عنوان مقاله [English]
Background and objective: Digital soil maps with fine resolution are one of the basic needs of users and decision-makers in agriculture, natural resource and environment. However, in our country, there is scarcity of this kind of data and producing fine resolution soil data is very costly. Therefore, downscaling digital soil maps arises as a suitable option in order to produce fine resolution soil data. Objectives of this study were to examine and evaluate downscaling digital maps of some soil surface properties from block supports 50,100 and 250 m to block support 10 m using direct approach across Merek sub catchment in Kermanshah province with an area of 24000 ha.
Material and methods: The first spatial structure information of soil surface properties including %sand, % silt, %clay, %organic carbon, % equivalent calcium carbonate and %gravel determined using legacy data(320 randomized point samples) and variography. Then, block kriging maps were produced with block support 50,100 and 250m. Terrain attributes, Landsat images, geology map, geomorphology and land use maps were used in this study as auxiliary variable. Correlation coefficient between auxiliary variables and target variables is calculated and auxiliary variables were significant at the 0.01 level selected as model inputs. Afterward, downscaling direct approach is used. In this approach, relationship between the soil properties and auxiliary variables with coarse resolution identified using generalized linear models (GLMs) and regression tree. Next, calibrated parameters and fine resolution covariates are applied to prediction soil properties in fine resolution. Models are trained on 75% of the block support data accordance with original data and evaluated on the remaining 25%, using k-fold validation (k=4) procedure.
Results: The results showed that amount of sand and gravel had minimum and maximum correlations with covariates, respectively. Considering all the pixel sizes, the highest correlation obtained among soil properties and elevation, direct duration, convexity, slope, topographic wetness index and mrvbf. Downscaling gravel map from block support 50m to 10m by GLMs showed best performance (RMSE=5.57%). Downscaling sand, clay, equivalent calcium carbonate and organic carbon from 250m block support and silt from of 50 m to 10 m block support using regression tree lead to estimate the lowest root mean square error (3.9%, 3%, 4.39%, 0.21% and 2.31% respectively). Besides, regression trees showed the best performance in downscaling of soil properties with different pixel size.
Conclusion: it seems that direct approach would be able to downscale digital maps of soil variables (such as silt, calcium carbonate equivalent, sand and organic carbon content) with acceptable accuracy and efficiency. Obviously, GLMs and regression tree can lead to strong results if the correlation between soil properties and auxiliary variables is high. It can be concluded that various auxiliary variables at diverse pixel sizes have different relationships with the target variable which affect the performance of the downscaling.