img

Wechat

Adv search

Journal of Desert Research ›› 2025, Vol. 45 ›› Issue (2): 191-204.DOI: 10.7522/j.issn.1000-694X.2024.00115

Previous Articles    

Hyperspectral remote sensing estimation of Fe2O3 content in desert soil based on dandelion-optimized random forest model

Xin Hu1(), Sawut Mamat1(), Feng Zhang1, Jintao Cui1, Aimaier Ainiwan2, Manlike Asiya2   

  1. 1.College of Geography and Remote Sensing Sciences / MOE Key Laboratory of Oasis Ecology / Key Laboratory of Smart City and Environment Modelling of Higher Education Institute / Xinjiang Field Scientific Observation and Research Station for the Oasisization Process in the Hinterland of Taklamakan Desert,Xinjiang University,Urumqi 830017,China
    2.Grassland Research Institute / Field Orientation Observation and Research Station of Grassland Ecological Environment on the Northern Slope of Tianshan Mountains,Xinjiang Academy of Animal Science,Urumqi 830057,China
  • Received:2024-07-30 Revised:2024-12-03 Online:2025-03-20 Published:2025-03-26
  • Contact: Sawut Mamat

Abstract:

The relationship between desert soil spectra reflectance and iron oxide (Fe2O3) content remains unclear, and effective monitoring methods are lacking. In this study, we focus on the Gurbantunggut Desert in Xinjiang, where desert soil samples were collected to obtain both Fe2O3 content and spectral data. Using fractional order differentiation (FOD) and continuous wavelet transform (CWT) to preprocess the original spectral data, we performed correlation analysis to identify the optimal spectral transformations for estimating Fe2O3 content in desert soils. Genetic algorithms (GA) were employed to extract sensitive spectral bands, and a Dandelion Optimization-based Random Forest (DO-RF) model was developed for Fe2O3 content estimation. The results indicate the following: (1) With the increase of Fe2O3 content, the reflectance of desert soil gradually decreases, showing a negative correlation between Fe2O3 content and soil spectral reflectance; (2) Both FOD and CWT can enhance the correlation between desert soil reflectance and its Fe2O3 content. Specifically, the highest correlations are achieved with FOD at the 1.2 order and CWT at a scale of 1, reaching 0.840 and 0.839 respectively; (3) GA effectively eliminates highly collinear redundant bands. Under a 1.2-order Fractional Order Derivative (FOD), it selects 31 optimal feature bands from 512 spectral bands, compressing them by 93.945%. Similarly, under a 1-scale Continuous Wavelet Transform (CWT), it identifies 13 optimal feature bands from 119 spectral bands, achieving an 89.076% compression; (4) The DO-RF model based on CWT processing exhibits the best accuracy and stability. The model validation coefficient of determination (R2) reaches 0.908, the root mean square error (RMSE) is 0.340, and the relative prediction deviation (RPD) is 3.390. Compared to the unoptimized RF, PLSR, and SVM, the R2 increases by 2.7%, 22.6%, and 4%, while the RMSE decreases by 6.6%, 27.8%, and 8.7%, and the RPD increases by 54.9%, 152.2%, and 68.6% respectively. These findings can serve as a reference for future satellite spectral remote sensing monitoring of Fe2O3 content in desert soil.

Key words: desert soil, Fe2O3 content, hyperspectral remote sensing, dandelion optimization, random forest

CLC Number: