Great attentions have been paid on landscape changes in recent years. As a basis of further researches on landscape functions and dynamics, contributes to analyzing spatial distribution characteristics of landscape components, landscape pattern analysis became an important topic of landscape ecology. In landscape pattern analysis, landscape metrics has been used as a common tool to exhibit the spatial distribution of landscape. During the past two decades, many processes have been used to analyze changes of landscape and the relationships with human influences and environmental factors. As a hotspot of quantitative methods of landscape analysis, gradient analysis, which is conducive to revealing the evolutionary laws of spatial landscape patterns has became a significant means in landscape ecology. However, the appropriate spatial scale is the key point to calculate the landscape metrics, and scale issues represent one of the foremost frontiers of landscape ecology. It is well known that the observed landscape pattern and its relationship with (landscape) process depend upon the scale. Great developments have been made on researches on pixel and spatial extent of landscape pattern evolution quantitatively and qualitatively. Little on the accuracy of the landscape analysis is related to both pixel and extent effect since there is the lackage of systematic study on the selection of optimization scale in landscape gradient pattern analysis. In this study, taking the dry valley of Minjiang River as study area and using ARCGIS9.3, ENVI4.8 and FRAGSTATS3.3, based on the image data of Land Resources Satellite ETM+ (2011), we aim to analyze the variation of landscape index with grain size. To obtain gradient patterns of the landscapes, 4 transects have been set up along both mainstream and tributaries of Minjiang River. A series of metrics at the landscape level (NP, PD, LPI, DIVISION, SHDI, and SHEI) were chosen and calculated using standard and moving window approach with different spatial scale, respectively. Then, an optimization-scale selection method, which integrates: the grain effect analysis of landscape index, the data loss assessment and the landscape index range effect curve analysis, was developed to obtain accuracy and efficient scale. By analyzing the grain inflexions of the landscape index comprehensively, we found that 30-90m and 110-160m were the appropriate grain ranges. Data loss assessment showed that 50m was the appropriate grain extent. In addition, based on moving window analysis, landscape index range effect curve analysis suggested that 250m was the most appropriate spatial extent for landscape pattern analysis. At the landscape level, 6 metrics were calculated by Moving window method: shrub land, accounting for 73.82% of the total landscape, was the matrix in dry valley of Minjiang River in the year 2011; the forestland and grassland's landscape heterogeneity decreased and relatively high the fragmentation degree of construct and farmland; water area has no obvious change; landscape metrics in the four transects present different amplitude and evident gradient diversity as landscape type change, the index change is larger in transition zone as compared to the single region; landscape pattern develops in the direction of heterogeneity, diversification and homogenization; and topography, precipitation, temperature and human activities were the factors for the gradient changes of landscape. As a kind of beneficial attempt, this study more finely analyzed the landscape pattern in the study area. The result of landscape pattern gradient analysis provided a novel way for discerning the landscape pattern change in the mountainous areas.