The Tseng-Wen Reservoir is a multipurpose reservoir designed for flood control, hydroelectric power generation, irrigation, water supply, recreation, and flow augmentation. The storage capacity of the Tseng-Wen reservoir is 608 × 10⁶ m³, but its effective water storage is 490 × 10⁶ m³, and the hydroelectric plant capacity is 50 MW. The Tseng-Wen Reservoir Basin is located in the upstream area of the Tseng-Wen River system in Chiayi County (Figure 1). The entire watershed area of this river basin is 1,176 km², in which the Tseng-Wen Reservoir watershed covers 481 km². The average slope of this river basin is approximately 1/57. Average rainfall in this watershed area is approximately 3,000 mm per year and annual average temperature is about 23.4° Celsius. Rich soil in the watershed is suitable for fruit and tea farms. Agriculture has restricted near the reservoir but tourism has developed in recent years.

Numerous major typhoons have struck Taiwan, such as Typhoon Mindulle in 2004, Haitang in 2005, Sepat in 2007, Kalmaegi in 2008, Morakot in 2009, and Fanapi in 2010 [17]. Especially, Typhoon Morakot struck Taiwan from 7 to 9 August, 2009, and produced record-breaking rainfall and catastrophic damage in Southern Taiwan. The typhoon produced copious amounts of rainfall peaking at 2,777 mm.

Multi-temporal Système Pour l'Observation de la Terre (SPOT) satellite images obtained after typhoons in 2001, 2003–2005, and 2007–2010 were used to quantify land cover changes in our study. Details for the dates are listed in Table 1. For atmospheric correction, Fast Line of Sight Atmospheric Analysis of Spectral Hypercube (FLAASH) is applied to correct the visible and near-infrared wavelengths in the satellite images [18]. Then, the NDVI maps were derived from the SPOT images taken in 2001 with a 20 m resolution, and with a 10 m resolution in 2003–2010. Moreover, the SPOT images were classified using supervised classification by the software package ERDAS IMAGINE. Land-use types were classified into the following six categories: forested land, built-up land, landslide, grassland, water, and bare land [19]. The reference maps were the aerial photographs by the Aerial Survey Office, Forestry Bureau, Taiwan.

Seasonal time series of water-quality data monitored in the reservoir were obtained from Taiwanese EPA Web sites [20]. The water quality data observed at three stations were obtained from 2001 to 2010 and data sampling frequency was three months. The sampling sites are shown in Figure 1. The water-quality variables, such as nitrate-nitrogen (NO₃-N), suspended sediments (SS), chemical oxygen demand (COD), dissolved oxygen (DO), total phosphorus (TP), and turbidity, were derived for analysis. The variables are used as general indicators of water quality. For example, the COD is commonly used to measure the amount of organic compounds in water. As the DO in water drops below a threshold, aquatic life is under stress. The presence of high nitrates and TP concentrations in water indicates possible pollution of the water. Turbidity is the haziness of a fluid caused by the SS, which are solid particles usually transported by flowing water.

Reservoir water chemistry was sampled at the outlets and downstream of the reservoir every three months. We refer to the water samples after typhoons in the fourth season and acquire annual values each year. We ran a series of models to examine the correlations between land-cover and the water quality variables.

Our basic statistical tool was stepwise multiple linear regression, with backward selection of variables and p = 0.1 to enter or remove variables. Cases with missing data were excluded. Statistical analyses were done using SPSS 10.0. The dependent (response) variables are NO₃-N and SS concentrations that are selected from the high-NDVI correlated water quality factors. The variable details are shown in Section 3.2. In addition, the independent variables are average NDVI at various spatial scales such as average NDVI in whole watershed, average NDVI in reservoir 1 km, 2 km and 3 km buffer zones, can be represented as: (1) Where NDVI(s,t) represent NDVI value varies with space s and time t; is average NDVI in whole watershed or reservoir buffer zone during time t; S is the domain set that defined as whole watershed or reservoir buffer zone; num is the number of the pixels in the set S.

Figure 2 shows that land use classification in 2001, 2004, 2007 and 2010. The forested land, grassland, bare land, build-up, and landslide accounted for 77.61%, 11.72%, 7.72%, 2.55% and 0.42% (excluding water) of the total watershed area in 2001, respectively. During 2001–2010, forest has decreased 4.81%, grassland has increased 2.35%, landslide has increased 2.12%, bare land has increased 0.55%, and built-up land has decreased 0.21% (Figure 3). The results matched previous studies [21,22] that many landslides in the Tseng-Wen reservoir watershed were caused by typhoons. Table 1 shows the statistics of NDVI images after typhoon events from 2001 to 2010. Results show that the lowest mean NDVI values (0.437) occurred on November 1, 2009, after Typhoon Morakot, and the second lowest NDVI values occurred on December 27, 2010 (0.494). The greatest impact on the landscape is from Typhoons Morakot and Fanapi. During the event (i.e., Typhoon Fanapi), the standard deviation of NDVI values was the largest. The analysis results of NDVI images (Figure 4) are sufficient to present land cover changes induced by disturbances, particularly by spatial structure, variability, and spatial correlation. The disturbances impacted the fragmentation and interspersion of the low NDVI patches and created heterogeneous patterns across the landscape within the affected area [16]. However, land cover change may be different in the spatial scales. The box plot shows that the range between the lower and upper quartiles in the NDVI decreases when the buffer zone increases (Figure 5).

Table 2 shows the correlation coefficients between the average NDVI in the watershed and average water quality factors during the whole year and after typhoon. The average value of the NDVI in the watershed is strongly correlated with NO₃-N during the entire year and is strongly correlated with SS after typhoon from 2001 to 2010. These water quality factors such as SS and NO₃-N concentrations are the indices for water quality assessment when considering land cover changes. The average SS concentrations after typhoons and the average annual NO₃-N in the sites are used by later descriptive statistics and regression analysis.

Table 3 lists quarterly measurements of SS, and NO₃-N data at three water quality-monitoring stations from 2001 to 2010 (Figure 6). The average concentrations for SS for observation Sites 1, 2, and 3 are in the ranges of 0.8–13.5 (ppm), 1.1–12.6 (ppm), and 1.0–9.8 (ppm), respectively. Moreover, the average concentration of NO₃-N for Sites 1, 2, and 3 are in the ranges of 0.01–1.62 (ppm), 0.01–1.45 (ppm), and 0.01–1.20 (ppm), respectively. The average SS values are 4.51, 4.85, and 4.52 (ppm), and the average NO₃-N values are 0.44, 0.47, and 0.45 (ppm) in these 3 sites. Figure 6(a) shows that SS concentrations are cyclical. Typhoons and heavy rainfalls trigger large sediment discharge into the rivers of Taiwan and cause high-suspended sediment concentrations during these events. Most nitrate concentrations in the water drains from agricultural land. However, Figure 6(b) shows NO₃-N concentrations vary with a decreasing trend.

Table 4 shows the regression models for water quality and land cover changes at various spatial scales. The factors comprise the average NDVI in the watershed, and the average NDVI in 1 km, 2 km, and 3 km buffer zones. Results show that annual nitrate-nitrogen concentrations are positively correlated with the NDVI with a 1 km buffer area. However, SS is negatively impacted by the average NDVI in the watershed, suggesting that typhoons impact land cover change in the watershed. For example, typhoons cause landslides, and these are a major source of soil erosion and sediment yield in the watershed. The average NDVI in the watershed adversely impacts the water quality, and therefore, increases sediments associated with water quality. The average NDVI in the watershed becomes a key factor influencing the SS concentration. Typhoon events are major natural disturbances causing NDVI changes and also cause serious landslides [16,23,24]. Both important factors affecting soil erosion and sediment delivery to river channels are changes in land use and climate. Due to destruction of vegetation and increased soil exposure in the watershed after rainstorms and typhoons, the NDVI values decreased. The rainstorms and typhoons cause divergent destruction of vegetation, and led to an increase in the potential for soil erosion.

Results show that the dominant explanatory variables in NO₃-N cases have an average NDVI with a 1 km buffer zone. In some reservoirs and lakes, the primary indicator of agriculture is dependent on NO₃-N concentration. The nitrate concentration is correlated with agricultural practices during the high-flow period. Human activity alters the patterns of nitrate concentrations during storm events in the agricultural catchment [25]. Since Taiwan joined the World Trade Organization (WTO) in 2002, imported agricultural products are cheaper than domestic ones, thus negating the need for extensive agriculture areas. This corresponds to the data that tea farms in Chiayi County decreased from 2,292 to 2,189 ha from 2005 to 2011 [26]. Forestry, agriculture, and anthropogenic activities impact the quality of water over short and long periods [27]. The SS and NO₃-N are typically sensitive in landslide and agriculture land areas. The previous results match that percent agricultural coverage had a significant influence on both SS and nitrate-N loading [7]. NDVI variation results imply that as land cover changes; hence, the multi-scale NDVI, which is one of the indices in the watershed of land cover changes, is associated with water quality and is hard to directly link with agriculture. However, the land use classification of SPOT images is also hard to identify the agriculture land. Further study could consider the high-resolution satellite images in the land use classification.