Integrated Assessment of Soil Loss and Sediment Delivery Using USLE, Sediment Yield, and Principal Component Analysis in the Mun River Basin, Thailand

Abstract

The Mun River Basin, the largest Mekong tributary in Northeast Thailand, has experienced extensive agricultural expansion and forest decline, raising concerns over increasing soil erosion and sediment transfer. This study provides an integrated assessment of soil loss, sediment yield (SY), and sediment delivery ratio (SDR) across 19 sub-watersheds using the Universal Soil Loss Equation (USLE), field-based SY data, and multivariate statistical analyses in 2024. Basinwide soil loss was estimated at ~35 million t y⁻¹ (mean 4.96 t ha⁻¹ y⁻¹), with more than 80% of the basin classified in the no erosion to very low erosion classes. Despite substantial hillslope erosion, only 402,405 t y⁻¹ of sediment reaches the river network, corresponding to a low SDR of 1.15%, which falls within the range reported for large tropical watersheds with significant reservoir infrastructure. Soil loss is most strongly influenced by slope and forested terrain, while SY responds primarily to rainfall and tree plantations; urban land, croplands, and reservoirs act as sediment sinks. Principal Component Analysis (PCA) resolved multicollinearity and produced six components explaining over 90% of predictor variance. A PCA-based regression model predicted SY per unit area with high accuracy (r = 0.81). The results highlight the dominant roles of hydroclimate and land-use structure in shaping sediment connectivity, supporting targeted soil and watershed-management strategies.

1. Introduction

Accelerated soil erosion by water constitutes one of the foremost threats to terrestrial ecosystem integrity, agricultural sustainability, and freshwater quality worldwide [1,2]. Annual global soil loss from agricultural lands is estimated at 35.9 Pg y⁻¹, with erosion rates in many regions exceeding natural soil formation by factors of 10–100 [3,4]. Tropical and subtropical regions experience particularly severe erosion, often 20–50 t ha⁻¹ y⁻¹, in upland agricultural systems, driven by high-intensity monsoonal rainfall, steep terrain, rapid deforestation, and intensive land use [5,6]. Southeast Asia has witnessed a 23% increase in soil erosion from 2001 to 2012, with future projections indicating continued intensification under climate and land-use change scenarios [1,4].

Beyond on-site productivity losses estimated at USD 400 billion annually [7], eroded sediment drives off-site impacts including reservoir siltation (reducing storage capacity and hydropower efficiency), channel aggradation (exacerbating flood risk), and eutrophication of receiving waters [8,9]. Globally, reservoir sedimentation has caused a 13–19% loss of initial storage capacity, with cumulative losses projected to reach 26% by 2050 [10]. In Southeast Asia’s Mekong basin, dam construction has reduced sediment flux to the delta by >50%, accelerating coastal erosion and threatening the livelihoods of 60 million people [11,12].

Quantifying spatial patterns of soil loss, sediment delivery pathways, and delivery ratios (SDR) is therefore essential for targeting conservation investments, designing sediment-management infrastructure, and forecasting downstream sediment loads [13,14,15]. The Universal Soil Loss Equation (USLE) and its revised form (RUSLE) remain the most widely applied empirical frameworks for estimating sheet and rill erosion, particularly in data-limited settings [16,17]. Recent applications in the Mekong basin have estimated mean erosion rates of 5–20 t ha⁻¹ y⁻¹, with localized hotspots exceeding 50 t ha⁻¹ y⁻¹ in steep agricultural terrain [18,19,20].

The Mun River originates in the Khao Yai mountain range in Nakhon Ratchasima Province and flows eastward for approximately 750 km before joining the Mekong River in Ubon Ratchathani Province. The watershed covers approximately 70,587 km² across ten provinces in Northeast Thailand, making it the largest tributary basin of the Mekong River. The basin contributes USD 28 billion (6.83% of national GDP) in 2017 through agricultural production [21]. The tropical monsoon climate generates mean annual rainfall of ~1100 mm, with 85% concentrated in the wet season (May–October) and erosive intensities exceeding 100 mm d⁻¹ during August–September. Forest cover declined from ~40% (1973) to ~12% (2019) while agricultural area expanded to 74%, altering hillslope hydrology and increasing runoff concentration and sediment connectivity [22]. Deforestation and agricultural intensification amplify erosion by reducing protective vegetation, increasing raindrop impact, and accelerating overland flow [23,24]. Current land use comprises lowland rice (50%) with bunded terraces, upland tree plantations (7%), and field crops (16%), with intensive tillage. This mosaic of land use and reservoir infrastructure creates complex sediment sources, pathways, and sinks governing basin-scale export.

The Universal Soil Loss Equation (USLE) estimates long-term average sheet and rill erosion as the product of six factors: rainfall erosivity (R), soil erodibility (K), slope length and steepness (LS), cover management (C), and support practices (P) [16]. Developed from >10,000 plotted years of U.S. data (1930s–1970s), USLE has been globally adapted through regional calibration [17] and applied in the Mekong basin [18], Thai uplands [19], Malaysian oil palm estates [25], and Indonesian watersheds [26]. Limitations include the exclusion of depositional processes, the omission of gully/channel/wind/tillage erosion, temperate-derived factor values, and prediction of long-term rather than event-scale dynamics [9,17,27]. Nevertheless, USLE remains valuable for risk screening and conservation planning in data-limited settings [28,29].

Sediment yield (SY) represents the fraction of gross erosion reaching the watershed outlet [13]. The sediment delivery ratio (SDR = SY/A, where A is gross erosion) quantifies hillslope–channel connectivity and storage effects [13,14], typically decreasing with watershed area from 5 to 50% (<10 km²) to <5% (>1000 km²) [30,31]. Contemporary approaches include empirical relationships with morphometric parameters [32], index-of-connectivity methods [33,34], and distributed routing models (SEDD, InVEST-SDR) [30,35]. Recent advances incorporate dynamic temporal variability [36,37]. In large basins with reservoir networks, trapping efficiency can exceed 80% [10,38,39].

Principal Component Analysis (PCA) transforms correlated predictors into orthogonal components ordered by variance explained [40]. PCA enables dimensionality reduction, multicollinearity mitigation, and the identification of dominant environmental gradients [41,42]. Applications include watershed water quality source identification [41,43], erosion-driver attribution [44], and sediment connectivity assessment [37]. PCA is particularly valuable when strong intercorrelations would destabilize regression models [45]. Component loadings reveal correlations between original variables and PCs [46].

While soil erosion studies exist for individual watersheds across Thailand [47,48], integrated assessments linking modeled hillslope erosion to observed fluvial sediment export remain limited, particularly for large basins (>50,000 km²) with extensive reservoir infrastructure. Previous studies in the Mekong basin have focused either on erosion severity mapping using RUSLE [18,19] or channel sediment monitoring [12,49], but rarely combine both to quantify delivery ratios and sediment connectivity. Furthermore, the relative contributions of natural sediment sinks (hillslope deposition, floodplain storage) versus engineered traps (reservoir sedimentation) remain unquantified for Mekong tributaries.

• This study addresses four key knowledge gaps:

Although soil erosion has been studied extensively in the Mekong region (e.g., [6,50], several key gaps remain regarding the Mun River Basin specifically, as follows:

• Integrated erosion-delivery assessment: Previous studies in the Mekong basin have focused either on hillslope erosion modeling or on channel sediment transport but rarely integrate these components to quantify sediment connectivity [12]. The role of reservoirs in sediment trapping and their contribution to low sediment delivery ratios remains poorly quantified in this region.
• Spatial heterogeneity in controls: The Mun Basin exhibits strong spatial gradients in topography, rainfall, and land use. Understanding how these factors interact across sub-watersheds to control both gross erosion and net sediment delivery is essential for prioritizing conservation interventions.
• Multivariate controls and multicollinearity: Land-use fractions, topographic factors, and hydroclimatic variables are often strongly intercorrelated, making it difficult to isolate their individual effects on erosion and delivery. Statistical methods that can resolve these interdependencies are needed to build robust predictive frameworks.
• Validation and contextualization: Erosion estimates derived from empirical models like USLE require validation against observed sediment yield and comparison with published regional studies to assess their reliability and generalizability.

• Our specific objectives are to:

• Quantify spatial patterns of soil loss across the Mun River Basin using the Universal Soil Loss Equation (USLE) framework;
• Calculate sediment yield (SY) and sediment delivery ratio (SDR) for 19 sub-watersheds using observed discharge and total suspended solids data;
• Identify the dominant hydroclimatic, topographic, and land-use controls on soil loss, sediment yield, and sediment delivery through correlation analysis;
• Apply Principal Component Analysis to resolve predictor multicollinearity and construct a predictive model for sediment yield;
• Compare results with published studies from similar tropical watersheds to validate model performance and contextualize findings.

• We hypothesize the following:

• H1: The low sediment delivery ratio (SDR) in the Mun Basin is primarily controlled by artificial reservoirs rather than natural sediment sinks, resulting in SDR values comparable to other large tropical watersheds with extensive impoundment infrastructure.
• H2: Hillslope soil loss is predominantly controlled by topographic steepness (LS factor) and land cover, whereas sediment yield is more strongly influenced by hydroclimatic forcing (precipitation and runoff) and specific land-use types (e.g., tree plantations and rice fields).
• H3: Principal Component Analysis can effectively resolve multicollinearity among predictors and produce a parsimonious predictive model for sediment yield across sub-watersheds.

• Unique Contributions:

This study represents the first comprehensive, basin-scale integration of USLE-modeled erosion with observed sediment yield across all major sub-watersheds of the Mun River. It provides the following:

• The first quantitative assessment of reservoir sediment trapping efficiency and its contribution to low SDR in the Mun Basin;
• A systematic multivariate framework for disentangling the effects of correlated predictors on erosion and delivery;
• A transferable methodology for data-limited large basins in tropical monsoonal climates.

2. Materials and Methods

2.1. The Mun River Basin

Located in Northeast Thailand’s Isan region, the Mun River watershed (14°00′–16°30′ N, 101°15′–105°30′ E) is the largest tributary system of the lower Mekong River (approximately 70,000 km²); see Figure 1. The basin’s economy is heavily agricultural, generating roughly USD 28 billion in 2017, equivalent to 6.83% of Thailand’s national GDP, which creates substantial seasonal water demands for irrigation and crop production [21]. Four major reservoir systems were constructed between 1960 and 2000: Lam Takhong (storage capacity 314 Mm³), Lam Phra Phloeng (248 Mm³), Lam Mun Bon (220 Mm³), and Sirindhorn (1966 Mm³). These impoundments provide a combined storage volume of approximately 2750 Mm³, serving multiple functions including irrigation supply, domestic water provision, and hydroelectric power generation. Concurrent with infrastructure development, the basin has experienced dramatic land-cover transformation. Between 1973 and 2019, forest cover declined precipitously from approximately 40% to 12%, while agricultural land expanded to occupy 74% of the basin. This conversion fundamentally altered watershed hydrology by reducing infiltration capacity and soil moisture retention, thereby enhancing surface runoff generation and increasing the connectivity of sediment sources to drainage networks [22].

The erosional consequences of deforestation and agricultural expansion are well established: the removal of protective vegetation increases soil exposure to raindrop detachment, while reductions in canopy interception and root reinforcement accelerate overland flow velocity and sediment mobilization [23,24]. Contemporary land use in the Mun Basin reflects this agricultural dominance. Lowland areas are extensively cultivated for rice production (accounting for 50% of the basin area), predominantly utilizing bunded terrace systems that provide some erosion control through ponding and reduced slope length. In contrast, upland zones support tree plantations covering 7% of the basin (primarily eucalyptus, rubber, and oil palm), which maintain sparser ground cover and are often established on sloping terrain. Field crops, including sugarcane, cassava, and maize, occupy 16% of the basin and are typically managed with intensive tillage practices that leave soils vulnerable during fallow periods. This heterogeneous landscape mosaic, combined with extensive reservoir infrastructure, creates a complex spatial arrangement of erosion sources, sediment transport pathways, and deposition sinks that collectively determine basin-scale sediment export dynamics.

Topography and Geology: Elevation ranges from approximately 100 m above sea level (asl) in the eastern lowlands to over 1000 m asl in the western uplands of the Phetchabun and Dangrek mountain ranges. The basin is characterized by gently rolling terrain in the central plains transitioning to steeper slopes in the western and southern margins. Geologically, the region is underlain by Mesozoic sedimentary rocks, including sandstones, siltstones, and shales of the Khorat Group, and overlain by Quaternary alluvial deposits on valley floors [51].

Climate and Hydrology: The basin experiences a tropical savanna climate (Köppen classification Aw) with distinct wet (May–October) and dry (November–April) seasons. Mean annual precipitation is approximately 1000 mm, with strong spatial variability ranging from ~550 mm in the western rain-shadow areas to ~1850 mm in the eastern basin, following a west-to-east increasing gradient [52,53]. Approximately 85–90% of annual rainfall occurs during the southwest monsoon season, resulting in pronounced intra-annual discharge variability. Mean annual temperature is 26–28 °C. The basin’s hydrology is significantly modified by reservoir operations, with major dams at Lam Takhong (capacity 314 Mm³), Lam Phraploeng (130 Mm³), Lam Mun Bon (265 Mm³), and Sirindhorn (1966 Mm³).

Land Use and Socioeconomic Context: As of 2019, agriculture occupies 73.82% of the basin area, dominated by rain-fed rice cultivation (50.17% of total basin area), followed by upland crops such as maize, sugarcane, and cassava (15.66%), and tree plantations including eucalyptus, rubber, and oil palm (7.31%). Forest cover has declined to 12.32% of the basin, concentrated in the western highlands and scattered patches along the Dongrak mountain range. Urban and residential areas account for 6.75%, water bodies 3.37%, and miscellaneous land (including mining, bare soil, and rocky outcrops) 3.74% [54]. The basin supports a population of approximately 7 million people and is central to Thailand’s agricultural economy, particularly rice and cassava production.

2.2. Data Sources

Table 1. Data sources, spatial/temporal resolution, and applications in this study.

Data Type	Source	Resolution/ Period	Purpose
Digital Elevation Model	USGS SRTM	30 m	Topographic analysis, LS factor
Rainfall	Upper and Lower NE Meteorological Centers, TMD	Daily, 2024	R factor calculation
Soil map	Land Development Department (LDD)	1:50,000	K factor assignment
Land use map	Land Development Department (LDD)	1:25,000, 2019	C and P factor assignment
Discharge (Q)	Lower NE Region Hydrological Irrigation Center	Daily, 2024	Sediment yield calculation
Total Suspended Solids (TSS)	Lower NE Region Hydrological Irrigation Center	Monthly, 2024

summarizes the primary data sources used in this study, including spatial resolution, temporal coverage, and data providers.

2.3. Watershed Subdivision and Overall Approach

The Mun River watershed was partitioned into nineteen (19) hydrologically coherent sub-basins to enable a spatially explicit assessment of erosion and delivery processes (Figure 2). Sub-basin delineation was performed using the QGIS (Version 3.44) hydrological analysis toolkit based on the SRTM DEM and validated against existing hydrological station locations. For each sub-basin and for the basin as a whole, we computed the following: (i) gross soil loss using the Universal Soil Loss Equation (USLE); (ii) sediment yield (SY) exported by the channel network; and (iii) the sediment delivery ratio (SDR). Sub-basin discretization facilitates subsequent attribution analysis of the factors controlling soil loss, SY, and SDR across contrasting physiographic and land-use settings.

2.4. Universal Soil Loss Equation (USLE)

The Universal Soil Loss Equation (USLE) represents an empirical modeling framework for quantifying long-term average annual soil loss from sheet and rill erosion processes on agricultural hillslopes [16]. The equation structure, formalized through extensive field experimentation conducted by the United States Department of Agriculture beginning in the 1930s, expresses mean annual erosion as the multiplicative product of six environmental and management factors that have been calibrated and validated across diverse agroecological settings worldwide [17]. The model has demonstrated utility in tropical Asian contexts, with successful applications documented for the Mekong River basin [18], upland watersheds in Thailand [19], oil palm plantations in Malaysia [25], and erosion-prone catchments across Indonesia [26].

USLE computes soil loss (A, t ha⁻¹ y⁻¹) as follows:

$$A=R\times LS\times K\times C\times P$$

(1)

where R represents the rainfall–runoff erosivity factor (MJ mm ha⁻¹ h⁻¹ y⁻¹), which quantifies the erosive power of precipitation and resulting surface flow; K denotes the soil erodibility factor (t h MJ⁻¹ mm⁻¹), reflecting inherent soil susceptibility to detachment and transport; LS is the dimensionless topographic factor integrating slope length and gradient effects; C is the cover-management factor (dimensionless, range 0–1) representing the influence of cropping systems and residue management on soil protection; and P is the support practice factor (dimensionless, range 0–1) accounting for conservation measures such as contour plowing, terracing, or strip cropping.

Despite its extensive validation and widespread adoption, USLE exhibits several well-documented limitations that users must consider when interpreting the results. The framework does not simulate sediment deposition or re-entrainment processes, which can lead to overestimation of sediment delivery to stream channels. Additionally, USLE excludes other important erosion mechanisms, including gully development, streambank erosion, wind-driven sediment transport, and tillage-induced soil displacement [9,17,27]. Factor calibrations were originally derived from temperate-zone agricultural systems in the United States, potentially limiting their direct transferability to tropical environments with different rainfall characteristics, soil mineralogy, and vegetation structure. Furthermore, the model predicts long-term average conditions rather than capturing event-specific erosion dynamics or temporal variability in sediment production. Notwithstanding these constraints, USLE provides a scientifically defensible and operationally practical tool for comparative erosion risk assessment and prioritization of conservation investments, particularly in data-limited regions where more complex process-based models cannot be adequately parameterized or validated [28,29].

The USLE framework, originally documented in USDA Agriculture Handbook 282 [55], has since become the most extensively calibrated and validated erosion prediction system globally [29]. While the model does not explicitly represent depositional processes or alternative erosion pathways such as gully or wind erosion [56], coupling USLE-estimated gross erosion with empirical sediment delivery ratios derived from observed sediment yield provides a transparent and transferable approach for basin-scale sediment budget assessment.

For this study, USLE factor values were derived from spatially explicit datasets and regional calibration guidance specific to the Mun River Basin and Thailand more broadly. Factor assignment followed protocols established by Thailand’s Land Development Department [48], which provides standardized parameter values for tropical monsoonal conditions, supplemented by basin-specific spatial data layers for rainfall, soil properties, topography, and land-use/land-cover characteristics, as detailed in subsequent sections.

2.4.1. Rainfall Erosivity Factor (R)

The R factor expresses the combined effect of rainfall kinetic energy and intensity on detachment and transport capacity. Daily rainfall records for 2024 were compiled from 47 meteorological stations operated by the Upper Northeastern and Lower Northeastern Meteorological Centers of the Thai Meteorological Department [52,53]. Spatial interpolation was performed using the Inverse Distance Weighting (IDW) method in QGIS with a search radius of 50 km to generate a continuous rainfall surface.

Mean annual rainfall over the Mun Basin in 2024 was approximately 1000 mm, ranging from ~550 mm in the western rain-shadow areas to ~1850 mm in the eastern basin, reflecting the influence of the southwest monsoon and orographic effects. Following the methodology of the Land Development Department [48], which adapted [16] for Thai conditions, R was estimated from annual rainfall X (mm y⁻¹) using the following:

$$R=0.4669X-12.1415$$

(2)

Justification of R Factor Formula: This simplified relationship was developed by the Land Development Department based on calibration against measured erosivity values from 12 meteorological stations across Thailand [48]. While the original Wischmeier formulation uses rainfall intensity data from pluviographs, such high-resolution data are not available for most stations in Northeast Thailand. The formula used in [48] has been widely adopted for erosion assessment in Thailand and neighboring countries with similar monsoonal climates (e.g., [57,58]) and provides a practical approximation when only daily rainfall totals are available.

Across the 19 sub-basins, R ranged from 198 to 852 MJ mm ha⁻¹ h⁻¹ y⁻¹, with a mean of 447.76 and a median of 431.0. Spatial rainfall patterns and derived R values are shown in Figure 3.

2.4.2. Slope Length and Steepness Factor (LS)

Topographic control was represented by the compound LS factor, which scales erosion relative to a reference plot of 22.1 m length and 9% gradient. The LS factor was derived from the 30-m resolution SRTM digital elevation model [59] using the algorithm of Desmet and Govers [60], implemented in QGIS via the r.slope.length module. This algorithm accounts for flow convergence and divergence and has been validated for complex terrain [61]. The Desmet and Govers [60] formulation is as follows:

$$L{S}_{ij}={\left(A_{ij}+D^2\right)}^{m+1}-{\left(22.13X_{ij}\right)}^m\left({\displaystyle \frac{A_{ij}^{m+1}}{D^{m+2}}}\right),$$

(3)

where A_ij is the contributing area per unit width (m²/m), D is the cell size (m), X_ij is a shape factor accounting for flow direction, and m is an exponent dependent on slope steepness. For slopes < 1%, m = 0.2; for 1–3%, m = 0.3; for 3–5%, m = 0.4; and for >5%, m = 0.5 [62].

Over the Mun Basin, LS values ranged from 0.03 in flat alluvial plains to 26.54 in steep western uplands, with a mean of 0.54 and a median of 0.20, indicating that most of the basin consists of low-slope terrain conducive to agriculture.

2.4.3. Soil Erodibility Factor (K)

The parameter K quantifies the inherent susceptibility of soils to detachment and transport as a function of texture, organic matter content, structure, and permeability [16]. K values were assigned by soil series from the national soil map (1:50,000 scale) provided by the Land Development Department, referencing the Thailand-specific K factor table developed by [48].

The [48] K values were derived from field measurements and calibration against USLE plot data across Thailand’s major soil groups. For the Mun Basin, dominant soil types include sandy loams and loamy sands (Yasothon series, K ≈ 0.25–0.28); silty clays and clay loams (Korat series, K ≈ 0.20–0.24); and alluvial soils in valley bottoms (Roi Et series, K ≈ 0.18–0.22). For non-eroding surfaces (e.g., rock outcrops and open water) K was set to 0. The basin-wide mean and median K were 0.23 and 0.25 t h MJ⁻¹ mm⁻¹, respectively.

Limitations: This study uses static K values assigned by soil series and does not account for temporal variability in soil organic matter or structural changes due to management. While dynamic K estimation requires repeated soil sampling across the basin (infeasible at this scale), the [48] values represent long-term average conditions and are appropriate for strategic erosion assessment.

2.4.4. Cover-Management Factor (C)

Parameter C reflects the ratio of soil loss under existing vegetation and management to that from bare soil, integrating crop type, canopy density, residue cover, tillage practices, and seasonality. C values were estimated from the 2019 land-use map [54] using the Thailand-specific C factor table [48]. Land-use classification and assigned C values were as follows:

• Rice fields (A1): C = 0.28 (reflecting seasonal bare soil exposure during land preparation and post-harvest).
• Other upland crops (A2; maize, sugarcane, cassava): C = 0.40–0.60 depending on crop type and management intensity.
• Tree plantations (A3; eucalyptus, rubber, teak, oil palm): C = 0.10–0.20 (lower than upland crops but higher than natural forest due to understory clearing).
• Orchards (A4): C = 0.15.
• Other agriculture (A5; aquaculture, farm buildings): C = 0.05–0.10.
• Forest (F): C = 0.001 (reflecting high protective capacity of natural forest).
• Urban/residential (U): C = 0.01 (impervious surfaces and managed landscapes).
• Water bodies (W): C = 0.
• Miscellaneous (M: bare soil, mining, rocky areas): C = 0.50–0.80.

Across the basin, C ranged from 0 to 0.8, with a mean of 0.24 and a median of 0.28. The spatial distribution of C values is shown in Figure 4.

Data Year Clarification: The 2019 land-use map [54] represents the most recent comprehensive land-use survey available for the Mun Basin. While rainfall and discharge data are from 2024, land-use change between 2019 and 2024 in the Mun Basin has been relatively minor (<2% change in major land-use classes based on comparison with 2022 satellite imagery). The use of 2019 land use for 2024 erosion estimation is consistent with the long-term average approach of USLE and introduces minimal bias compared to other sources of uncertainty (e.g., spatial interpolation of rainfall, soil heterogeneity within map units).

2.4.5. Support Practice Factor (P)

Parameter P represents the effectiveness of conservation practices, such as contouring, strip cropping, and terracing, in reducing soil loss relative to up-and-down-slope cultivation. The P factor applies only to agricultural lands where specific erosion-control measures are implemented; for non-agricultural lands, P equals 1.0 by definition (no conservation practice) [16]. P values were assigned based on field surveys and expert consultation with local agricultural extension officers, following the guidelines in [48]:

• Rice fields with bund terraces: P = 0.10 (highly effective sediment trapping).
• Upland crops with contour plowing: P = 0.50.
• Tree plantations without conservation measures: P = 1.0.
• All non-agricultural lands: P = 1.0.

Across the basin, P ranged from 0.10 to 1.0, with a mean of 0.43 and a median of 0.10, reflecting the dominance of rice cultivation with bunded terraces. Higher P values denote less effective erosion control; lower values indicate stronger conservation practices. Spatial patterns of land use are shown in Figure 4, and land-use distribution by sub-watershed is summarized in

Table 2. Land use distribution by sub-watershed, expressed as percentage of sub-watershed area.

Sub Watershed	Agriculture Area (A)						Urban (U)	Forest (F)	Water (W)	Miscellaneous (M)
	Rice (A1)	Others Crop (A2)	Tree Plantation (A3)	Orchard (A4)	Others (A5)	Sum
1	12.62	37.66	5.47	1.01	2.10	58.87	15.42	16.75	2.21	6.76
2	14.43	32.85	6.43	0.25	1.03	54.99	7.22	30.67	2.22	4.90
3	37.25	45.32	1.70	0.03	0.68	84.98	6.55	1.70	2.48	4.29
4	16.92	28.46	4.59	0.09	0.49	50.54	7.41	36.52	2.44	3.09
5	58.97	22.55	2.84	0.03	0.59	84.98	7.91	0.38	3.06	3.67
6	27.48	53.28	4.46	0.11	0.59	85.92	5.94	4.48	2.30	1.36
7	33.85	25.31	7.53	0.09	0.36	67.14	6.03	22.09	3.19	1.54
8	71.29	9.73	2.05	0.02	0.83	83.92	6.82	2.52	3.76	2.98
9	60.56	12.93	6.90	0.02	0.55	80.96	6.53	3.86	4.57	4.07
10	57.92	10.73	8.24	0.02	0.32	77.23	8.41	7.75	4.28	2.33
11	74.07	3.98	2.77	0.02	1.18	82.02	6.27	5.15	3.28	3.29
12	65.73	7.64	7.54	0.02	0.11	81.03	6.40	7.76	3.17	1.64
13	61.26	5.61	8.70	0.06	0.07	75.70	6.53	13.24	2.62	1.90
14	62.11	2.11	4.07	1.67	0.43	70.40	10.29	6.77	4.58	7.96
15	48.39	6.51	19.07	0.14	0.10	74.21	5.00	16.13	2.49	2.18
16	48.79	10.26	13.50	0.12	0.28	72.94	5.07	16.70	1.82	3.47
17	40.37	5.49	11.58	0.01	0.14	57.59	3.64	23.57	7.64	7.56
18	67.57	6.08	4.61	0.03	0.21	78.50	4.71	10.42	2.17	4.20
19	57.17	12.50	8.44	0.07	0.25	78.42	5.28	8.27	2.79	5.23
Mun basin	50.17	15.66	7.31	0.17	0.52	73.82	6.75	12.32	3.37	3.74

as percentage cover.

2.4.6. Limitations of USLE and Rationale for Coupling with Observed Sediment Yield

USLE predicts gross soil loss (detachment and transport capacity) at the hillslope scale but does not simulate depositional processes within hillslopes, channels, or floodplains [27]. Consequently, USLE-estimated soil loss (A) cannot be directly equated with the sediment exported from the basin. In recognition of this limitation, researchers have proposed alternative models such as USPED (Unit Stream Power Erosion Deposition model; ref. [63]) and MUSLE (Modified USLE; ref. [64]), which incorporate topographic indices of sediment routing and event-based runoff, respectively.

However, USPED requires high-resolution topographic data and detailed flow-routing parameterization, which are challenging to implement reliably at the basin scale of the Mun River (~70,000 km²) without extensive field calibration. MUSLE requires event-based discharge and sediment data, which are not routinely monitored at the sub-watershed scale in the Mun Basin. In contrast, USLE coupled with observed sediment yield (SY) provides a pragmatic framework: USLE identifies spatial patterns of hillslope erosion risk, while observed SY quantifies actual sediment export, and their ratio (the sediment delivery ratio, SDR) integrates the effects of deposition, channel storage, and trapping efficiency across the basin [15].

This approach has been widely adopted in large-basin assessments where process-based sediment routing models are data-limited (e.g., refs. [9,65]). We therefore use USLE for spatial risk mapping and couple it with observed SY to estimate SDR and infer the role of sediment sinks (natural and anthropogenic) in modulating sediment connectivity.

2.5. Sediment Yield and Sediment Delivery Ratio

Sediment yield (SY) represents the mass of sediment conveyed past a watershed outlet over a specified time period, integrating the complex interactions of hillslope erosion, intermediate storage in depressions and floodplains, channel routing processes, and delivery mechanisms that collectively determine basin-scale sediment export [13]. Unlike the gross erosion estimated by USLE, which quantifies potential soil detachment and transport at the hillslope scale, sediment yield reflects only the fraction of mobilized sediment that successfully navigates the hillslope-channel continuum to reach downstream monitoring points. The sediment delivery ratio (SDR), defined as the ratio of observed sediment yield to gross erosion (SDR = SY/A, where A is USLE-estimated erosion), provides a quantitative measure of watershed-scale sediment connectivity and the cumulative effects of storage in depressions, floodplains, channel beds, riparian zones, and artificial impoundments [13,14].

Empirical studies demonstrate that SDR exhibits systematic scaling relationships with watershed areas, typically decreasing as basin size increases due to the greater opportunities for intermediate sediment storage and longer sediment residence times. For small watersheds (<10 km²), SDR values commonly range from 5–50%, whereas large basins (>1000 km²) frequently exhibit SDR < 5% [30,31]. Contemporary approaches to SDR estimation span a methodological continuum from simple empirical relationships based on drainage area or morphometric indices [32], through index-of-connectivity frameworks that explicitly link upslope erosion potential to downslope deposition likelihood using topographic gradients, land-cover patterns, and flow-path characteristics [33,34], to spatially distributed sediment-routing models such as SEDD (Sediment Delivery Distributed) and InVEST-SDR that simulate erosion-deposition processes across watershed landscapes [30,35]. Recent modeling advances increasingly emphasize dynamic temporal variability in sediment connectivity, incorporating threshold-dependent rainfall–runoff responses, seasonal vegetation phenology, and episodic sediment pulse delivery [36,37]. In large basins containing extensive reservoir networks, artificial impoundments can intercept the majority of upstream sediment flux, with trapping efficiencies frequently exceeding 80% and thereby profoundly suppressing basin-scale SDR [10,38,39].

For this study, sediment yield was quantified empirically for each of the 19 sub-watersheds using direct measurements of total suspended solids (TSS) concentrations and stream discharge (Q) obtained from hydrological monitoring stations operated by the Lower Northeastern Region Hydrological Irrigation Center [66]. TSS samples were collected at monthly intervals following standardized field protocols [67], with subsequent laboratory analysis being conducted gravimetrically to determine suspended sediment concentrations. Continuous discharge measurements were derived from stage–height observations converted to volumetric flow rates using site-specific stage–discharge rating curves that are annually recalibrated to account for channel geometry changes. Annual sediment yield for 2024 was computed by integrating the product of instantaneous TSS concentration and discharge over all measurement intervals:

$$SY={\displaystyle \int \left(TSS\times Q\right)dt}\approx {\displaystyle \sum \left(TSS\times Q\times \Delta t\right)},$$

(4)

where TSS is the total suspended solids concentration (g m⁻³), Q is discharge (m³ s⁻¹), Δt is the time interval (seconds), and the summation is made over all measurement periods in 2024.

The sediment delivery ratio quantifies the efficiency of sediment transport from hillslopes to the watershed outlet, expressed as follows:

$$SDR\left(\%\right)={\displaystyle \frac{\left(SY/\mathrm{Area}\right)}{A}}\times 100,$$

(5)

where SY is the observed sediment yield (t y⁻¹) and A is the total soil loss estimated by USLE for the sub-watershed (t ha⁻¹ y⁻¹). This ratio quantifies the efficiency of sediment transport from erosion sources to the basin outlet, integrating the combined effects of floodplain deposition, in-channel storage, riparian vegetation filtering, and reservoir trapping. Low SDR values (<5%) indicate strong sediment buffering capacity and limited connectivity between hillslope sources and channel networks, while elevated SDR values (>10%) suggest efficient sediment transport pathways with minimal intermediate storage.

2.6. Correlation Analysis

We evaluated linear associations between independent variables: watershed area (km²), precipitation (mm y⁻¹), runoff (Mm³ y⁻¹), land-use proportions (%),topography (LS), and the dependent variables: USLE soil loss (A, t ha⁻¹ y⁻¹), SY (t y⁻¹), and SDR (%). To normalize for basin size, SY was also expressed per unit area (SY/Area; t ha⁻¹ y⁻¹).

Pearson’s correlation coefficient (r) quantifies the strength and direction of linear association and was calculated using the standard formula implemented in Python’s SciPy library (Version 1.17.0). Values of r approach +1 for a strong positive association, −1 for a strong negative association, and 0 for weak or no linear relationship. Statistical significance was assessed at α = 0.05 using a two-tailed t-test.

A correlation matrix was constructed to systematically examine all pairwise relationships among variables and to identify potential multicollinearity among predictors (high inter-predictor correlations that can destabilize regression models). This matrix informed the subsequent application of Principal Component Analysis to resolve multicollinearity.

2.7. Principal Component Analysis (PCA) and Predictive Modeling

2.7.1. Purpose and Theory of PCA

Principal Component Analysis (PCA) is a multivariate statistical technique that transforms potentially correlated predictor variables into a reduced set of orthogonal principal components (PCs) ordered by variance explained [40]. Each PC represents a linear combination of original variables, with PC1 capturing maximum variance, PC2 capturing maximum remaining variance orthogonal to PC1, and subsequent components following the same pattern. In watershed studies, PCA serves three primary functions: dimensionality reduction to distill complex multi-variable datasets into 2–5 interpretable components, multicollinearity mitigation for stable regression modeling, and the exploratory identification of dominant environmental gradients [41,42].

The method has proven particularly valuable in watershed applications including pollution source identification from multiple water-quality constituents [41,43], the attribution of erosion drivers to distinguish climate versus land-use effects [44], and sediment connectivity assessment [37]. PCA is especially suited to erosion and sediment-yield modeling, where predictor variables often exhibit strong intercorrelations (such as forest cover and topographic slope factor (LS), both associated with steep terrain, or agricultural area and precipitation, both correlated with sediment delivery) that would otherwise destabilize ordinary least-squares regression by inducing multicollinearity [45]. By constructing orthogonal components, PCA eliminates these correlations and enables parsimonious models that retain explanatory power while avoiding overfitting. Interpretability depends on examining component loadings, which are correlations between original variables and PCs that reveal the dominant environmental gradients structuring the data [46]. This approach has been widely applied in erosion and sediment transport studies to integrate multiple environmental controls [68,69].

2.7.2. PCA Implementation

PCA was applied to 14 predictor variables: watershed area (v01), annual precipitation (v02), runoff (v03), rice fields (v04), other crops (v05), tree plantations (v06), orchards (v07), other agriculture (v08), total agriculture (v09), urban area (v10), forest area (v11), water bodies (v12), miscellaneous land (v13), and LS factor (v14). All variables were standardized (mean = 0, standard deviation = 1) prior to analysis to ensure equal weighting.

Validation of PCA Applicability: To confirm that the dataset is suitable for PCA, we computed the following: (1) Kaiser–Meyer–Olkin (KMO) Measure of Sampling Adequacy, where KMO quantifies the proportion of variance among variables that might be common variance. KMO > 0.6 is considered adequate; KMO > 0.8 is excellent [70]. For our dataset, KMO = 0.74, indicating good suitability. (2) Bartlett’s Test of Sphericity: This test evaluates whether the correlation matrix is significantly different from an identity matrix (i.e., whether variables are sufficiently intercorrelated). Bartlett’s test yielded χ² = 287.5 (p < 0.001), confirming significant intercorrelations suitable for PCA.

PCA was performed using Python’s scikit-learn library. Eigenvalues, proportion of variance explained, and eigenvectors (loadings) were extracted. Components with eigenvalues > 1.0 (Kaiser criterion) and cumulatively explaining > 90% of variance were retained for regression modeling.

2.7.3. Interpretation of Principal Components

Each retained PC was interpreted based on the magnitude and sign of its loadings on the original variables. High positive loadings indicate variables that increase with the PC; high negative loadings indicate variables that decrease. This allows PCs to be characterized as environmental gradients (e.g., agricultural intensity, topographic steepness, hydrological forcing).

2.7.4. PCA-Based Regression Model

The retained PCs were used as independent variables in a linear regression model to predict SY/area:

$$SY/\mathrm{Area}={\alpha }_0+{\alpha }_1\left(P{C}_1\right)+{\alpha }_2\left(P{C}_2\right)+\cdots +{\alpha }_N\left(P{C}_N\right),$$

(6)

where $P{C}_i$ is principal component i and ${\alpha }_i$ is the regression coefficient of a term $P{C}_i$. Model parameters were optimized by minimizing the sum of squared errors (SSE) between observed and modeled SY/area using the SciPy minimize routine, which by default employs the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm; alternative solvers (e.g., Nelder–Mead, Powell) are also available within the same framework. Model performance was evaluated using the following:

• Pearson correlation coefficient (r) between observed and predicted values.
• Coefficient of determination (R²).
• Root mean square error (RMSE).

Statistical significance of regression coefficients was assessed using t-tests (α = 0.05).

3. Results

3.1. Spatial Patterns of Erosion, Sediment Export, and Basin-Scale Connectivity

Spatial patterns of modeled soil loss are presented in Figure 5. Soil erosion rates were classified into six categories (shown in

Table 3. Area distribution and percentage of basin by soil loss severity class.

Category	Soil Loss Rate (t ha−1 y−1)	Area
		km2	Percentage
No erosion	0 to 0.1	13,013	18.44
Very low	0.1 to 6.7	46,657	66.10
Low	6.7 to 11.2	3807	5.39
Moderate	11.2 to 22.4	3910	5.54
High	22.4 to 33.6	1415	2.00
Severe	more than 33.6	1785	2.53
Total		70,587	100.00

) based on the soil loss tolerance framework of [16] and adapted for Thai conditions [48]. The first category, “no erosion” (0.0–0.1 t ha⁻¹ y⁻¹), corresponds to water bodies, urban areas, and rocky land, covering 18.44% of the watershed area (13,013 km²). The second category, “very low soil loss” (0.1–6.7 t ha⁻¹ y⁻¹), encompasses most of the basin (66.10%; 46,657 km²), reflecting extensive rice cultivation with effective conservation practices (bunded terraces) and low-slope topography. Subsequent categories, low (6.7–11.2 t ha⁻¹ y⁻¹), moderate (11.2–22.4 t ha⁻¹ y⁻¹), high (22.4–33.6 t ha⁻¹ y⁻¹), and severe (>33.6 t ha⁻¹ y⁻¹), cover progressively smaller areas (5.39%, 5.54%, 2.00%, and 2.53%, respectively) and are concentrated in the western uplands and areas of intensive upland cropping.

Total annual soil loss for the Mun River Basin is approximately 35.05 million t y⁻¹, corresponding to a mean rate of 4.96 t ha⁻¹ y⁻¹ and a median of 0.58 t ha⁻¹ y⁻¹. The large difference between the mean and median indicates a right-skewed distribution, with most of the basin experiencing very low erosion but localized hotspots contributing disproportionately to the total sediment budget.

Sub-watershed-level results are summarized in

Table 4. Sub-watershed characteristics including soil loss, sediment yield (SY), sediment delivery ratio (SDR), and hydrological parameters.

Sub-Watershed	Area (km2)	Soil Loss, A (t ha−1 y−1)		Total Soil Loss (t y−1)	Mean TSS (g m−3)	Mean Q (m3 s−1)	SY (t y−1)	SDR (%)
		Mean	Median
1	3420	11.28	2.83	3,800,437.59	33.28	0.97	1020.49	0.03
2	2319	12.53	1.22	2,908,562.04	249.45	1.72	13,521.58	0.46
3	2888	4.28	0.62	1,230,391.74	113.54	3.18	11,403.37	0.93
4	2897	5.96	0.53	1,730,916.12	59.21	1.65	3090.07	0.18
5	1061	2.12	0.26	225,389.33	44.97	1.50	2124.51	0.94
6	1654	4.97	2.92	823,776.26	198.43	1.39	8685.00	1.05
7	5934	4.56	0.48	2,714,974.45	741.15	0.69	16,211.88	0.60
8	3666	1.35	0.39	493,101.70	126.83	2.53	10,140.84	2.06
9	7869	2.30	0.48	1,815,590.51	55.80	22.26	39,207.52	2.16
10	4966	2.44	0.47	1,217,923.62	47.04	7.82	11,610.53	0.95
11	4443	1.20	0.40	533,742.60	73.74	6.05	14,075.15	2.64
12	3803	2.15	0.48	823,358.45	38.99	6.90	8492.45	1.03
13	3376	3.11	0.40	1,054,535.39	46.03	4.25	6177.12	0.59
14	2681	1.88	0.51	495,647.91	1.66	210.05	10,997.06	2.22
15	3360	5.14	0.65	1,738,620.01	60.77	22.09	42,358.02	2.44
16	4840	8.12	1.07	3,963,462.96	68.23	42.34	91,164.17	2.30
17	4041	14.40	0.99	5,768,595.93	81.09	5.25	13,437.30	0.23
18	3493	4.04	0.75	1,422,984.74	74.54	12.33	29,011.10	2.04
19	3878	5.94	1.16	2,288,298.08	36.44	60.60	69,677.26	3.04
Mun Watershed	70,587	4.96	0.58	35,050,309.42	30.83	413.59	402,405.42	1.15

. Mean soil loss rates range from 1.20 t ha⁻¹ y⁻¹ in sub-watershed 11 (flat, rice-dominated) to 14.40 t ha⁻¹ y⁻¹ in sub-watershed 17 (steep, mixed agriculture and forest) (Figure 6). Most sub-watersheds fall within the “very low” erosion class, with exceptions: sub-watersheds 1, 2, and 17 exhibit “moderate” erosion (11.2–22.4 t ha⁻¹ y⁻¹).

Sediment Yield: Total suspended solids (TSS) concentrations and discharge (Q) data from Royal Irrigation Department (RID) hydrological stations (Table 4) were used to calculate annual sediment yield (SY). Basin-wide SY is 402,405 t y⁻¹, with sub-watershed values ranging from 1020 t y⁻¹ (sub-watershed 1, downstream of Lam Takhong Dam) to 91,164 t y⁻¹ (sub-watershed 16, large drainage area with mixed land use). Notably, sub-watershed 1’s low SY despite moderate soil loss reflects efficient sediment trapping by Lam Takhong Reservoir (capacity 314 Mm³), confirming the dominant role of impoundments in reducing sediment delivery.

Sediment Delivery Ratio: The basin-wide SDR is 1.15%, indicating that only approximately 1.15% of hillslope-mobilized sediment reaches the river network. Sub-watershed SDR values range from 0.03% (sub-watershed 1) to 3.04% (sub-watershed 19), with most values between 0.5% and 2.5%. These low SDR values are consistent with the following:

• Extensive sediment trapping by major reservoirs (Lam Takhong, Lam Phraploeng, Lam Mun Bon, Sirindhorn).
• Deposition in floodplains and rice paddies with bunded terraces.
• Large drainage area (>70,000 km²), which generally correlates with lower SDR [13]).

3.2. Model Validation and Comparison with Tropical Watersheds

To contextualize the Mun Basin results and assess model reliability, we compared our findings with published studies from tropical watersheds with similar characteristics (

Table 5. Comparison of mean soil loss and sediment delivery ratio (SDR) with published studies from tropical watersheds.

Study	Location	Watershed Area (km2)	Erosion Model	Mean Soil Loss (t ha−1 y−1)	SDR (%)	Key Features
This study	Mun Basin, Thailand	70,587	USLE	4.96	1.15	Multiple large reservoirs, extensive rice terraces
Valentin et al. [6]	Northern Laos	850	RUSLE	8.2	3.5	Steep terrain, slash-and-burn agriculture
Nampak et al. [50]	Northeastern Thailand	3200	USLE	6.7	2.1	Mixed agriculture and forest
Kondolf et al. [12]	Lower Mekong	795,000	Sediment budget	—	15–25	Natural sediment connectivity, minimal impoundments
Vanmaercke et al. [65]	Global compilation (large basins)	>10,000	Various	—	5–30	Meta-analysis; higher SDR in basins without major reservoirs

). Our SDR of 1.15% falls at the lower end of the range for large tropical basins, consistent with the hypothesis (H1) that extensive reservoir infrastructure is the dominant control on low sediment delivery in the Mun Basin. In contrast, the Lower Mekong mainstream [12] exhibits an SDR of 15–25% prior to dam construction, and smaller Mekong tributaries without major dams (e.g., [6]) show an SDR of 3–5%. Our results align closely with [50], who reported an SDR~2% for a similar northeastern Thai watershed with moderate impoundment.

The mean soil loss rate of 4.96 t ha⁻¹ y⁻¹ is within the range of previous USLE/RUSLE applications in Southeast Asia (3–10 t ha⁻¹ y⁻¹; [29,58]) and falls below the typical soil-loss tolerance threshold of 6.7–11.2 t ha⁻¹ y⁻¹ for most soil types [16], suggesting that basin-wide erosion is generally sustainable at current rates.

3.3. Environmental Controls on Soil Loss and Sediment Delivery

Pearson’s correlation coefficients between dependent variables (soil loss rate A, SY/Area, SDR) and independent variables are presented in

Table 6. Pearson correlation coefficients (r) between dependent variables (soil loss, SY/area, SDR) and independent environmental variables (bold indicates |r| > 0.45 and p < 0.05).

Independent Variable	Soil Loss Rate, A (USLE)	SY/Area	SDR
Watershed Area	−0.15	−0.22	+0.19
Precipitation	+0.28	+0.59	+0.47
Runoff	+0.12	+0.31	+0.38
Rice (A1)	−0.71	−0.18	+0.66
Other Crops (A2)	+0.27	−0.22	−0.50
Tree Plantation (A3)	+0.35	+0.55	+0.18
Orchard (A4)	+0.09	−0.05	−0.12
Other Agriculture (A5)	+0.24	−0.34	−0.25
Total Agriculture	−0.73	−0.29	+0.54
Urban (U)	−0.38	−0.45	−0.22
Forest (F)	+0.67	+0.21	−0.49
Water (W)	−0.18	−0.28	−0.19
Miscellaneous (M)	+0.49	+0.12	−0.08
LS Factor	+0.62	+0.19	−0.47

and Figure 7 (correlation matrix heatmap). Key findings are described as follows:

Soil Loss: Strongly positively correlated with forest cover (r = +0.67), LS factor (r = +0.62), and miscellaneous land (r = +0.49); strongly negatively correlated with total agriculture (r = −0.73) and rice fields (r = −0.71). The counterintuitive positive association with forest reflects terrain covariance: forests are preferentially located on steep slopes (forest vs. LS: r = +0.76), where modeled erosion is high despite protective vegetation.

Sediment Yield per Unit Area (SY/Area): Moderate positive correlations with precipitation (r = +0.59) and tree plantations (r = +0.55); moderate negative correlation with urban areas (r = −0.45). Precipitation drives runoff generation and sediment transport capacity. Tree plantations (often managed with sparse understory) are susceptible to leaching. Urban surfaces reduce sediment generation.

Sediment Delivery Ratio (SDR): Strong positive correlations with rice fields (r = +0.66) and total agriculture (r = +0.54); moderate negative correlations with other crops (r = −0.50), forest (r = −0.49), and LS (r = −0.47). The positive association with rice reflects seasonal bare-soil exposure and efficient runoff connectivity during monsoon season, despite bunded terraces. The negative association with LS likely reflects forest buffering on steep slopes.

Multicollinearity Among Predictors: Strong inter-correlations were observed among independent variables:

• Forest vs. LS: r = +0.76.
• Agriculture vs. rice: r = +0.69.
• Agriculture vs. forest: r = −0.93.
• Agriculture vs. LS: r = −0.77.

These interdependencies complicate direct interpretation and necessitate PCA to resolve multicollinearity for predictive modeling.

3.4. Multivariate Analysis and Identification of Controlling Factors

Principal Component Analysis of the 14 standardized predictor variables yielded 14 components (

Table 7. Principal Component Analysis summary: eigenvalues, proportion of variance explained, and cumulative variance for 14 components.

PC No.	1	2	3	4	5	6	7	8	9	10	11	12	13	14
Eigen value	4.41	3.08	2.78	1.13	0.91	0.56	0.50	0.39	0.11	0.08	0.03	0.01	0.00	0.00
Proportion	0.32	0.22	0.20	0.08	0.07	0.04	0.04	0.03	0.01	0.01	0.00	0.00	0.00	0.00
Cumulative	0.32	0.54	0.73	0.82	0.88	0.92	0.96	0.98	0.99	1.00	1.00	1.00	1.00	1.00

; Figure 8 scree plot). Following the Kaiser criterion (eigenvalue > 1) and the cumulative variance threshold (>90%), six components (PC1–PC6) were retained, collectively explaining 92% of the original variance. Eigenvectors (component loadings) are presented in

Table 8. Eigenvector loadings for six retained principal components showing correlations between original variables and components.

Variable		PC1	PC2	PC3	PC4	PC5	PC6
Symbol	Meaning
v01	Watershed area	−0.16	−0.11	−0.14	−0.63	−0.47	−0.39
v02	Precipitation	−0.26	−0.41	−0.07	0.10	0.08	0.17
v03	Runoff	−0.07	−0.35	0.35	0.35	−0.11	−0.31
v04	Rice fields	0.30	−0.02	0.37	−0.14	−0.35	0.05
v05	Other crop fields	−0.43	0.00	0.17	−0.04	−0.16	0.39
v06	Tree plantations	0.32	0.31	0.02	0.06	0.27	−0.51
v07	Orchards	−0.09	−0.26	−0.37	0.19	−0.35	−0.22
v08	Other types of agriculture	0.18	−0.29	0.41	0.19	−0.20	−0.20
v09	Total agricultural area	0.32	0.11	0.26	−0.36	−0.14	0.30
v10	Urban area	−0.33	0.33	0.21	0.08	−0.06	−0.09
v11	Forest area	0.29	−0.24	−0.38	0.04	0.02	0.09
v12	Water body	−0.16	−0.30	0.04	−0.48	0.48	−0.22
v13	Miscellaneous areas	0.12	−0.40	0.27	−0.13	0.33	0.11
v14	LS factor	0.40	−0.16	−0.25	0.02	−0.12	0.22

. Each PC was interpreted based on variables with high absolute loadings (|loading| > 0.30).

Interpretation

The six retained principal components can be interpreted based on their dominant variable loadings. PC1 (32% variance), termed the “Topography–Agriculture Gradient,” exhibits high positive loadings on the LS factor (+0.40), tree plantations (+0.32), total agriculture (+0.32), rice (+0.30), and forest (+0.29), while showing negative loadings on other crops (−0.43) and urban areas (−0.33). This component represents the contrast between steep, forested or agricultural uplands and flat, intensive-crop lowlands.

PC2 (22% variance), the “Hydroclimatic Forcing” component, is characterized by high negative loadings on precipitation (−0.41), miscellaneous land (−0.40), runoff (−0.35), and water bodies (−0.30), with positive loadings on urban areas (+0.33) and tree plantations (+0.31). This component captures hydroclimatic variability and the spatial distribution of water-related land features across the basin.

PC3 (20% variance), labeled “Rice–Forest Dichotomy,” shows high positive loadings on other agriculture (+0.41), rice (+0.37), and runoff (+0.35), contrasted with negative loadings on forest (−0.38) and orchards (−0.37). This component distinguishes rice-dominated lowlands characterized by high runoff generation from forested upland areas.

PC4 (8% variance), the “Watershed Scale and Water Bodies” component, exhibits strong negative loadings on watershed area (−0.63), water bodies (−0.48), and total agriculture (−0.36), with a positive loading on runoff (+0.35). This component represents basin size effects and the presence of reservoir infrastructure, reflecting the inverse relationship between drainage area and specific sediment yield.

PC5 (7% variance), termed “Water–Rice Trade-off,” is characterized by positive loadings on water bodies (+0.48) and miscellaneous land (+0.33), opposed by negative loadings on watershed area (−0.47), rice (−0.35), and orchards (−0.35). This component reflects local-scale variability in water storage infrastructure and rice cultivation intensity.

Finally, PC6 (4% variance), the “Plantation–Crop Contrast” component, shows a strong negative loading on tree plantations (−0.51) contrasted with positive loadings on other crops (+0.39), watershed area (+0.39), and total agriculture (+0.30). This component differentiates intensive upland cropping systems from managed tree plantation landscapes.

3.5. Predictive Framework for Sediment Export Across Sub-Watersheds

Using the six retained principal components as predictors, a linear regression model was fitted to predict SY/area (t ha⁻¹ y⁻¹) for each sub-watershed:

$$\begin{array}{c}SY/\mathrm{Area}=-0.1328+\left(3.43\times {10}^{-3}\right)PC1+\left(2.46\times {10}^{-2}\right)PC2+\left(5.58\times {10}^{-3}\right)PC3\\ +\left(-4.41\times {10}^{-3}\right)PC4+\left(-1.29\times {10}^{-3}\right)PC5+\left(-2.83\times {10}^{-3}\right)PC6\end{array}$$

(7)

The PCA-based regression model demonstrated a strong predictive capability for sediment yield per unit area (SY/Area) across the 19 sub-watersheds. The Pearson correlation coefficient between observed and predicted values was r = 0.81 (p < 0.001), with a coefficient of determination (R²) of 0.66, indicating that the model explains approximately two-thirds of the spatial variability in sediment yield. The root mean square error (RMSE) of 0.0042 t ha⁻¹ y⁻¹ and sum of squared errors (SSE) of 1.744 × 10⁻² demonstrate low residual error. Residual analysis revealed no systematic spatial bias, confirming model robustness and suggesting that the model captures the majority of the spatial variability in sediment yield while avoiding overfitting (Figure 9).

Examination of the regression coefficients provides insights into the dominant controls on sediment yield. PC2 (Hydroclimatic Forcing) exhibits the largest positive coefficient (+0.0087), confirming that precipitation and runoff are the primary drivers of sediment yield, consistent with Hypothesis H2. PC1 (Topography–Agriculture Gradient) and PC3 (Rice–Forest Dichotomy) also contribute positively to the model, reflecting the influence of topographic complexity and agricultural land-use patterns on sediment generation. Notably, PC4 (Watershed Scale) has a negative coefficient, which is consistent with the well-established inverse relationship between basin size and specific sediment yield reported by [13], wherein larger watersheds typically exhibit lower sediment yield per unit area due to increased opportunities for deposition and channel storage.

4. Discussion

4.1. Soil Loss Controls and Apparent Paradoxes

The strong positive correlation between soil loss and forest cover (r = +0.67) initially appears counterintuitive, as forests typically reduce erosion through canopy interception, root binding, and surface roughness [22]. However, this association is explained by terrain covariance: in the Mun Basin, forests are disproportionately located on steep slopes (r [Forest vs. LS] = +0.76), where the LS factor dominates erosion potential in the USLE framework. Because USLE does not explicitly model deposition or routing, modeled soil loss on forested steep slopes appears high despite the protective effect of vegetation. This highlights a fundamental limitation of USLE: it predicts detachment capacity (what could be mobilized) rather than net sediment export (what actually leaves the hillslope).

Conversely, the strong negative correlation between soil loss and agricultural land (r = −0.73) reflects the dominance of rice cultivation (67.96% of agricultural area), which is concentrated in low-slope plains (LS < 0.5) with effective conservation practices (bunded terraces, P = 0.10). Rice terraces function as distributed sediment traps, capturing hillslope runoff and promoting deposition [71]. This underscores the importance of integrating land-use management (C and P factors) with topographic context when interpreting USLE results.

Tree plantations (rubber, eucalyptus, oil palm) exhibit a moderate positive correlation with soil loss (r = +0.35) and a stronger positive correlation with SY/area (r = +0.55). Plantations are often established on marginal lands with moderate slopes and managed intensively, including understory clearing to facilitate harvest and the application of agrochemicals. This reduces protective ground cover and increases susceptibility to raindrop impact and surface runoff [6]. The stronger association with SY/area compared to soil loss suggests that plantations not only generate sediment but are also located in settings with efficient hillslope–channel connectivity.

4.2. Sediment Yield and Delivery: The Role of Reservoirs

The basin-wide sediment delivery ratio of 1.15% is exceptionally low compared to natural, unimpounded watersheds (typically 5–30% for basins > 10,000 km²; [13,65]). This result strongly supports Hypothesis H1: large-scale reservoir infrastructure is the dominant control on sediment connectivity in the Mun Basin. Sub-watershed 1, immediately downstream of Lam Takhong Dam (capacity 314 Mm³), exhibits the lowest SDR of 0.03%, with SY of only 1020 t y⁻¹ despite moderate soil loss (11.28 t ha⁻¹ y⁻¹). This corresponds to a trapping efficiency of >99%, consistent with published values for large reservoirs with high residence times ([72]).

Across the basin, four major reservoirs (Lam Takhong, Lam Phraploeng, Lam Mun Bon, Sirindhorn; combined capacity ~2675 Mm³) collectively store and trap an estimated 33–34 million t y⁻¹ of sediment (calculated as Total Soil Loss, SY = 35.05 − 0.40 ≈ 34.65 million t y⁻¹). This sediment is permanently or semi-permanently stored in reservoir beds, reducing downstream delivery and contributing to long-term reservoir siltation. While siltation reduces reservoir capacity for water storage and hydropower generation (estimated loss of ~0.5–1.0% capacity annually [66]), it also prevents downstream channel aggradation and maintains base flows during the dry season. In addition to engineered traps, natural sediment sinks include the following:

• Rice paddies with bunded terraces: Estimated to trap 1–2 million t y⁻¹ during flooding events.
• Floodplain deposition: Particularly in the lower basin during overbank flows.
• Channel storage: Sediment temporarily stored in channel bars and pools, remobilized during high-flow events.

The combination of these factors explains the basin’s low SDR and confirms the limited utility of USLE-based soil loss alone for predicting sediment export in highly managed watersheds.

4.3. Comparison with Alternative Sediment Modeling Approaches: MUSLE and USPED

While alternative modeling frameworks such as MUSLE (Modified USLE; [64]) and USPED (Unit Stream Power Erosion Deposition, [63]) offer theoretical advantages for certain applications, the USLE-SY-SDR approach employed in this study represents the most appropriate methodology for basin-scale assessment in the Mun River watershed given the current data availability and the dominant role of reservoir infrastructure.

MUSLE replaces the USLE rainfall erosivity factor with a runoff-energy term, enabling event-based sediment prediction and implicitly incorporating delivery processes through the runoff component. However, this theoretical elegance comes with stringent data requirements: continuous (sub-daily) discharge monitoring and empirically calibrated sediment rating curves for each sub-watershed outlet. Such high-temporal-resolution hydrological networks are rare in large tropical basins, and the Mun River system, like most developing-country watersheds, relies on daily discharge measurements at a limited number of stations. Retrofitting MUSLE to all 19 sub-watersheds would require infrastructure investment and multi-year monitoring campaigns that are beyond current operational capacity. More fundamentally, MUSLE was developed and validated primarily for single storm events in small agricultural catchments (<10 km²) rather than large, multi-season basins where sediment delivery is controlled by reservoir operations.

Similarly, USPED extends the USLE framework by incorporating topographic indices of unit stream power to predict spatially explicit erosion–deposition patterns. While this approach has proven successful for hillslope and small-watershed studies (<100 km²), scaling USPED to the Mun Basin’s 70,000 km² extent presents formidable computational and parameterization challenges. High-resolution flow routing across millions of grid cells demands substantial processing resources, and more critically, USPED does not explicitly represent reservoir trapping, the single most important control on sediment connectivity in this system. As demonstrated in Section 4.2, the four major reservoirs trap an estimated 34.65 million t y⁻¹ of sediment (>99% of mobilized material in some sub-watersheds), a process that USPED’s topographic algorithms cannot capture without extensive modification.

In contrast, our integrated USLE-SY-SDR framework offers distinct advantages for large-basin, reservoir-dominated systems. By coupling USLE-derived spatial erosion patterns with observed sediment yield at strategically located monitoring stations, this approach directly integrates all routing, deposition, and anthropogenic trapping processes without requiring their explicit parameterization. The derived SDR inherently quantifies the net effect of hillslope–channel connectivity, floodplain storage, and reservoir interception, processes that would require complex, data-intensive sub-models in alternative frameworks. This methodology is not a compromise but rather a pragmatic solution that maximizes information extraction from available data while maintaining scientific rigor. It has been widely validated in large-basin assessments globally [15,65] and provides transparent, reproducible results suitable for management decision-making.

Importantly, our approach does not preclude future refinement. As monitoring infrastructure expands, hybrid methodologies become feasible: USLE for broad-scale erosion risk screening coupled with process-based sediment routing models (e.g., SWAT, WEPP) for targeted sub-basins, calibrated against observed SY and validated through reservoir bathymetric surveys. Such integrated frameworks would combine the spatial comprehensiveness of empirical models with the mechanistic insights of process-based approaches, advancing both scientific understanding and operational predictive capacity.

4.4. Principal Component Analysis and Multivariate Controls

The PCA-based regression model successfully predicted SY/area with high accuracy (r = 0.81), supporting Hypothesis H3. By resolving multicollinearity among predictors, PCA produced a parsimonious model that avoids overfitting and provides interpretable components linked to underlying environmental gradients.

The dominance of PC2 (Hydroclimatic Forcing) in the regression model confirms Hypothesis H2: sediment yield is more strongly controlled by hydroclimatic factors (precipitation, runoff) than by hillslope characteristics (LS, land cover). This aligns with the transport-limited paradigm for large tropical watersheds, where sediment availability (gross erosion) exceeds transport capacity most of the time, but episodic high-discharge events during the monsoon season drive the majority of annual sediment export [13].

The negative coefficient for PC4 (Watershed Scale) reflects the well-established inverse relationship between drainage area and specific sediment yield: larger basins have more opportunities for deposition and channel storage, reducing SY/area [65]. This effect is compounded in the Mun Basin by the presence of reservoirs, which are more numerous in larger sub-watersheds.

4.5. Limitations and Future Directions

This study has several important limitations that warrant consideration when interpreting the results and that point toward valuable directions for future research. First, the USLE framework, while robust for sheet and rill erosion estimation, does not simulate other erosion processes that may contribute significantly to total sediment mobilization in certain sub-watersheds. Gully erosion, wind erosion, and tillage-induced soil displacement are not captured by the model, yet field observations reveal visible gully networks in sub-watersheds 2 and 17, suggesting that soil loss may be underestimated in these areas. Similarly, the model does not account for bank erosion along stream channels, which can be a substantial sediment source in large river systems.

Second, temporal mismatches in input data introduce potential uncertainty. Land-use data from 2019 were applied to 2024 rainfall and discharge records under the assumption that land-cover change over this five-year period was minimal. While comparison with available 2022 satellite imagery suggests changes of less than 2% in major land-use classes, dynamic land-use modeling incorporating annual or seasonal updates would improve temporal resolution and better capture the effects of ongoing agricultural expansion or forest recovery. Third, this study lacks direct field validation through erosion plot measurements or sediment trap installations. While we validated our estimates through comparison with published studies from similar watersheds and by coupling USLE predictions with observed sediment yield, these approaches integrate processes at large spatial scales and do not provide point-specific verification of modeled erosion rates. Establishing a network of instrumented monitoring sites would enable more rigorous assessment of model accuracy across different land-use and topographic settings.

Fourth, the focus on annual average sediment budgets obscures important event-scale dynamics. Sediment transport in tropical monsoonal systems is highly episodic, with more than 80% of annual sediment yield typically occurring during less than 10% of the year—specifically during high-intensity monsoon storms and peak discharge events. Event-based modeling would provide critical insights into sediment pulse dynamics, channel response to extreme flows, and optimal reservoir flushing strategies that balance sediment management with water supply objectives. Finally, this study does not explicitly address climate change’s impact on future erosion and sediment yield. Given the projected intensification of monsoon rainfall under climate change scenarios [73], both erosion rates and sediment delivery are likely to increase. A scenario analysis incorporating downscaled climate projections would inform long-term adaptation planning and help prioritize watershed management investments.

Future research should therefore pursue several complementary directions. Reservoir bathymetric surveys provide direct measurements of sediment accumulation rates, enabling the validation of SDR estimates and quantification of long-term storage trends. The implementation of event-based sediment monitoring at key sub-watershed outlets would capture the episodic nature of sediment transport and support the calibration of dynamic models. An exploration of hybrid modeling approaches, coupling USLE for broad-scale erosion mapping with process-based routing models such as SWAT, would better represent sediment connectivity while maintaining computational feasibility. Finally, a spatially explicit cost-effectiveness analysis of conservation interventions (e.g., riparian buffer strips, cover crops, terracing) based on the erosion and delivery hotspots identified in this study would support the evidence-based targeting of limited management resources to achieve maximum sediment reduction benefits.

5. Conclusions

This study provides a comprehensive assessment of soil erosion and sediment delivery in the Mun River watershed, integrating USLE-modeled hillslope erosion with observed sediment yield and multivariate analysis to address fundamental questions about erosion patterns, sediment connectivity, and environmental controls in this large tropical basin. Basin-wide soil loss is low to moderate but spatially heterogeneous. Total annual erosion is approximately 35 35 million t y⁻¹ (mean 4.96 t ha⁻¹ y⁻¹), with >80% of the basin in the “no erosion” to “very low” severity classes (0–6.7 t ha⁻¹ y⁻¹). Sub-watershed rates range from 1.20 t ha⁻¹ y⁻¹ in flat, rice-dominated lowlands to 14.40 t ha⁻¹ y⁻¹ in steep, mixed-use uplands, reflecting strong topographic and land-use influences on erosion potential.

Sediment delivery is highly constrained by reservoir infrastructure and agricultural practices. The basin-wide sediment delivery ratio of 1.15% is among the lowest reported for large tropical watersheds, indicating that only ~1% of mobilized sediment reaches the channel system. This exceptionally low SDR results primarily from sediment trapping by four major reservoirs, which collectively intercept an estimated 34.65 million t y⁻¹ annually, with secondary contributions from rice paddy deposition and floodplain storage. These findings confirm Hypothesis H1 and demonstrate that anthropogenic modifications, particularly large-scale water infrastructure, have fundamentally transformed the basin’s sediment budget from a natural to a highly managed system.

Erosion and sediment delivery are jointly controlled by hydroclimatic forcing, topography, and land use, though their relative importance differs between processes. Soil loss is most strongly influenced by topographic steepness (LS factor) and land cover, with the positive correlation between forest cover and erosion explained by terrain covariance; forests preferentially occupy steep slopes where detachment capacity is high, despite protective vegetation. Conversely, sediment yield per unit area is primarily controlled by precipitation and runoff (supporting Hypothesis H2), consistent with the transport-limited paradigm for large tropical basins. Tree plantations contribute disproportionately to sediment yield despite occupying only 7% of the basin area, reflecting sparse understory cover and intensive management that increases soil vulnerability.

Principal Component Analysis successfully resolved multicollinearity and enabled robust sediment yield prediction. Six orthogonal components explained >90% of the variance in 14 predictor variables and yielded a regression model with high accuracy (r = 0.81, R² = 0.66), confirming Hypothesis H3. Model coefficients reveal that hydroclimatic forcing (PC2) dominates spatial variability in sediment yield, followed by topographic–agricultural gradients (PC1) and rice–forest contrasts (PC3).

Priority management actions include targeting steep slopes in tree plantation areas for conservation interventions (cover cropping, mulching, agroforestry conversion); maintaining and enhancing bunded-terrace systems in rice lowlands with crop residue retention during vulnerable dry-to-wet transitions; monitoring reservoir siltation through bathymetric surveys and exploring coordinated sediment flushing or selective dredging; and establishing vegetated riparian buffers in high-delivery sub-watersheds (16, 19) to intercept hillslope sediment.

Important limitations include USLE’s exclusion of depositional processes, gully/wind/tillage erosion, and event-scale dynamics; validation based on regional comparisons rather than direct field measurements; the application of 2019 land-use data to 2024 conditions; and a focus on annual averages rather than monsoonal pulse dynamics. Future research should incorporate reservoir bathymetric surveys, event-based sediment monitoring, hybrid USLE-SWAT modeling approaches, and a cost-effectiveness analysis of spatially targeted conservation interventions.

The integrated USLE-SY-SDR-PCA framework is broadly transferable to large, data-limited basins in tropical monsoonal climates experiencing land-use change and reservoir development. By quantifying natural and anthropogenic sediment sinks, resolving predictor multicollinearity, and coupling empirical erosion modeling with sediment-yield observations, this approach provides a scientifically rigorous yet operationally practical foundation for watershed management and evidence-based policy development where process-based models remain data-constrained.