Rainfall and runoff, the main erosion agents, detach soil particles from its matrix and transport gravitationally the detached materials or sediments to surrounding rivers. In the rivers, sediments flow further downstream with streamflow. Suspended sediment load (SSL) is a major portion of the total load transported by streams [1] and commonly accounts for 85% to 95% [2]. Sediment has been becoming an important issue involving in sustainable development of water resources system. The Mekong2Rio conference also addressed the concern on food security which could be adversely affected by alteration of the sediment natural flow [3]. Construction of water storages (e.g. dam-reservoirs) can provide solutions to food security issues through increased irrigation and at the same time improve access to energy through hydropower generation. However, such developments could affect on fisheries through the loss of sediment trapped behind dam walls, for example. The Lower Mekong Basin (LMB) contains over 100 hydropower projects (HPP) and if there are no any effective countermeasures taken into account, their development could trap sediment around 26 Mt/year, 60% of the total basin production [4].

Quantification of sediment load is necessary not only during the project development stage but also along the course of operation until decommissioning [5-12]. It is interesting with regard to reservoir sedimentation, fish habitat, river utilization as well as biological sustainability in the whole river basin [13]. The most reliable way in estimating sediment load is the use of its observed records, but sediment sampling is very difficult and requires high experienced professionals because of its significant fluctuation within the river section [9] and userunfriendly measurement tools. Moreover, it is time consuming and costly [11,14]. These constraints have led to low frequency of sediment observation around the world and especially in developing and remote regions [1] such as the LMB. In response to this problem, modeling approach, based on different hydrological variables and terrain attributes, has been taken into consideration.

The unit stream power (USP) theory of Yang [15], the SHESED model of Wicks and Bathurst [16] and others known as the physically-based models, could be universally used to predict sediment yield of a watershed but it requires a lot of detailed information including hydrological, hydraulic and geological characteristics of the river basin, and as well as sediment characteristics itself. Preparation of such dataset will be difficult and costly. Furthermore, the stream power approach cannot predict well the SSL because, in rivers, the finest fraction of SSL is often a non-capacity load [17,18]. Similarly for process-based models such as the modified universal soil loss equation (MUSLE), introduced by Williams [19], and its family (USLE and RUSLE), they also require huge amount of input data. Data consumption and complex sediment transport mechanism have driven both physicallyand process-based models to be based on many simplifying assumptions and empirical relationships, particularly for rainfall and runoff erosive effects [7,20]. In consequence, their application in data scarce areas would yield high uncertain results or be completely infeasible.

Alternative approach to the processand physicallybased techniques is the utilization of data-driven models. Artificial neural network (ANN) is the most well-known and powerful data-driven method and it has been proved to be useful in modeling complex hydrologic processes or non-linear systems such as sediment transport [10,21, 22]. ANN forecasts outputs using experiences learned from historical data. This method is widely used because it does not require detailed information of the physical process controlling the system and generally applicable using available hydrological data. Tayfur [23] stated that ANN is a very practical and promising modeling tool for the study of sediment transport processes in data shortage regions. Kisi and Shiri [13] employed ANN to estimate daily suspended sediment concentration (SSC) in Eel River, California, and obtained a very satisfactory result with determination coefficient (R²) ranging between 0.82 and 0.95. In predicting daily SSL in Mississippi, Missouri and Rio Grande River in USA, Melesse et al. [11] found that ANN (0.65 ≤ R² ≤ 0.96) for most cases is superior to other data-driven models: multiple linear/nonlinear regressions and autoregressive integrated moving average. In Longchuanjiang River, the Upper Yangtze Catchment in China, monthly SSL was modeled well by ANN with R² varying from 0.66 to 0.89 in validation stage [24].

Singh et al. [12] compared two different models for predicting monthly SSL of Nagwa watershed in India and the results showed that ANN is better than MUSLE for larger R² 8% in calibration stage and 13% in validation stage. Similar study conducted by Talebizadeh et al. [25] demonstrated that ANN is superior to MUSLE in estimating low and medium values but inferior in case of high values. In comparing with various physically-based models including USP, the performance of ANN is comparable and in some cases better [23]. Additionally, ANN could provide detailed information for design purposes and management practices in civil and environmental engineering sector [10], and hysteretic analysis of sediment transport [26] which no other methods have been confirmed their applicability yet. However, one drawback of ANN is the need of long time series data for system training.

To our knowledge, there are many studies considering ANN for SSL simulation but very limited researches assessing its applicability in ungauged catchments (UC). The objectives of this research are to: 1) simulate monthly average SSL (SSL_m) of four catchments using ANN; 2) assess the application of the calibrated ANN models in three ungauged catchment representatives (UCR) before using them to predict SSL_m of three actual ungauged catchments (AUC) in the Tonle Sap River Basin; and 3) estimate annual SSL (SSL_A) of each AUC for the case of with and without hydropower dam-reservoirs. All ten catchments considered in this study are located totally in the LMB. UC in this context refers to catchment having no sediment observation.