Determination Of Soil Sampling Density For Slope Positions At Agricultural Landscape In Purple Soil Hilly Region

Soil sampling is essential for acquiring soil information and spatial distribution; however, sampling across an area is usually time-and labor-consuming as well as costly. It is desirable for a survey to collect the minimum number of soil samples necessary to estimate the values of soil properties within a specified area. In purple soil hilly region, southwestern China, the dry sloping soils developed on the identical parent material are generally planted with the same crop and under the uniform management practice; even so soil still exhibits evident heterogeneity at spatial space. The fact indicates that topography imposes an influential effect on soil properties, and the effect must be taken into consideration when establishing a sampling scheme. In addition, complex terrains of this region make it more difficult to collect soil samples compared with an area with flat terrain. Therefore, it is necessary to determine an appropriate sampling density to guide soil sampling in this area. A method of determining soil sampling density was developed by combining digital soil mapping, combinational optimizing algorithm, and slope classification. The proposed method was applied to a farmland area (covering an area of approximately2km2) located in Yongxing, Jiangjin, Chongqing (calibration area), which is representative of a purple soil hilly region; and sampling density of soil properties (soil pH, organic matter, alkali-hydrolyzable N, available P, and available K) for slope positions (ridge, shoulder, backslope, footslope, and valley) was determined. Then the determined soil sampling density was used to select soil samples in another area (validation area) which is similar in environmental conditions with calibration area. The main results are as follows:1. Spatial prediction of soil properties within a farmland in purple soil hilly regionPearson correlation coefficients showed that soil pH, organic matter, and alkali-hydrolyzable N strongly correlated with terrain attributes, while the relations between soil available P, available K and topographical indicators were poor. This information indicated that topography markedly influenced variation of soil pH, organic matter, and alkali-hydrolyzable N; however, the effect of terrain attributes on soil available P, and available K were weak. One-way analysis of variance (ANOVA) found significant differences for soil pH, organic matter, alkali-hydrolyzable N, and available P with land use types (crop land and paddy field), while not for available K. The results revealed that land use types significantly affected variance of these soil properties (except available K). Prediction models (based on terrain attributes and based on combination of terrain attributes and land use types) could account for3.1～72.4%of the variability in soil properties (soil pH, organic matter, alkali-hydrolyzable N, available P, and available K). The models in predicting organic matter, alkali-hydrolyzable N, and pH had a good predictive ability, however, that of available P and available K prediction models was poor because of P and K were mainly controlled by parent materials while not by topography in this study area. The result was consistent with findings of previous researchers and could be accepted.Comparisons between prediction models based on terrain attributes and based on combination of terrain attributes and land use types indicated that inclusion of land use types in the prediction models was not always improving the prediction accuracy. For example, the precision of the model based on combination of terrain attributes and land use types in predicting pH was poorer than that of the model based on terrain attributes; and for alkali-hydrolyzable N, the predictive ability of the two types of prediction models were nearly the same; there were only a little improvements in predicting soil organic matter and available P of inclusion of land use types in prediction models. It could be explained that topography determined the spatial distribution of land use types to a great extent in the current study area, thus topography and land use jointly influenced soil properties and their effects were mostly interactered. If the interactive effects were filtered out, the contribution of the land use types in explaining the variability of soil properties would be very limited and could be ignored.2. Optimizing spatial distribution of soil sampling points and soil-landscape model using simulated annealing (SA) algorithmSimulated annealing combined with multiple linear regression was used to optimize the spatial distribution of original200soil sampling points in the calibration data set. The original200soil sampling points were optimized in spatial distribution from2to199for each soil property. An optimized soil-landscape model and corresponding prediction error (mean squared error) were given to each combination of optimized soil sampling points. From the prediction error, it could be seen that6,7,7,3, and3optimized soil sampling points could be used to replace the original200soil sampling points to predict the spatial variation of pH, organic matter, alkali-hydrolyzable N, available P, and available K, respectively.As also seen from the prediction errors, accuracies of the prediction models (for soil pH, organic matter, alkali-hydrolyzable N, available P, and available K, respectively) were the most high when the number of optimized soil sampling points was136,124,48, and95, respectively. Values of adjusted determination coefficient (R2adj) of optimized soil-landscape models were all larger than that of the original soil-landscape models. The values of R2adj of optimized soil-landscape models were improved by263.82,192.25,18.86,8.38, and4.56%for available P, available K, pH, alkali-hydrolyzable N, and organic matter, respectively, compared with those of the original soil-landscape models. On the contrary, prediction errors of the optimized soil-landscapes were all lower than those of the original soil-landscape models. Values of mean absolute error (MAE) of the optimized models were reduced by11.83,11.20,3.99,1.79, and1.63%for pH, available P, organic matter, available K, and alkali-hydrolyzable N, respectively, compared with those of the original soil-landscape models. Values of root mean squared error (RMSE) of the optimized models were reduced by12.14,11.06,9.29,6.88, and3.93%for available P, available K, pH, alkali-hydrolyzable N, and organic matter, respectively, compared with those of the original soil-landscape models. Values of akaike information criterion (AIC) of the optimized models were reduced by4.33,2.78,1.87,1.39, and1.27%for soil pH, available P, available K, organic matter, and alkali-hydrolyzable N, respectively, compared with those of the original soil-landscape models.3. Field scale slope position segmentation at agricultural landscape in purple soil hilly regionField scale slope positions segmentation was carried out by using similarity-based approach at agricultural landscape in purple soil hilly region. The classified slope positions (ridge, shoulder, backslope, footslope, and valley) generally followed the actual feature of the landform. To further validate the usefulness of the similarity-based model in complex terrain area, one-way ANOVA was applied to examine differences of soil properties (soil pH, organic matter, alkali-hydrolyzable N, available P, and available K) among slope positions. The results of ANOVA showed that significant differences for soil properties were found with slope positions. In addition, the relationships between soil thickness, soil types and quantified spatial gradient were also investigated and obvious trends were found. Finally, correspondence analysis (CA) was employed to examine the relations between slope positions and spatial distribution of land use types. The results showed that topography controlled the spatial distribution of land use types. All these information verified the validity of similarity-based approach used in the complex terrain area. Compared with the traditional slope position classification method, the similarity-based approach not only could produce the "harden" map of the slope positions, but also could quantitatively describe the spatial gradient of slope information. The latter could provide detailed information for simulation process, such as slope erosion or predictive soil mapping, at fine scale. Areas with little fuzziness correspond well to soil species, while areas with high ambiguity correspond to miscellaneous soil species at the transitional areas of slope positions. This indicated that uncertainties in slope position classification must be considered in soil-landscape modeling, otherwise it would lead to wrong inference.4. Determination of soil sampling density for slope positions at agricultural landscape in purple soil hilly regionGeostatistical method was employed to fit semi variogram of soil properties (soil pH, organic matter, alkali-hydrolyzable N, available P, and available K) for slope positions (ridge, shoulder, backslope, footslope, and valley) and the whole study area. The parameter range of the semi variogram could reflect the degree of soil variation. The larger a range the less variable a soil property would be, and vice versa. The ranges of soil properties for slope positions were different from each other, which indicated that variation of soil properties among slope positions were not the same. This verified the justifiability of the proposed method determining soil sampling density for slope positions. Further, ranges of soil properties for slope positions were all larger than that of the whole study area. This showed that soil properties variability was more homogeneous within slope positions than that of the whole study area, which implied that slope position segmentation could improve efficiency in soil sampling.The proposed method combining simulated annealing, multiple linear regression, and slope position classification was used to determine soil sampling density for slope positions (ridge, shoulder, backsloe, footslope, and valley) with15%proper relative error. The sampling densities of pH were8.6,9.4,10.4,10.4, and11.5ha for ridge, shoulder, backslope, footslope, and valley, respectively. The sampling densities of soil organic matter were8.6,7.5,5.8,6.2, and5.7ha for ridge, shoulder, backslope, footslope, and valley, respectively. The sampling densities of soil alkali-hydrolyzable N were8.6,6.2,5.8,5.2, and11.5ha for ridge, shoulder, backslope, footslope, and valley, respectively. The sampling densities of soil available P were2.9,2.5,2.3,1.8, and3.8ha for ridge, shoulder, backslope, footslope, and valley, respectively. The sampling densities of soil available K were4.3,3.7,3.3,2.8, and5.7ha for ridge, shoulder, backslope, footslope, and valley, respectively. Further, the determined soil sampling density was applied to another area which was similar in environmental conditions with the calibration area. The results showed that the determined soil sampling density could meet the given accuracy requirement.