Predicting Soil Available Phosphorus by Hyperspectral Regression Method Based on Gradient Boosting Decision Tree

Xiu Jin; Xianzhi Zhu; Shaowen Li; Wencai Wang; Haijun Qi

doi:10.3788/LOP56.131102

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 13 >Page 131102 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 13, 131102 (2019)

Predicting Soil Available Phosphorus by Hyperspectral Regression Method Based on Gradient Boosting Decision Tree

Xiu Jin, Xianzhi Zhu, Shaowen Li^*, Wencai Wang, and Haijun Qi

Author Affiliations

School of Information & Computer, Anhui Agricultural University, Hefei, Anhui 230036, China

show less

DOI: 10.3788/LOP56.131102 Cite this Article Set citation alerts

Xiu Jin, Xianzhi Zhu, Shaowen Li, Wencai Wang, Haijun Qi. Predicting Soil Available Phosphorus by Hyperspectral Regression Method Based on Gradient Boosting Decision Tree[J]. Laser & Optoelectronics Progress, 2019, 56(13): 131102 Copy Citation Text

show less

Fig. 1. Stacking method

Download full size

Fig. 2. Indoor hyperspectral acquisition system

Download full size

Fig. 3. Hyperspectral reflectance of soil. (a) Original spectra; (b) smoothing spectra

Download full size

Fig. 4. fRMSE values of different LV numbers in linear and nonlinear PLS

Download full size

Fig. 5. Parameter optimization of GBDT model. (a) Rloss=Fls, Rn_estimators=100; (b) Rloss=Fhuber, Rn_estimators=200; (c) Rloss=Fquantile, Rn_estimators=200; (d) Rloss=Flad, Rn_estimators=310

Download full size

Fig. 6. Results of different model integration algorithms. (a) Results of random forest based on modeling set; (b) results of random forest based on testing set; (c) results of boosting tree based on modeling set; (d) results of boosting tree based on testing set; (e) results of GBDT based on modeling set; (f) results of GBDT based on testing set

Download full size

Type	Sample	Max /(mg·kg^-1)	Min /(mg·kg^-1)	Average /(mg·kg^-1)	Standard deviation /(mg·kg^-1)
Total	193	34.96	0.03	10.56	9.36
Training	144	34.96	0.03	10.94	9.49
Testing	49	32.24	0.60	9.01	8.99

Table 1. Statistical parameters of soil available phosphorus content

Modeling method	Training set		Testing set		Prediction level(testing set)	Parameter
Modeling method	f_RPD			R²	f_RPD	R²
PLS	1.66	0.73	1.65	0.68	B	R_LVs=11
RBF-PLS	1.58	0.71	1.79	0.73	B	R_LVs=11, R_gamma =0.016
Sigmoid-PLS	1.55	0.70	1.75	0.73	B	R_LVs =10, $\begin{matrix} {R_{gamma}}_{} \end{matrix}$ =0.00085,R_cofe0cofe0=4.5
SVR	1.60	0.74	1.53	0.69	B	C=10000
RBF-SVR	1.70	0.76	1.66	0.72	B	C=2000000, R_gamma =0.0028
Sigmoid-SVR	1.59	0.73	1.55	0.70	B	C=10¹¹, R_gamma =0.000001,R_cofe0=0
Ridge	1.60	0.74	1.50	0.69	B	R_Alpha=0.001
RBF-Ridge	1.55	0.74	1.50	0.70	B	R_Alpha=0.00006, R_gamma =0.01
Sigmoid-Ridge	1.52	0.72	1.50	0.69	B	R_Alpha=4×10^-7,R_gamma =0.0005, R_cofe0=0.9

Table 2. Testing results of optimal single model

Ensemble method	Training set		Testing set		Prediction level(testing set)	Parameter
Ensemble method	f_RPD			R²	f_RPD	R²
Random forest	2.10	0.84	2.08	0.84	A	R_{n_estimators}, R_{max_depth}=5
Boosting tree	2.86	0.90	2.12	0.82	A	R_{n_estimators} =300, R_{learning_rate}=0.01,R_{max_depth}=5, R_loss=F_linear
GBDT	2.56	0.88	2.55	0.86	A	R_{n_estimators} =310, R_{learning_rate}=0.29,R_{max_depth}=4, R_loss=F_lad