采煤拉张裂隙区土壤有效磷高光谱估测

Hyperspectral Estimation of Soil Available Phosphorus in Coal Mining-Induced Tensile Fissure Zones

关键词:  土壤有效磷  采煤拉张裂隙  竞争性自适应重加权采样算法  高光谱遥感  估测模型

Key words:  Soil Available Phosphorus (AP)  Coal Mining-induced Tensile Fissure  Competitive Adaptive Reweighted Sampling (CARS) Algorithm  Hyperspectral Remote Sensing  Estimation Model

摘要:

【目的】采煤活动引发的地表裂隙破坏土壤结构,影响土壤理化性质与养分分布,严重制约了矿区农业可持续发展。开展采煤拉张裂隙区土壤有效磷(Available Phosphorus, AP)含量快速估测,对于矿区复垦土壤肥力评价和养分管理具有重要意义。【方法】以淮北朱庄煤矿3522工作面采动裂隙区为研究区,采集了90个土壤样本,测定了其有效磷含量与高光谱数据,设计了9种光谱预处理组合,并以此分别建立偏最小二乘回归(Partial Least Squares Regression, PLSR)、随机森林(Random Forest, RF)和支持向量回归(Support Vector Regression, SVR)三种全波段估测模型。同时,基于各模型的最优预处理组合,采用竞争性自适应重加权采样算法(Competitive Adaptive Reweighted Sampling, CARS)进行特征波段筛选,对比分析全波段与特征筛选模型的预测性能。【结果】实验结果表明:全波段建模中,PLSR模型表现最优,最优预处理为FD-SG-SNV,验证集决定系数R2为0.694,预测残差偏差比RPD为1.843；RF和SVR模型的最优R2分别为0.476、0.665。经CARS特征筛选后,PLSR模型验证集R2降至0.525,而SVR和RF精度略有提升,CARS-SVR验证集R2为0.669,RPD为1.770。【结论】在本研究样本条件下,全波段PLSR模型表现相对较优,说明其对小样本高维光谱数据中分散弱信息具有一定整合能力。CARS-SVR模型在特征筛选后仍保持了接近全波段模型的预测效果,表明特征筛选对非线性模型具有一定的辅助作用。研究结果可为采煤拉张裂隙区土壤有效磷高光谱估测及建模策略选择提供依据。

Abstract:

[Purpose] Coal mining activities induce surface fissures and soil structural damage, altering soil physical and chemical properties and disrupting nutrient distribution, which seriously hinders ecological restoration and sustainable agricultural development in mining areas. As a key limiting nutrient for crop growth, soil available phosphorus (AP) is a critical indicator for evaluating reclaimed soil fertility. Therefore, rapid and accurate estimation of AP content in mining tensile fissure zones is of great significance for precise soil nutrient management in post-mining lands. [Method] This study selected the mining-induced tensile fissure zone of the 3522 working face in Zhuzhuang Coal Mine, Huaibei, as the study area. A total of 90 soil samples were collected to determine AP content and hyperspectral reflectance. Nine spectral preprocessing combinations were designed to optimize spectral data quality. Three mainstream modeling algorithms, including partial least squares regression (PLSR), random forest (RF), and support vector regression (SVR), were used to construct full-band AP estimation models. Based on the optimal preprocessing scheme of each model, the competitive adaptive reweighted sampling (CARS) algorithm was applied to screen characteristic spectral bands and remove redundant variables. The predictive performance of full-band and feature-screened models was systematically compared. [Result] The results showed that the full-band PLSR model based on FD-SG-SNV preprocessing obtained the best performance, with a validation R2 of 0.694 and an RPD of 1.843. The optimal validation R2 of RF and SVR full-band models were 0.476 and 0.665, respectively. After CARS feature selection, the validation accuracy of PLSR decreased to R2 of 0.525, while the prediction performance of nonlinear SVR and RF models was slightly improved. The CARS-SVR model maintained favorable prediction accuracy, with a validation R2 of 0.669 and an RPD of 1.770. [Conclusion] The full-band PLSR model exhibited superior performance in this study, demonstrating its capability to integrate dispersed weak spectral information from small-sample and high-dimensional spectral data. The CARS-SVR model achieved stable prediction results comparable to full-band modeling, indicating that feature selection can effectively optimize the performance of nonlinear machine learning models. This study provides a reliable reference for hyperspectral estimation of soil available phosphorus and scientific model selection in coal mining-induced fissure areas.