基于XGBoost模型的湟水流域耕地土壤养分遥感反演

doi:10.12118/j.issn.1000-6060.2023.034

摘要/Abstract

摘要：

湟水流域是河湟谷地重要的组成部分，协同环境因素预测土壤养分空间分布对农业土壤养分管理尤为重要。土壤养分反演研究中对于参数对模型结果的影响和模型适用性的研究较少。选取研究区地形因子、土壤pH及光谱反射率共28个因子，结合贝叶斯优化算法构建人工神经网络（ANN）、支持向量机（SVM）和极端梯度提升（XGBoost）3种机器学习模型预测耕地土壤养分空间分布，计算决定系数（R²）、均方根误差（RMSE）和相对分析误差（RPD）评价3种模型的精度。结果表明：（1）基于贝叶斯优化超参数的XGBoost模型对全氮（TN）含量预测精度优于其他模型（R²=0.893，RMSE=0.359，RPD=2.470），预测土壤有机质（SOM）、速效磷（AP）和速效钾（AK）含量时，XGBoost模型验证集R²分别为0.801、0.509、0.442。（2）对比3种模型的寻优次数和误差发现，BOA-XGBoost模型参数优化次数少、效率高，具有更好的鲁棒性。对于不同的养分，ANN和SVM模型预测精度存在差异，SVM模型预测SOM含量时精度更高（RPD=1.580），而ANN模型预测TN时精度最佳（RPD=2.460）。基于贝叶斯算法进行超参数优化构建的XGBoost模型预测精度高，可以达到良好的预测效果，可为湟水流域精准农业施肥提供参考。

关键词: 土壤养分, XGBoost, 空间分布, 环境因子, 湟水流域

Abstract:

The Huangshui River Basin is an important part of the Huangshui Valley. Additionally, collaborative environmental factors that predict the spatial distribution of soil nutrients are particularly important for managing soil nutrients. Moreover, less attention is paid to the effect of model parameters on the results obtained from soil nutrient inversion studies. In this study, the Huangshui River Basin in Qinghai Province (China) was selected as the study area, and 28 factors, including elevation, aspect, slope, plane curvature, section curvature, relief degree of land surface, topographic wetness index, soil pH, and spectral reflectance, were selected. In addition, these factors were combined with the Bayesian optimization algorithm (BOA) to construct artificial neural network (ANN), support vector machine (SVM), and extreme gradient boosting (XGBoost) machine learning models for predicting the spatial distribution of four soil nutrients in farmlands: soil organic matter (SOM), total nitrogen (TN), available phosphorus (AP), and available potassium (AK), respectively. Further, the prediction accuracy of these three models was evaluated based on the model coefficient of determination (R²), root-mean-square error (RMSE), and relative percent deviation (RPD). The results revealed that: (1) All four soil nutrients exhibited a moderate degree of variability, with TN showing the highest variability of 69.481%. The XGBoost model based on the Bayesian optimized hyperparameter combination was better than other models in predicting the TN content (R², RMSE, and RPD were 0.893, 0.359, and 2.470, respectively). The R² values of the XGBoost model validation set for estimating the SOM, AK, and AP contents were 0.801, 0.509, and 0.442, respectively, and the corresponding RPD values were 2.152, 1.210, and 1.274, respectively. Moreover, this model exhibited a better prediction capability. (2) The comparison of the number of optimizations and errors of the three models revealed that the BOA-XGBoost model exhibited minimum number of parameter optimizations, higher efficiency, and better robustness. The ANN and SVM models demonstrated different prediction accuracies for different nutrients; additionally, the SVM model predicted the SOM content with high accuracy (RPD=1.580), while the ANN model predicted TN efficiently (RPD=2.460). Based on Landsat 8 remote sensing images, the XGBoost inversion model developed by combining 28 factors of the Huangshui River Basin was found to be more suitable for application in soil nutrient inversion research; furthermore, it can more accurately describe the spatial distribution pattern of the soil nutrient inversion in the Huangshui River Basin, better ensure precise agriculture fertilization, improve the fertilizer utilization rate and crop yield, and provide a reference for precise agriculture fertilization in the Huangshui River Basin.

Key words: soil nutrient, XGBoost, spatial distribution, environmental factor, Huangshui River Basin

刘尊方, 雷浩川, 盛海彦. 基于XGBoost模型的湟水流域耕地土壤养分遥感反演[J]. 干旱区地理, 2023, 46(10): 1643-1653.

LIU Zunfang, LEI Haochuan, SHENG Haiyan. Remote sensing inversion of soil nutrient on farmland in Huangshui River Basin based on XGBoost model[J]. Arid Land Geography, 2023, 46(10): 1643-1653.

图/表 11

图1

图2

表1

表2

表3

表4

图3

表5

表6

图4

图5

参考文献 34

[1]	Wang R M, Zou R Y, Liu J M, et al. Spatial distribution of soil nutrients in farmland in a hilly region of the Pearl River Delta in China based on geostatistics and the inverse distance weighting method[J]. Agriculture, 2021, 11(1): 50, doi: 10.3390/agriculture11010050.
[2]	Alemu L, Mesfin B. Performance of mid infrared spectroscopy to predict nutrients for agricultural soils in selected areas of Ethiopia[J]. Heliyon, 2022, 8(3): e09050, doi: 10.1016/j.heliyon.2022.e09050.
[3]	刘焕军, 张美薇, 杨昊轩, 等. 多光谱遥感结合随机森林算法反演耕作土壤有机质含量[J]. 农业工程学报, 2020, 36(10): 134-140.
	[Liu Huanjun, Zhang Meiwei, Yang Haoxuan, et al. Invertion of cultivated soil organic matter content combining multi-spectral remote sensing and random forest algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering, 2020, 36(10): 134-140. ]
[4]	郑淼, 王翔, 李思佳, 等. 黑土区土壤有机质和全氮含量遥感反演研究[J]. 地理科学, 2022, 42(8): 1336-1347. doi: 10.13249/j.cnki.sgs.2022.08.002
	[Zheng Miao, Wang Xiang, Li Sijia, et al. Remote sensing inversion of soil organic matter and total nitrogen in black soil region[J]. Scientia Geographica Sinica, 2022, 42(8): 1336-1347. ] doi: 10.13249/j.cnki.sgs.2022.08.002
[5]	Gulhane V A, Rode S V, Pande C B. Wavelet for predicting soil nutrients using remotely sensed satellite images[J]. International Journal of Computer Applications, 2017, 174(4): 35-38.
[6]	李冠稳, 高小红, 杨灵玉, 等. 不同粒径土壤有机质含量可见光-近红外光谱估算研究——以湟水流域为例[J]. 土壤通报, 2017, 48(6): 1360-1370.
	[Li Guanwen, Gao Xiaohong, Yang Lingyu, et al. Estimating soil organic matter contents from different soil particle size using visible and near-infrared reflectance spectrum: A case study of the Huangshui Basin[J]. Chinese Journal of Soil Science, 2017, 48(6): 1360-1370. ]
[7]	雷浩川, 刘尊方, 于晓晶, 等. 基于Landsat 5影像的青海省大通县土壤表层全氮空间格局反演[J]. 青海大学学报, 2021, 39(6): 79-86.
	[Lei Haochuan, Liu Zunfang, Yu Xiaojing, et al. Spatial pattern inversion of soil surface total nitrogen in Datong County of Qinghai Province based on Landsat 5 image[J]. Journal of Qinghai University, 2021, 39(6): 79-86. ]
[8]	肖云飞. 青海省土壤有机质、全碳、全氮高光谱遥感估算研究[D]. 西宁: 青海师范大学, 2019.
	[Xiao Yunfei. Estimation of soil organic matter, total carbon and total nitrogen by hyperspectral remote sensing in Qinghai Province[D]. Xining: Qinghai Normal University, 2019. ]
[9]	李冠稳, 高小红, 肖能文, 等. 基于sCARS-RF算法的高光谱估算土壤有机质含量[J]. 发光学报, 2019, 40(8): 1030-1039. doi: 10.3788/fgxb
	[Li Guanwen, Gao Xiaohong, Xiao Nengwen, et al. Estimation soil organic matter contents with hyperspectra based on sCARS and RF algorithms[J]. Chinese Journal of Luminescence, 2019, 40(8): 1030-1039. ] doi: 10.3788/fgxb
[10]	杨晓宇, 包妮沙, 曹粤, 等. 基于无人机成像光谱技术的农田土壤养分估测及制图[J]. 地理与地理信息科学, 2021, 37(5): 38-45.
	[Yang Xiaoyu, Bao Nisha, Cao Yue, et al. Estimation and mapping of soil nutrient in farmland based on UAV imaging spectrometry[J]. Geography and Geo-information Science, 2021, 37(5): 38-45. ]
[11]	郑曼迪, 熊黑钢, 乔娟峰, 等. 基于高光谱的不同人类干扰程度下荒漠土壤有机质含量估算模型[J]. 干旱区地理, 2018, 41(2): 384-392.
	[Zheng Mandi, Xiong Heigang, Qiao Juanfeng, et al. Hyperspectral based estimation model about organic matter in desert soil at different levels of human disturbance[J]. Arid Land Geography, 2018, 41(2): 384-392. ]
[12]	Miran N, Rasouli Sadaghiani M H, Feiziasl V, et al. Predicting soil nutrient contents using Landsat OLI satellite images in rain-fed agricultural lands, northwest of Iran[J]. Environmental Monitoring and Assessment, 2021, 193(9): 607, doi: 10.1007/s10661-021-09397-0. pmid: 34455498
[13]	张欢, 高小红. 复杂地形区土壤有机质空间变异性分析及制图[J]. 水土保持研究, 2020, 27(5): 93-100.
	[Zhang Huan, Gao Xiaohong. Analysis of spatial variability and mapping of soil organic matter contents in complex terrain areas[J]. Research of Soil and Water Conservation, 2020, 27(5): 93-100. ]
[14]	Liu Z F, Lei H C, Lei L, et al. Spatial prediction of total nitrogen in soil surface layer based on machine learning[J]. Sustainability, 2022, 14(19): 11998, doi: 10.3390/su141911998.
[15]	Dharumarajan S, Lalitha M, Niranjana K, et al. Evaluation of digital soil mapping approach for predicting soil fertility parameters: A case study from Karnataka Plateau, India[J]. Arabian Journal of Geosciences, 2022, 15(5): 386, doi: 10.1007/s12517-022-09629-8.
[16]	杨练兵, 陈春波, 郑宏伟, 等. 基于优化随机森林回归模型的土壤盐渍化反演[J]. 地球信息科学学报, 2021, 23(9): 1662-1674. doi: 10.12082/dqxxkx.2021.200711
	[Yang Lianbing, Chen Chunbo, Zheng Hongwei, et al. Retrieval of soil salinity content based on random forests regression optimized by Bayesian optimization algorithm and genetic algorithm[J]. Journal of Geo-information Science, 2021, 23(9): 1662-1674. ] doi: 10.12082/dqxxkx.2021.200711
[17]	Wang L Y, Hu P, Zheng H W, et al. Integrative modeling of heterogeneous soil salinity using sparse ground samples and remote sensing images[J]. Geoderma, 2023, 430: 116321, doi: 10.1016/j.geoderma.2022.116321.
[18]	Liu F, Qin T L, Yan D H, et al. Classification of instream ecological water demand and crucial values in a semi-arid river basin[J]. Science of the Total Environment, 2020, 712: 136409, doi: 10.1016/j.scitotenv.2019.136409.
[19]	Dong B Q, Qin T L, Wang Y, et al. Spatiotemporal variation of nitrogen and phosphorus and its main influencing factors in Huangshui River Basin[J]. Environmental Monitoring and Assessment, 2021, 193(5): 292, doi: 10.1007/s10661-021-09067-1. pmid: 33891180
[20]	Ruhollah T M, Ram N, Sood K, et al. Artificial bee colony feature selection algorithm combined with machine learning algorithms to predict vertical and lateral distribution of soil organic matter in South Dakota, USA[J]. Carbon Management, 2017, 8(3): 277-291. doi: 10.1080/17583004.2017.1330593
[21]	杨欣, 袁自然, 叶寅, 等. 基于无人机高光谱遥感的冬小麦全氮含量反演[J]. 光谱学与光谱分析, 2022, 42(10): 3269-3274.
	[Yang Xin, Yuan Ziran, Ye Yin, et al. Winter wheat total nitrogen content estimation based on UAV hyperspectral remote sensing[J]. Spectroscopy and Spectral Analysis, 2022, 42(10): 3269-3274. ]
[22]	Miao J, Zhen J N, Wang J J, et al. Mapping seasonal leaf nutrients of mangrove with Sentinel-2 images and XGBoost method[J]. Remote Sensing, 2022, 14(15): 3679, doi: 10.3390/rs14153679.
[23]	Komolafe A A, Olorunfemi I E, Oloruntoba C, et al. Spatial prediction of soil nutrients from soil, topography and environmental attributes in the northern part of Ekiti State, Nigeria[J]. Remote Sensing Applications: Society and Environment, 2021, 21: 100450, doi: 10.1016/j.rsase.2020.100450.
[24]	代子俊. 近30年湟水流域土壤养分时空变异及影响因素[D]. 西宁: 青海师范大学, 2018.
	[Dai Zijun. Spatio-temporal variation of soil nuterents in Huangshui River Basin and its affecting factors in the past 30 years[D]. Xining: Qinghai Normal University, 2018. ]
[25]	方慧婷, 蒙继华, 程志强. 基于遥感与作物模型的土壤速效养分时空变异分析[J]. 中国农业科学, 2019, 52(3): 478-490. doi: 10.3864/j.issn.0578-1752.2019.03.008
	[Fang Huiting, Meng Jihua, Cheng Zhiqiang. Spatio-temporal variability of soil available nutrients based on remote sensing and crop model[J]. Scientia Agricultura Sinica, 2019, 52(3): 478-490. ] doi: 10.3864/j.issn.0578-1752.2019.03.008
[26]	刘尊方, 雷浩川, 雷蕾. 湟水流域土壤有机质和速效磷空间布局分析[J]. 科学技术与工程, 2022, 22(34): 15095-15102.
	[Liu Zunfang, Lei Haochuan, Lei Lei. Analysis on spatial distribution of soil organic matter and available phosphorus in Huangshui River Basin[J]. Science Technology and Engineering, 2022, 22(34): 15095-15102. ]
[27]	Chen B L, Zheng H W, Luo G P, et al. Adaptive estimation of multi-regional soil salinization using extreme gradient boosting with Bayesian TPE optimization[J]. International Journal of Remote Sensing, 2022, 43(3): 778-811. doi: 10.1080/01431161.2021.2009589
[28]	胡亚男, 高小红, 申振宇, 等. 基于野外实测Vis-NIR光谱的土壤肥力估算研究——以湟水流域为例[J]. 土壤通报, 2021, 52(3): 575-584. doi: 10.1046/j.1365-2389.2001.00419.x
	[Hu Ya’nan, Gao Xiaohong, Shen Zhenyu, et al. Estimating fertility index by using field-measured Vis-NIR spectroscopy in the Huangshui River Basin[J]. Chinese Journal of Soil Science, 2021, 52(3): 575-584. ] doi: 10.1046/j.1365-2389.2001.00419.x
[29]	Ji W S, Liu Y Q. Research on quantitative evaluation of remote sensing and statistics based on wireless sensors and farmland soil nutrient variability[J]. Computational Intelligence and Neuroscience, 2022, 2022: 3646264, doi: 10.1155/2022/3646264.
[30]	郭静, 龙慧灵, 何津, 等. 基于Google Earth Engine和机器学习的耕地土壤有机质含量预测[J]. 农业工程学报, 2022, 38(18): 130-137.
	[Guo Jing, Long Huiling, He Jin, et al. Predicting soil organic contents in cultivated land using Google Earth Engine and machine learning[J]. Transactions of the Chinese Society of Agricultural Engineering, 2022, 38(18): 130-137. ]
[31]	胡贵贵, 杨粉莉, 杨联安, 等. 基于主成分和机器学习的土壤有机质含量空间预测建模[J]. 干旱区地理, 2021, 44(4): 1114-1124.
	[Hu Guigui, Yang Fenli, Yang Lian’an, et al. Spatial prediction modeling of soil organic matter content based on principal components and machine learning[J]. Arid Land Geography, 2021, 44(4): 1114-1124. ]
[32]	李莹莹, 赵正勇, 杨旗, 等. 基于GF-1遥感数据预测区域森林土壤有机质含量[J]. 土壤, 2022, 54(1): 191-197.
	[Li Yingying, Zhao Zhengyong, Yang Qi, et al. Prediction of soil organic matter content based on artificial neural network model and GF-1 remote sensing data[J]. Soils, 2022, 54(1): 191-197. ]
[33]	孙铭岳. 基于多源遥感的黄河三角洲耕地土壤盐分和有机质含量反演[D]. 泰安: 山东农业大学, 2022.
	[Sun Mingyue. Inversion of soil salinity and organic matter on arable land in the Yellow River Delta based on multi-source remote sensing[D]. Tai’an: Shandong Agricultural University, 2022. ]
[34]	杜军, 古军伟, 邱士可, 等. 基于支持向量回归的土壤全氮含量高光谱估测研究[J]. 河南科学, 2020, 38(10): 1585-1590.
	[Du Jun, Gu Junwei, Qiu Shike, et al. Prediction of total nitrogen by using hyperspectral data based on support vector regression[J]. Henan Science, 2020, 38(10): 1585-1590. ]

土壤养分	样点数/个	最小值/g·kg^-1	最大值/g·kg^-1	平均值/g·kg^-1	标准差/g·kg^-1	变异系数/%
SOM	107	4.310	59.789	28.377	11.168	39.356
TN	107	0.372	5.874	1.232	0.856	69.481
AP	107	0.001	0.162	0.055	0.032	58.182
AK	107	0.048	0.488	0.204	0.087	42.647

土壤养分	高程	坡向	坡度	平面曲率	剖面曲率	地形湿度指数	地形起伏度	pH
SOM	0.422^**	0.223^*	-0.022	0.101	0.052^*	0.010^*	0.095^*	-0.338^**
TN	0.595^**	0.238^*	-0.174^*	-0.015	0.089^*	0.052^*	0.276^**	-0.485^**
AP	-0.048	-0.066^*	-0.028	0.080^*	-0.036	-0.048	-0.098^*	0.026
AK	0.052	0.041	-0.080^*	0.029	0.015	-0.015	-0.013	-0.106^*

波段	SOM	TN	AP	AK	波段	SOM	TN	AP	AK
b1	-0.390^**	-0.343^**	0.118^*	-0.039	lg5	-0.283^**	-0.158	0.088	-0.034
b2	-0.391^**	-0.334^**	0.108^*	-0.040	lg6	-0.259^**	-0.185	0.082	-0.125^*
b3	-0.393^**	-0.311^**	0.083	-0.070	lg7	-0.251^**	-0.247^*	0.069	-0.169^*
b4	-0.379^**	-0.280^**	0.085	-0.084^*	1/b1	0.278^**	0.424^**	-0.134	0.046
b5	-0.267^**	-0.084	0.068	-0.041	1/b2	0.334^**	0.485^**	-0.151^*	0.034
b6	-0.194^*	-0.018	0.033	-0.120	1/b3	0.376^**	0.491^**	-0.142^*	0.044
b7	-0.174^*	-0.073	0.008	-0.180^*	1/b4	0.385^**	0.489^**	-0.143^*	0.063
lg1	-0.387^**	-0.495^**	0.150^*	-0.024	1/b5	0.300^**	0.247^*	-0.112	0.029
lg2	-0.392^**	-0.452^**	0.139^*	-0.019	1/b6	0.317^**	0.361^**	-0.128	0.119^*
lg3	-0.394^**	-0.407^**	0.113	-0.051	1/b7	0.316^**	0.422^**	-0.126	0.139^*
lg4	-0.387^**	-0.386^**	0.114	-0.071

模型	参数	土壤养分
模型	参数	SOM	TN	AP	AK
ANN	隐藏层节点/个	80	100	120	120
	最大迭代次数/次	1000	1500	500	500
	激活函数	S型函数	S型函数	修正线性单元函数	修正线性单元函数
SVM	惩罚系数	3.12	2.50	0.35	1.10
SVM	核参数	0.42	0.73	0.91	0.92
XGBoost	树的最大深度	4	5	4	4
	最大树数目	50	30	75	96
	学习率	0.03	0.07	0.05	0.04
	最小叶节点样本权重和	4	4	2	1

土壤养分	模型	建模集		验证集
土壤养分	模型	R²	RMSE	R²	RMSE	RPD
SOM	ANN	0.753	5.363	0.643	7.401	1.214
	SVM	0.822	5.122	0.632	5.648	1.580
	XGBoost	0.910	3.791	0.801	4.321	2.152
TN	ANN	0.885	0.245	0.803	0.306	2.460
	SVM	0.871	0.415	0.735	0.252	1.886
	XGBoost	0.958	0.235	0.893	0.359	2.470
AP	ANN	0.491	0.040	0.382	0.030	1.002
	SVM	0.468	0.023	0.441	0.029	1.213
	XGBoost	0.692	0.022	0.509	0.026	1.210
AK	ANN	0.514	0.063	0.419	0.064	1.321
	SVM	0.486	0.061	0.354	0.072	1.260
	XGBoost	0.692	0.043	0.442	0.055	1.274