关键词: 煤矿区, 土壤全磷, 土壤速效磷, 高光谱反演, 卷积神经网络

Abstract: Soil phosphorus management is an issue that be worthy of attention in the improvement of cultivated land quality and the protection of water environment in coal mining areas. With hyperspectral remote sensing technology, exploring the accurate inversion method of soil phosphorus content can facilitate the monitoring of cultivated land soil phosphorus content in coal mining areas. In this study, we collected soil samples of cultivated land from the Macun mining area in Jiaozuo, Henan Province, and got the spectral data of soil samples by a portable ground object spectrometer. The original spectrum and its decomposed data by continuous wavelet transform (CWT) were selected as independent variables, while the feature wavebands were selected by the competitive adaptive reweighted sampling (CARS). The modeling effects of deep learning \[convolutional neural network (CNN)\] and common machine learning methods \[random forest (RF) and BP neural network (BPNN)\] were compared and analyzed. Results showed that CWT method enhanced the correlation between spectral reflectance and soil phosphorus content, and thus improved the model accuracy. CNN exhibited a strong feature of learning ability, and its prediction accuracy was much higher than those of BPNN and RF. Combining the CARS algorithm for feature wavebands screening and the CNN for feature extraction could further improve the efficiency of the model. The relative percentage difference (RPD) of the optimal model of soil total phosphorus and available phosphorus content were increased by 7.53 and 1.25, respectively. Among the wavelet transform L₁-L₁₀ data filtered by CARS, L₉-CARS-CNN model had the highest accuracy in the soil total phosphorus and available phosphorus. The RMSE and RPD of its verification set were 1.33 mg·kg^-1 and 25.29 for soil total phosphorus, and 1.52 mg·kg^-1 and 11.95 for available phosphorus, respectively. Therefore, combining the CARS for feature wavebands extraction and the CNN for model building can accurately retrieve soil phosphorus content, which will have broad application prospects.

Key words: coal mining area, soil total phosphorus, soil available phosphorus, hyperspectral inversion, convolutional neural network