$

) was employed to evaluate the overestimated and underestimated errors of rainfall amounts by the SREs. The ME represents the accumulated estimation errors of SREs, categorized into Hit Bias (HB), Miss Bias (MB), and False Bias (FB). HB arises when both SREs and rain gauge measurements detect rainfall, but the estimated amounts differ, resulting in underestimation or overestimation. HB is positive when the SREs rainfall exceeds the rainfall measured by the gauges, indicating an overestimation by the SREs. Conversely, a negative HB signifies that the SREs are underestimating the rainfall. MB occurs when rain is recorded by rain gauges but not detected by the SREs. MB, which aggregates the missed rainfall by the SREs, is always negative. Conversely, FB happens when SREs indicate rainfall that is not confirmed by the rain gauge measurements and is always positive. Bias decomposition enhances the robustness of our assessment to gain valuable insights into the accuracy of the SREs as well as contributes to more informed decision-making in various fields where bias evaluation is crucial. Results approaching zero signify excellent agreement between SREs and gauge measurements. HB outcomes range from negative infinity to positive infinity, while MB values range from negative infinity to zero. MB and HB quantify the combined depth of both missed and accurately detected rainfall events. Results approaching zero signify excellent agreement, with HB outcomes spanning from negative infinity to positive infinity, while MB values range from negative infinity to zero. Bias is expressed as a percentage of the gauge rainfall amount, enabling direct comparison with biases reported from different stations [39].

2.4. Assessment of Spatio-Temporal Rainfall Patterns

Following the evaluation of the SREs, the annual and seasonal rainfall patterns were assessed based on long-term CHIRPS rainfall estimates (1981–2020). The long-term mean annual and seasonal CHIRPS rainfall maps were calculated to visualize regional spatial and temporal patterns using cell (grid) statistics in a GIS environment. Moreover, the coefficient of variation (CV = σ_R/µ_R, where σ_R is the standard deviation of rainfall and µ_R is the mean rainfall) was used to evaluate the inter-annual and inter-seasonal variation in the rainfall time series. We examined the temporal variability of annual rainfall by computing the standardized rainfall (SR_i = [R_i − µ_R]/σ_R), where SR_i is the standardized annual rainfall for year i, R_i is the annual rainfall in year i, µ_R is the long-term mean annual rainfall (CHIRPS) over the period of assessment (1981–2020) and σ_R is the standard deviation of annual rainfall (CHIRPS) over the period of assessment (1981–2020). The standardized annual rainfall helps identify the dry and wet/normal years [40,41].

2.5. Assessment of Meteorological Drought

In this study, we used the Standardized Precipitation Index (SPI) to assess the occurrence of meteorological drought in Bolivia. The SPI, proposed by [42], has gained widespread use due to its strong theoretical foundation, robustness, and versatility in drought analysis. SPI’s primary advantage is its ability to analyze drought impacts across various temporal scales, allowing for the identification of different drought types [43]. Following the identification of dry and wet/normal years through the analysis of standardized annual rainfall, we investigated meteorological drought occurrence for selected months of the main rainy season (December, January, and February) and selected dry and wet/normal years using SPI. The SPI was computed using long-term (1981–2020) CHIRPS data to demonstrate its utility for drought assessment in data-sparse regions. The steps employed to calculate the SPI were: (i) calculate monthly mean (μ) and standard deviation (σ) of rainfall for the target months; (ii) compute SPI for the target year and months as SPI = (X − µ)/σ, where X is the monthly rainfall of a particular year of interest, μ is the mean rainfall for that month over the period of assessment (1981–2020), and σ is the standard deviation of rainfall for that month over the period of assessment (1981–2020); and (iii) classify and interpret the SPI values into drought categories as: SPI ≥ 2: Extremely wet, 1.5 ≤ SPI < 2: Very wet, 1 ≤ SPI < 1.5: Moderately wet, −1 < SPI < 1: Near normal, −1.5 ≤ SPI < −1: Moderately dry, −2 ≤ SPI < −1.5: Severely dry, and SPI < −2: Extremely dry.

3. Results

3.1. Categorical Validation Statistics

The results from the categorical validation statistics are presented in

Table 5. Detection skill scores for daily rainfall of CHIRPS and IMERG rainfall estimates, encompassing the Lowland, Midland, and Highland stations. The variable N indicates the count of data pairs considered in the comparison. Average performances for each evaluation metric and terrain class are in bold.

		CHIRPS				IMERG
	Station	POD	FAR	FBI	HSS	POD	FAR	FBI	HSS
Highland	Laykacota (N = 6839)	0.45	0.41	0.76	0.28	0.67	0.54	1.45	0.08
	Oruro (N = 6771)	0.44	0.49	0.85	0.30	0.57	0.60	1.44	0.16
	Potosí (N = 5692)	0.47	0.57	1.09	0.29	0.55	0.70	1.79	0.10
	Average	0.46	0.49	0.90	0.29	0.60	0.61	1.56	0.11
Midland	Cochabamba (N = 6894)	0.48	0.49	0.94	0.30	0.65	0.59	1.58	0.18
	Tarija (N = 6771)	0.41	0.52	0.86	0.27	0.54	0.58	1.29	0.21
	Sucre (N = 6921)	0.46	0.53	0.97	0.27	0.53	0.63	1.43	0.10
	Average	0.45	0.51	0.92	0.28	0.57	0.60	1.43	0.16
Lowland	Santa Cruz (N = 6874)	0.37	0.36	0.58	0.26	0.52	0.46	0.97	0.19
	Trinidad (N = 6870)	0.49	0.34	0.74	0.34	0.60	0.51	1.22	0.11
	Cobija (N = 6798)	0.48	0.32	0.71	0.29	0.61	0.42	1.05	0.12
	Average	0.45	0.34	0.68	0.30	0.58	0.46	1.08	0.14

, with values closer to the optimal score highlighted in bold. The probability of detection (POD) indicates a higher performance for the IMERG product over CHIRPS consistently across the three different topographical regions. The skill of the IMERG product to detect the rainfall days was >50% of the observed rainfall days for all stations under investigation. The CHIRPS rainfall product’s skill to detect rainfall day detection was <50% of the observed rainfall days in all stations we considered for evaluation. In contrast, the proportion of rainfall days when there was in fact no rain (false alarm ratio, FAR) for IMERG rainfall was high compared to CHIRPS. That means the CHIRPS product performed better than the IMERG product consistently across the three regions as we evaluated with FAR categorical metrics. The CHIRPS rainfall product in the Lowland region performs better than the Midland and Highland regions with the best score of 32% FAR for Cobija station. The frequency bias index (FBI) value for the IMERG product showed an overestimation by CHIRPS across the three regions with the highest overestimated value of 1.79 at Potosí station. The CHIRPS product with an FBI value of 0.97 was the best value recorded score at Sucre station of the Midland region. Although the CHIRPS rainfall product performed better compared to the IMERG product, there was an underestimation of results in general across all stations. The CHIRPS rainfall product at Santa Cruz station located in Lowland region had the highest underestimated 0.58 value of FBI. Moreover, the assessment of the Heidke skill score (HSS) results showed a better performance for the CHIRPS estimates compared with the IMERG product.

The analysis considering the three distinct geographical zones showed varying degrees of performance between the IMERG and CHIRPS rainfall products. Specifically, in terms of rainfall detection skill, the IMERG product exhibited superior performance, with detection rates of 60% in the Highland, 57% in the Midland, and 58% in the Lowland regions (Table 5). Conversely, the CHIRPS product demonstrated lower detection rates, with 46% in the Highland, 45% in the Midland, and 45% in the Lowland regions. When it comes to detecting non-rain events, the CHIRPS rainfall product outperformed the IMERG product across all three zones. Notably, CHIRPS captured a substantial portion of total rainfall, with 90% and 92% accuracy in the Highland and Midland regions, respectively. In contrast, there was a notable underestimation of 32% in the Lowland region for the CHIRPS rainfall product. On the contrary, the IMERG product tended towards overestimation, revealing 56% and 43% overestimations of total rainfall in the Highland and Midland regions, respectively, but displayed better accuracy with only an 8% overestimation in the Lowland region.

3.2. Continuous Validation Statistics

The results of rainfall totals assessment (error metrics) at daily and monthly timescales are tabulated in

Table 6. Continuous validation statistics at a daily timescale. The variable N indicates the count of data pairs considered in the comparison. Average performances for each evaluation metric and terrain class are in bold.

		CHIRPS				IMERG
	Station	ME	MAE	d	Bias	ME	MAE	d	Bias
Highland	Laykacota (N = 6839)	0.17	1.71	0.58	1.13	0.77	2.08	0.58	1.54
	Oruro (N = 6771)	−0.18	1.33	0.57	0.84	0.13	1.56	0.59	1.10
	Potosí (N = 5692)	−0.14	1.21	0.57	0.87	0.09	1.50	0.56	1.09
	Average	−0.05	1.42	0.57	0.95	0.33	1.71	0.58	1.24
Midland	Cochabamba (N = 6894)	0.15	1.60	0.52	1.13	0.93	2.31	0.48	1.75
	Tarija (N = 6771)	0.15	1.98	0.57	1.10	0.22	2.11	0.60	1.14
	Sucre (N = 6921)	−0.03	1.98	0.64	0.98	−0.23	2.12	0.59	0.87
	Average	0.09	1.85	0.58	1.07	0.31	2.18	0.56	1.25
Lowland	Santa Cruz (N = 6874)	−0.71	3.94	0.57	0.79	0.34	4.78	0.60	1.09
	Trinidad (N = 6870)	0.13	5.23	0.66	1.03	0.67	5.77	0.71	1.13
	Cobija (N = 6798)	0.23	5.77	0.59	1.05	−0.12	5.76	0.67	0.98
	Average	−0.12	4.98	0.61	0.96	0.30	5.44	0.66	1.07

and

Table 7. Continuous validation statistics at a monthly timescale. The variable N indicates the count of data pairs considered in the comparison. Average performances for each evaluation metric and terrain class are in bold.

		CHIRPS				IMERG
	Station	ME	MAE	d	Bias	ME	MAE	d	Bias
Highland	Laykacota (N = 228)	5.63	12.36	0.95	1.13	23.25	24.37	0.89	1.55
	Oruro (N = 228)	−5.58	12.48	0.93	0.85	29.11	32.43	0.81	1.79
	Potosí (N = 213)	−0.71	10.76	0.93	0.98	7.11	10.21	0.96	1.24
	Average	−0.22	11.87	0.94	0.99	19.82	22.34	0.89	1.53
Midland	Cochabamba (N = 228)	4.9	15.27	0.93	1.13	28.28	28.86	0.90	1.75
	Tarija (N = 228)	5.66	16.38	0.94	1.12	7.17	14.34	0.96	1.15
	Sucre (N = 228)	−1.08	17.35	0.44	0.98	−7.15	17.07	0.44	0.87
	Average	3.16	16.33	0.77	1.08	9.43	20.09	0.77	1.26
Lowland	Santa Cruz (N = 228)	−23.17	40.25	0.83	0.79	10.85	26.67	0.95	1.1
	Trinidad (N = 228)	5.65	40.81	0.95	1.04	21.62	35.09	0.97	1.14
	Cobija (N = 228)	8.79	42.3	0.58	1.06	−2.4	33.46	0.58	0.98
	Average	−2.91	41.12	0.79	0.96	10.02	31.74	0.83	1.07

, respectively. It was found that the CHIRPS rainfall product consistently outperformed the IMERG rainfall product within the Highland and Midland regions (Table 6). In the Lowland zone, neither the CHIRPS nor the IMERG rainfall products exhibited a consistently superior performance, as evidenced by the absence of a predominant sound signal (Table 6). The CHIRPS product at Sucre station within the Midland region shows the lowest value of ME (−0.03 mm) followed by the Trinidad station in Lowland region with 0.13 mm of ME (Table 6). On the other hand, IMERG showed lower ME values of 0.09 mm and 0.13 mm at Potosí and Oruro stations of the Highland region, respectively. Similarly, the IMERG product at Cobija station located in the Lowland region showed a lower value of ME (−0.12 mm) compared to the other stations in a similar region (Table 6). In an assessment considering the three different topographic regions, CHIRPS rainfall data represent better performance than IMERG rainfall estimates. Moreover, a difference is presented between the predominance of performance in the analysis of continuous statistical validation considering the daily and monthly scales. In the first case, the predominance in performance is evident for the CHIRPS satellite; however, when the values are aggregated to a monthly scale, IMERG showed a better predominant performance in some indices. Furthermore, the ME and bias index results indicated overestimation of rainfall in the three topographic regions.

When the daily SREs are compared considering the MAE, the results showed better performance for both CHIRPS and IMERG in the Highland and Midland compared to the Lowland stations. The lowest value of MAE (1.21 mm) was found for the CHIRPS rainfall product at Potosí station of the Highland region. Furthermore, ~6 mm of the maximum MAE was found in the Lowland region at Trinidad station for IMERG and Cobija station for CHIRPS rainfall estimates. The index of agreement metrics (d) for CHIRPS was 0.57, 0.58, and 0.61 in the Highland, Midland, and Lowland regions, respectively, whereas corresponding d values for IMERG within these regions were found to be 0.58, 0.56, and 0.66 (Table 6). These results indicate that in the Highland and Midland regions, CHIRPS and IMERG display comparable levels of agreement, with CHIRPS slightly outperforming IMERG in the Midland region. Conversely, IMERG demonstrates higher agreement with ground observations in the Lowland and Highland regions. The bias index evaluation showed a reasonably good agreement between the rainfall estimated by the SREs and the gauge measurements for all nine stations. For instance, the comparison between the altitude groups indicates better performance for CHIRPS in the Highland region than the IMERG rainfall product. Additionally, CHIRPS shows a better performance in both the Midland and Lowland regions when compared to the IMERG dataset (Table 6). In addition to the quantitative comparisons, daily SREs and gauge data were compared visually using scatter plots of daily rainfall depth over Highland, Midland, and Lowland stations (Figure 3). The visual comparison of daily rainfall amounts showed marked scatter with several instances of both over- and underestimation. Several points are spread along the x-axis indicating that the SREs missed several rainy days (S = 0 and G > 0). These missed daily rainfall amounts are mostly within 0–50 mm, but also exceed 50 mm particularly in the Lowland region (Figure 3). In the Midland and Lowland regions, CHIRPS markedly underestimated daily rainfall amounts exceeding 50 mm. Similarly, there are many data points spread along the y-axis indicating false rainy days (S > 0 and G = 0). These false rainfall amounts are mostly within 0–50 mm, with larger false rainfall amounts observed in the Midland and Lowland regions especially for IMERG (Figure 3).

Moreover, to assess whether temporal aggregation improves the performance of both satellite rainfall datasets, continuous statistical validation has been carried out considering a monthly timescale (Table 7). The CHIRPS rainfall product exhibits superior performance, indicated by (low ME), across all three regions when compared to the IMERG product. The IMERG rainfall product at Cobija station in the Lowland region shows a better ME (−2.40 mm) than other stations and CHIRPS data within the same region (Table 7). The MAE at monthly timescale revealed that the rainfall data from the CHIRPS product perform better than IMERG with relatively good agreement with the observed rainfall. Moreover, bias index metrics revealed that the CHIRPS rainfall product performs better in the Highland and Midland regions, whereas the IMERG product performs better in the Lowland region. Monthly SREs and gauge data were also compared visually using scatter plots of monthly rainfall over the Highland, Midland, and Lowland stations (Figure 4). The monthly comparison (Figure 4) shows that the degree of scatter of points markedly reduced compared to the daily comparisons (Figure 3), indicating that the performances of the SREs improved when rainfall estimates were evaluated at a monthly timescale compared with evaluations at a daily timescale. The improvement in the performance of SREs at monthly timescales was also evidenced by high index of agreement values at monthly (Table 7) compared to daily (Table 6) evaluations.

Moreover, both CHIRPS and IMERG were able to represent the seasonal rainfall cycle reasonably well (Figure 5). CHIRPS and IMERG capture the spatial differences in the amounts of rainfall reasonably well, i.e., the Highland and Midland regions receive less rainfall compared with the Lowland region (Figure 5). Figure 5 also shows that IMERG tended to overestimate rainfall, especially in Highland areas, whereas CHIRPS exhibited its highest accuracy in estimating rainfall amounts in such regions. Overall, CHIRPS showed a better performance in estimating the monthly rainfall amounts across the three regions of Bolivia (Table 7 and Figure 5).

3.3. Analysis of Bias Decomposition at a Daily Timescale

The biases of the CHIRPS and IMERG products were decomposed into HB, MB, and FB to assess not only the correctly detected rainfall but also the magnitude of missed rainfall as well as false rainfall detections (

Table 8. Bias decomposition statistics at a daily timescale. The variable N indicates the count of data pairs considered in the comparison. Average performances for each evaluation metric and terrain class are in bold.

		CHIRPS			IMERG
	Station	HB	MB	FB	HB	MB	FB
Highland	Laykacota (N = 6839)	15.3	−37.2	34.8	8.6	−1.3	46.6
	Oruro (N = 6771)	−10.9	−37.5	32.5	−19.9	−4.1	34.2
	Potosí (N = 5692)	−24.1	−28.0	38.8	−32.3	−9.4	50.1
	Average	−6.6	−34.2	35.4	−14.5	−4.9	43.6
Midland	Cochabamba (N = 6894)	10.6	−38.1	40.4	22.5	−4.1	56.1
	Tarija (N = 6771)	2.6	−36.2	43.8	−15.5	−8.6	37.9
	Sucre (N = 6921)	−10.9	−30.3	39.3	−40.0	−8.2	35.0
	Average	0.8	−34.9	41.2	−11.0	−7.0	43.0
Lowland	Santa Cruz (N = 6874)	1.9	−43.3	20.0	10.5	−18.2	17.1
	Trinidad (N = 6870)	6.5	−28.2	24.6	−3.3	−4.6	21.1
	Cobija (N = 6798)	9.4	−30.6	26.0	−16.9	−3.8	18.4
	Average	5.9	−34.0	23.6	−3.3	−8.9	18.9

). The positive and negative values of the HB index indicate rainfall overestimation and underestimation, respectively. For CHIRPS, HB values in the Highland region range from an overestimation of rainfall by 15.3% at Laykacota station to an underestimation by 24.1% at Potosí station. Similarly, HB values for IMERG range from an overestimation of rainfall by 8.6% at Laykacota station to an underestimation by 32.3% at Potosí station. In the Midland region, HB values for CHIRPS range from an overestimation of rainfall by 10.6% at Cochabamba station to an underestimation by 10.9% at Sucre station. For IMERG, HB values range from an overestimation of rainfall by 22.5% at Cochabamba station to an underestimation by 40% at Sucre station. Furthermore, in the Lowland region, HB values for CHIRPS range from an overestimation of rainfall by 1.9% at Santa Cruz station to 9.4% at Cobija station, whereas for IMERG, they range from an overestimation of rainfall by 10.5% to an underestimation by 16.9% at the respective stations. Overall, CHIRPS showed the best performance in terms of HB in the Midland region (an overestimation by 0.8%) whereas for IMERG it was in the Lowland region with an overall underestimation of 3.3%. All the stations have negative MB and positive FB values for both CHIRPS and IMERG. MB values were relatively higher for CHIRPS, ranging from −28% (Potosí in the Highland region) to −43.3% (Santa Cruz in the Lowland) while IMERG showed relatively lower MB values of −1.3% (Laykacota in the Highland) to −18.2% (Santa Cruz in the Lowland). On average, CHIRPS tended to miss about 34% of the rainfall detected by the gauge measurements across the three regions. Overall, IMERG tended to have better performance in terms of the MB index, especially in the Highland region with an average value of about −5%. Table 8 also shows that FB values were relatively higher for IMERG, ranging from 17.1% (Santa Cruz in the Lowland) to as high as 56% in Cochabamba in the Midland. Overall, IMERG showed a high FB value of about 43% in the Midland and Highland regions. For CHIRPS, the FB values range from 20.0% in Santa Cruz (Lowland station) to 43.8% in Tarija (Midland station). Overall, CHIRPS’ FB values were comparable in magnitude with its MB values. It is worth noting that the main component contributing to bias is the one with high magnitude, but MB and FB have opposite signs and cancel each other. The bias decomposition analysis highlights that CHIRPS rainfall estimates generally exhibit a lower deviation from the gauge-based measured rainfall compared to the IMERG rainfall product (Table 8). This suggests that CHIRPS provides a closer estimation of rainfall values in comparison to the observed dataset, across the three topographic regions.

3.4. Assessment of Spatio-Temporal Rainfall Patterns

Since the SREs evaluation results revealed that CHIRPS has a better performance in comparison to IMERG across different terrains of Bolivia, we assessed the spatio-temporal rainfall pattern considering long-term (1981–2020) rainfall data from CHIRPS. The mean annual rainfall ranges from 116 mm to 3103 mm (Figure 6a). The area with the highest rainfall amounts is the Lowland region followed by the Midland while the Highland region receives the least amount of rainfall. About 20% of Bolivia receives a mean annual rainfall between 116 mm and 514 mm, mainly located in the southern parts (60% of the Highland region). The quantile map of mean annual rainfall (Figure 6a) also shows relatively high amounts of rainfall (1686–3103 mm) in the northern parts of Bolivia, covering 40% of the country. Large parts of the Midland region receive a moderate amount of annual rainfall (514–936 mm), with higher values (1686 to 3103 mm) observed in the northeast part (Figure 6a). The CV values of annual rainfall (Figure 6b) were relatively high (>50%) for the southwestern part, while large swaths of Bolivia showed CV values below 50%, indicating the less inter-annual variability of rainfall. The spatial patterns of the main wet season (from December to February) rainfall resemble that of the mean annual rainfall except for the decrease in rainfall values (Figure 6c). During the wet season, about 20% of Bolivia receives seasonal rainfall of more than 764 mm, mainly located in the northern Lowland region, whereas the wet season rainfall over large parts of southern Bolivia is below 436 mm (Figure 6c). The CV for the wet season rainfall (Figure 6d) shows CV values less than 50% indicating less inter-seasonal rainfall variability. Overall, areas with less wet season rainfall have high CV values. During the dry season (from June to August), the western part of the Lowland region presents the highest rainfall values, between 108 and 588 mm (Figure 6e). The CV of dry season rainfall revealed that high variability was observed in large parts of the Lowland region compared to the Midland and Highland regions (Figure 6f).

Figure 7 shows the standardized annual rainfall for each of the three regions. During the period 1981–2020, both negative and positive anomalies of rainfall were observed, indicating the occurrence of wet/normal and dry years. The interannual variability of each region has different trends; strong positive anomalies (years of high rainfall, representing wet/normal years) were observed in 1984 and 2018, especially in the Midland and Highland regions. A long dry period was observed from 1992 to 2000, with the years 1994 and 1995 being the ones with high negative anomalies (dry years). Strong negative anomalies (years of low rainfall/dry years) were also observed in the years 2004 and 2016 (Figure 7). The drought years (1994/1995 and 2015/2016) and the wet/normal year (2017/2018) were selected for further assessment of drought occurrence in Bolivia.

3.5. Assessment of Meteorological Drought

Since CHIRPS outperformed IMERG in estimating rainfall amounts across Bolivia, we investigated meteorological drought considering the CHIRPS rainfall product. An assessment of meteorological drought using SPI was carried out considering the wet months (from December to February) of two dry years (of 1994/1995 and 2015/2016) and one wet/normal year (2017/2018). The SPI results of December 1994 showed that about 20% of Bolivia experienced moderate to severe drought conditions, mainly observed in large parts of the Midland region (Figure 8). The SPI results of January 1995 revealed drought conditions over large swaths of Bolivia, with around 70% of the nation experiencing drought (34% moderately, 31% severely, and 9% extremely dry), mainly in the Lowland and Highland regions (Figure 8). In February of 1995, moderate to severe drought conditions were observed predominantly in the southern parts of the Lowland and Midland regions (Figure 8). In December 2015, moderate to extreme drought conditions were observed in the northern and central parts of Bolivia, and these drought conditions shifted to southwestern Bolivia (15% of the area located in the south of the Highland) in January 2016 and to the southeastern regions (21% of Bolivia) in February 2016 (Figure 8). During the wet/normal year of December 2017, 55% of Bolivia was close to a normal year, with about 32% of the country in northern Bolivia experiencing moderately to extremely wet conditions. In January 2018, SPI results showed values close to normal for about 80% of Bolivia’s territory with specific areas identified as moderately wet (16%) and very wet (3%), which are largely found in the Lowland region. In February 2018, even though large swaths of Bolivia received above-average rainfall, the central parts of the country experienced moderate to severe drought conditions (Figure 8).

4. Discussion

For effective rainfall detection, SREs should exhibit a combination of high POD, high HSS, low FAR, and a balanced FBI [19]. These metrics collectively reflect the product’s ability to accurately identify rainfall while minimizing false alarms and biases, thus enhancing its utility and reliability for users. Table 5 shows the categorical validation statistics indexes (i.e., POD, FAR, FBI, and HSS). The HSS results of CHIRPS and IMERG (Table 5) do not match that criterion, as their values are low for both rainfall products within three different altitudinal zones. However, CHIRPS gave higher HSS scores in comparison with IMERG results. Although it is not possible to assert that both SREs provide an excellent accuracy of rainfall estimates in the regions evaluated, it is possible to establish that the IMERG satellite presents better skill to detect rainfall days than CHIRPS, with values over 50% of POD across the three topographic regions. Our analysis showed that the IMERG rainfall product has higher FAR values than CHIRPS across the three altitudinal regions. The FBI index showed a similar trend in Highland and Midland, with slight underestimation of rainfall for the CHIRPS product and overestimation for IMERG. The Lowland region presented a slight overestimation by IMERG and underestimation by CHIRPS.

The continuous validation statistics were carried out at daily and monthly scales (Table 6 and Table 7). Similar trends for the mean error and bias were observed for both the CHIRPS and IMERG rainfall products at daily and monthly timescales. The trends were different with the MAE and index of agreement; the daily and monthly evaluation results present reasonable performance for CHIRPS, while IMERG showed the best results at monthly evaluations. This agrees well with previous studies [44,45] which concluded that IMERG performs better at monthly than daily scale. It is important to note that the index of agreement result was relatively closer to the best score for monthly than daily assessments. In general, the results showed reasonable performance at daily and monthly scales with an overestimation of rainfall by the SREs, especially in the Highland region. Studies in other regions by Popovych and Dunaieva [46] and Geleta and Deressa [47] also reported a slight overestimation of rainfall by CHIRPS under different topographic settings. Overall, we observed differences in the performance of the SREs depending on the altitudinal ranges of each region. Similarly, [48] reported that the accuracy of CHIRPS estimates was dependent on geographic and climatic characteristics across contrasting topographies of South America. The study conducted in the Highland area of Bolivia (Katari basin) using CHIRPS satellite data [49] concludes that CHIRPS data tend to overestimate rainfall. The ME results for Laykacota station in both CHIRPS and IMERG assessments showed overestimation of rainfall by the SREs, likely due to its geographical proximity to the Katari basin. In contrast to our results, findings from a study by Fenta et al. [19] in Ethiopia showed underestimation by CHIRPS in both Highland and Lowland areas. Another study by Belay et al. [4] reported that a slight overestimation of rainfall occurrence by CHIRPS for the Lowland region and underestimation for the Highland region. Bias decomposition into HB, MB, and FB revealed that MB and FB values were comparable for CHIRPS whereas FB values had higher contribution for the marked overestimation of rainfall by IMERG. Bias decomposition results also suggest better performance for CHIRPS compared with IMERG across the three regions of Bolivia. The types of sensors and retrieval algorithms employed by the SREs could play an essential role in the magnitudes of the HB, MB, and FB indices of CHIRPS and IMERG. Wang et al. [50] reported that SREs that employ passive microwave algorithms (e.g., IMERG) struggle to accurately distinguish between rain clouds and surfaces that emit microwave signals like rainfall. This can cause snow and sand surfaces to be mistaken for rainfall, leading to overestimation of rainfall by IMERG. Additionally, rain detected at higher altitudes by the SREs may evaporate before reaching the ground, particularly in dry regions, further contributing to rainfall overestimation [9,45,51]. The underestimation of rainfall by SREs is likely due to several factors: (i) high-intensity rainfall is often underestimated because the rainfall is averaged over a pixel area, and such events typically occur over spatial domains smaller than the pixel sizes of CHIRPS and IMERG [9,19]; (ii) SREs may struggle to accurately estimate rainfall from warm-rain processes in mountainous regions since the cloud-top temperature tends to be warmer than the rainfall detection thresholds (e.g., CHIRPS), leading to rainfall underestimation [51]; and (iii) the low sampling frequency of the satellites causes SREs to miss short-duration, high-intensity rainfall events [45].

The spatial distribution of average rainfall considering long-term (1981–2020) CHIRPS data revealed a clear difference in rainfall amounts among the three altitudinal regions. The high rainfall areas are in the northeast part of the country in the Lowland region, in contrast to the Highland region which is characterized by low rainfall. A similar pattern of rainfall distribution was presented by Blacutt et al. [15], even though in that assessment the rainfall data came from different satellites (CFR, MERRA, and TRMM3B42). During the wet season (from December to February), the TRMM satellite identified a similar area in the Midland northern part with high rainfall values, whereas during the dry season (from June to August) the areas of low rainfall values agree with our findings for the Highland region [15]. The meteorological drought assessment performed for the critical years (2015/2016 and 1994/1995) in comparison with the normal year 2017/2018 revealed a spatio-temporal variability in the occurrences of drought across Bolivia. Our results agree well with the findings of Vicente-Serrano et al. [43] who reported the occurrence of severe drought across different regions of Bolivia, especially around 1995. Furthermore, the period from 2003 to 2006 was also predominantly dry, albeit with a lower drought magnitude compared to the previous drought events [43]. Also, the selection of the drought years assessed was based on the standardized rainfall analysis, and this criterion proved to be valid compared to the “Bolivian drought monitor” (http://monitorsequias.senamhi.gob.bo/#/data/timeseries, accessed on 12 April 2024). Even though this official document assesses meteorological, hydrological, and agricultural drought, the years 2015/2016 were identified with extreme drought conditions, coinciding with our findings.

5. Conclusions

IMERG demonstrated superior accuracy in detecting rainfall occurrence compared to CHIRPS over Bolivia’s complex terrain. However, IMERG’s higher false alarm ratio presents challenges, particularly in distinguishing true rainfall events from false detections. On the other hand, CHIRPS offered more accurate estimations of rainfall amounts across the three regions (Altiplano [Highland], Valles [Midland], and Llanos [lowland]), showing minimal random errors and relative biases below 10%. Bias decomposition revealed that the high FB values of IMERG significantly contributed to the pronounced overestimation of rainfall, particularly in the Midland and Highland regions. Both CHIRPS and IMERG were able to represent the seasonal rainfall cycle reasonably well. CHIRPS and IMERG depict relatively similar spatial distributions of rainfall, i.e., the Highland and Midland regions receive less rainfall compared with the Lowland region. IMERG tended to overestimate rainfall, especially in Highland areas, whereas CHIRPS exhibited its highest accuracy in estimating rainfall amounts in such regions. The dependence of the SREs’ performance on elevation highlights the importance of considering regional differences in elevation when evaluating SREs, as terrain characteristics can significantly influence rainfall dynamics. Our study demonstrates the effectiveness of CHIRPS data by analyzing spatio-temporal rainfall patterns and evaluating meteorological drought occurrences in Bolivia, where long-term gauge data availability is limited. SREs can be alternative sources of rainfall data to complement gauge data and enhance our ability to assess, monitor, and respond to drought occurrences, ultimately contributing to improved resilience and adaptation to climate variability and change. This study provides valuable insights to choose appropriate SREs for informed decision-making in water-related applications, particularly in poorly gauged regions with complex terrain. We note that the bias levels of the SREs are still too high to ignore and future research needs to focus on correction of the SREs’ bias before they are integrated into operational applications related to water resources management (e.g., water resources potential assessment, water balance studies, etc.).