shahrokh fatehi; jahangard mohammadi; Mohammad Hassan Salehi; aziz momeni; Norair Toomanian; Azam Jafari
Abstract
Introduction: Spatial scale is a major concept in many sciences concerned with human activities and physical, chemical and biological processes occurring at the earth’s surface. Many environmental problems such as the impact of climate change on ecosystems, food, water and soil security requires not ...
Read More
Introduction: Spatial scale is a major concept in many sciences concerned with human activities and physical, chemical and biological processes occurring at the earth’s surface. Many environmental problems such as the impact of climate change on ecosystems, food, water and soil security requires not only an understanding of how processes operates at different scales and how they can be linked across scales but also gathering more information at finer spatial resolution. This paper presents results of different downscaling techniques taking soil organic matter data as one of the main and basic environmental piece of information in Mereksubcatchment (covered about 24000 ha) located in Kermanshah province. Techniques include direct model and point sampling under generalized linear model, regression tree and artificial neural networks. Model performances with respect to different indices were compared.
Materials and Methods: legacy soil data is used in this research, 320 observation points were randomly selected. Soil samples were collected from 0-30 cm of the soil surface layer in 2008 year. After preliminary data processing and point pattern analysis, spatial structure information of organic carbon determined using variography. Then, the support point data were converted to block support of 50 m by using block ordinary kriging. Covariates obtained from three resources including digital elevation model, TM Landsat imagery and legacy polygon maps. 23 relief parameters were derived from digital elevation model with 10m × 10m grid-cell resolution. Environmental information obtained from Landsat imagery included, clay index, normalized difference vegetation index, grain size index. The image data were re-sampled from its original spatial resolution of 30*30m to resolution of 10m*10m. Geomorphology, lithology and land use maps were also included in modelling process as categorical auxiliary variables. All auxiliary variables aggregated to 50*50 grid resolutions using mean filtering. In this study Direct and point sampling downscaling techniques were used under different statistical and data mining algorithms, including generalized linear models, regression trees and artificial neural networks. The direct approach was implemented here using generalized linear models, regression trees and artificial neural networks in following three steps, (i) creating the spatial resolution of 50m*50m averaged over 10m*10m grid resolution environmental variables within each coarse grid resolution, (ii) establishing relationships between these coarse grid resolutions of 50m*50m environmental variables and soil organic carbon using GLMs, regression tree and neural networks and (iii) using parameter values gained in step 2 in combination with the original 10m*10mgrid resolution environmental variables to produce predictions of soil organic carbon with10m*10m grid resolution. In point sampling approach, within each coarse resolution (50m*50m), a fixed number of fine grid resolution (10m*10m) were randomly selected to calibrate models at high resolution. In this study, 5 fine grid resolutions (20% fine grid cell within each coarse grid cell) randomlywere sampled at. Then, each selected point overlied on an underlying fine-resolution grid and recorded its environmental variables and averaged fine grid resolution (10m*10m) within their corresponding coarse grid resolution (50m*50m). To calibrate model parameters, these averaged environmental variables were used. The calibrated parameters applied to fine-resolution environmental data in order to predict soil organic carbon at spatial resolution of 10m*10m. The prediction accuracy of the resulting soil organic carbon maps was evaluated using a K-fold validation approach. For this purpose, the entire dataset was divided into calibration (n = 240) and validation (n = 80) datasets four times at random. Prediction of soil organic carbon using calibration datasets and their validation was conducted for each split, and the average validation indices are reported here. The obtained values of the observed and predicted SOC were interpreted by calculating Adjusted R2 and the root mean square error (RMSE).
Results and Discussion: Point pattern analysis showed the sampling design is, generally, representative relative to geographical space .A semi-variogram was used to drive the spatial structure information of soil organic carbon. We used an exponential model to map soil organic carbon using block kriging. Grid resolution block kriging map was 50m*50m. Since the distribution of organic carbon variable and covariates were normal or close to normal for run generalized linear models selected Gaussian families and identity link function. The validation results of this model in point sampling was slightly (Adjusted R2=0.57 and RMSE=0.22) better than the direct method (Adjusted R2 =0.47 and RMSE=0.26).The results of modelling using regression tree in point sampling approach (Adjusted R2 =0.57and RMSE=0.22) is very close to the direct method (Adjusted R2 =0.57 and RMSE=0.23).In implementation of neural networks, the combination of the number of neurons and learning rate for direct downscaling method were obtained 10 and 0.10, respectively and for point sampling downscaling method were, 20 and 0.1 The results of validation obtained from the implementation of this model in point sampling approach (Adjusted R2 =0.45 and RMSE=0.27) is very close to the direct method (Adjusted R2 =0.47 and RMSE=0.28).Validation results indicated that in both downscaling approaches, regression tree (Adjusted R2=0.57, root mean square root (RMSE) =0.22-0.23) has higher accuracy and efficiency better than generalized linear models (Adjusted R2=0.49-0.57, RMSE=0.22-0.26) and neural network (Adjusted R2=0.45-0.47, RMSE=0.27-0.28).
Conclusion: In general, the results showed that the efficiency and accuracy of the sampling point approach is slightly better than the direct approach. Validation results indicated that in both downscaling approaches, regression tree has higher accuracy and performed better than neural network and generalized linear models. However, it is required to perform more research on the different ways of downscaling digital soil maps in the future.
