| Accurate and detailed soil spatial information is vital for environment modeling,ecological assessment,and decision making.The rapid development of remote sensing satellites provides a huge opportunity to improve the performance of soil prediction models,but the potential of Gaofen-1(GF-1)remote sensing data in three-dimensional soil mapping has not been fully explored.In this study,the Luoding City was used as the model building area,and the GF-1 multi-spectral derived remote sensing data were combined with other environmental data(basic soil maps,digital elevation model(DEM)derived terrain-hydrology data,a basic artificial neural network(ANN)model based on basic soil data,a terrain-hydrology ANN_A model with basic soil data+terrain-hydrology variables,a full variable ANN_B model with basic soil data+terrain-hydrology variables+remote sensing variables and a remote sensing ANN_Cmodel with basic soil data+GF-1 remote sensing variables were constructed to explore the role of GF-1 derived remote sensing variables in predicting forest soil nutrients.In addition,Xinxing County is used as the model promotion area,by comparing the results of directly applying the optimal model in the model building area and constructing a simple and efficient linear promotion model,to explore the promotion performance of GF-1-derived remote sensing variables in predicting forest soil nutrient models.The prediction objects were organic matter(OM),alkali-hydro-nitrogen(AN),available phosphorus(AP),available potassium(AK),total nitrogen(TN),total phosphorus(TP),and total potassium(TK)at five soil depths(D1:0-20 cm,D2:20-40 cm,D3:40-60 cm,D4:60-80cm,and D5:80-100 cm).The results showed that:(1)The full-variable ANN_B model coupled with remote sensing data has the best prediction accuracy,followed by the ANN_A model of terrain-hydrology without remote sensing data,and the ANN_C model with only remote sensing data has the worst prediction accuracy.The prediction accuracy(R~2)of the ANN_Amodel for the seven soil properties was between 0.61 and 0.78,the prediction accuracy(R~2)of the ANN_B model was between 0.66 and 0.88,and the prediction accuracy(R~2)of the ANN_Cmodel was between 0.42 and 0.76.In short,ANN_B>ANN_A>ANN_C.The improved accuracy of the fully variable ANN_B model coupled with remote sensing data and the good prediction accuracy of ANN_C with only remote sensing data indicated that GF-1-derived vegetation remote sensing variables are effective for soil prediction.(2)The vegetation remote sensing variables have better prediction performance for the upper(0-40 cm)soil nutrients.Compared with the terrain-hydrology ANN_A model without remote sensing data,the prediction accuracy(R~2)of the full-variable ANN_B model with remote sensing data can be significantly improved by 0.10-0.14 at the 0-40 cm(except AK and TP),but the accuracy improvement are poor(0.02-0.08)at 40-100 cm.Vegetation remote sensing ANN_C models with only remote sensing variables have poorer accuracy with soil depth,which also proves that the vegetation remote sensing variables perform better in the upper soil.In addition,except for the two soil indicators AK and TP,the R_C~2 of other soil indicators D1 soil layer is close to or even better than R_A~2,OM is less than 0.01,AN is less than 0.02,AP is greater than 0.01,TN is the same,and TK is greater than 0.02.It shows that the prediction ability of vegetation remote sensing variables for topsoil(0-20 cm)is equivalent to terrain-hydrology variables.(3)The vegetation remote sensing variables have poor predictive power for AK and TP.In the upper soil depth,the prediction accuracy of the full-variables ANN_B model which combination of remote sensing data for AK and TP improved only 0.05-0.07(R~2).The vegetation remote sensing ANN_C model with only remote sensing variables predicted the mean R~2 of OM,AN,AP,AK,TN,TP,and TK to be 0.63,0.63,0.60,0.53,0.58,0.51,and 0.63,respectively,with AK and TP having the lowest accuracy.These indicate that the vegetation remote sensing variables selected in this study could not capture the spatial variability of AK and TP well.It may be that the spatial variability of AK and TP are mainly affected by other vegetation variables,or maybe that the range of TP content is too narrow,making the model difficult to fit.(4)The linear model well corrects the result of the direct promotion of the ANN model.This study found that the accuracy(ROA)of the full-variable ANN_Bmodel coupled with remote sensing data lost 23%～40%in the directly promoted area(Xinxing County),but there was a significant correlation between the predicted value of directly promoted and the measured value.In addition,zoning was performed based on soil type,and 50%of the samples in each type area were used to establish a linear model,and the other 50%of the samples were used for independence test,and it was found that the ROA increased by 13%-28%.(5)The three-dimensional spatial distribution of forest soil OM,AN,AP,AK,and TN in the model building area(Luoding City)gradually decreased with soil depth,while TP and TK increased slightly with soil depth.In addition,only AN is abundant in Luoding City,and other soil indicators are relatively lacking,especially AP and TP are extremely lacking. |
