基于GF-1遥感数据预测区域森林土壤有机质含量

Prediction of Soil Organic Matter Content Based on Artificial Neural Network Model and GF-1 Remote Sensing Data

为探索国产卫星GF-1预测土壤有机质(SOM)的能力,本研究以广东省云浮市的罗定市为研究区,以GF-1多光谱遥感影像衍生的9个遥感变量和DEM提取的9个地形水文变量为预测因子,建立2种人工神经网络模型(A模型:地形水文;B模型:地形水文+遥感),对5个土壤深度(L1:0~20 cm,L2:20~40 cm,L3:40~60 cm,L4:60~80 cm,L5:80~100 cm)的SOM进行预测。结果表明:5个深度的B模型全都比A模型的精度高,尤其是L1、L2土层,精度提升明显,其R^(2)分别提高了13%和10%;而深层土壤(L3、L4、L5)的精度提升较小,仅提高了4%、5%和4%。另外,两个评价指标RMSE和ROA±10%也表现出相似的趋势。总体而言,GF-1遥感数据显著改善了上层(0~40 cm)森林土壤人工神经网络模型的预测精度,对下层(40~100 cm)森林土壤模型改善尺度较低,是预测森林土壤SOM含量可观的新遥感数据源。

To explore the capability of GF-1 satellite to predict soil organic matter(SOM),Luoding City of Yunfu City,Guangdong Province was taken as the study area,and 9 multi-spectral remote sensing variables retrieved from GF-1 and 9 terrain variables derived from DEM were used as predictors to establish two kinds of artificial neural network models(Model A:terrain;Model B:terrain&remote sensing)for predicting soil organic matter(SOM)at five soil depths(L1:0–20 cm,L2:20–40 cm,L3:40–60 cm,L4:60–80 cm,and L5:80–100 cm).The results showed that the accuracies of SOM full-variable B model at five depths was higher than those of A model with topographic variables only.Especially for the L1 and L2 layers of soil,the accuracy was obviously improved.The R^(2) of the L1 and L2 layers of SOM were increased by 13%and 10%respectively.However,the accuracies of deep soils(L3,L4,L5)were only improved by 4%,5%and 4%,respectively,and RMSE and ROA±10%also showed a similar trend.The results show that GF-1 remote sensing image can be used as a new data source to predict SOM.