*2.2. Descriptions of SWAT and SWAT-MAD*

The SWAT model is a continuous-time, semi-distributed, process-based, and basinscale agro-hydrologic model [17]. The primary model components consist of hydrology, crop growth, and water quality and the major data needed for setting up the model for a basin are elevation, land use, soil, climate, and management practices [23]. The SWAT model has been commonly used to simulate basin-scale best management practices on hydrologic cycles and crop production worldwide [24–26]. Recently, a more representative MAD auto-irrigation method was developed by Chen et al. [21] and integrated into the SWAT model (hereafter referred to as SWAT-MAD). The MAD auto-irrigation method triggers irrigation according to a pre-defined allowable depletion percentage of plant available

water, determined by the crop-specific maximum rooting depth and soil-specific characteristics [21]. The ArcSWAT (version 2012.10\_2.19; revision 664; Stone Environmental, Inc., Montpelier, VT) for the ArcGIS 10.2.2 platform was used in this study. The SWAT Calibration and Uncertainty Procedures (SWAT-CUP 2012) with the Sequential Uncertainty Fitting version-2 (SUFI-2) [27] was used for the model calibration and validation for streamflow with the goal of maximizing Nash–Sutcliffe efficiency (*NSE*). The *NSE* [28], coefficient of determination (*R*2) [29], and percent bias (*PBIAS*) [30] were used to evaluate the performance of the SWAT-MAD model in the DMFB basin.

The SWAT-MAD model was calibrated and validated for streamflow data at two USGS gages and county-level crop yields of both irrigated and dryland cotton. The SWAT-MAD model was also evaluated by county-level seasonal irrigation requirements of cotton and percolation amount. The calibrated parameter values for the SWAT-MAD model are listed in Table S1. A detailed description of the SWAT-MAD model setup, calibration, and validation for the DMFB basin is provided in the Supplementary Materials. The SWAT-MAD model calibration and validation performance statistics for monthly streamflow at the stream gages (Table 1) were well above the "satisfactory" range suggested by Moriasi et al. [31]. The *R*<sup>2</sup> and overall *PBIAS* were 0.21 and 2.3% when comparing SWAT-MAD simulated and observed irrigated cotton lint yield in Lynn County [1] in the DMFB basin. The simulated irrigation for cotton by the MAD auto-irrigation method (346.9 mm) was very close to the local survey data [32]. The SWAT-MAD model simulated percolation amount was also comparable with the values from local reports and literature [33].

**Table 1.** Performance statistics for monthly streamflow prediction on two USGS gages in the Double Mountain Fork Brazos basin using the SWAT-MAD model.

