Predicting Spatial Variations in Soil Nutrients with Hyperspectral Remote Sensing at Regional Scale

Ying-Qiang Song, Xin Zhao, Hui-Yue Su, Bo Li, Yue-Ming Hu, Xue-Sen Cui2018 · DOI

TL;DR

This paper presents a novel approach to predict soil nutrient levels (Nitrogen, Phosphorus, Potassium) across a region using hyperspectral remote sensing data. By combining advanced machine learning models with spatial interpolation techniques, the study achieves significantly improved accuracy in mapping soil nutrient distribution, offering a more efficient and effective method for agricultural management.

What Problem Does This Paper Address?

Accurately understanding the spatial distribution of soil nutrients is crucial for optimizing agricultural productivity, ensuring food security, and implementing sustainable fertilization practices. Traditional methods of soil nutrient assessment rely on extensive field sampling and laboratory analysis, which are time-consuming, costly, and provide limited spatial coverage. While remote sensing offers a more efficient alternative, multispectral data has limitations in capturing the subtle spectral variations associated with soil nutrient content. Hyperspectral remote sensing, with its finer spectral resolution, holds promise for more precise nutrient prediction, but effectively integrating this data with predictive models, especially at a regional scale, remains a challenge. This paper aims to bridge this gap by exploring and comparing various machine learning and hybrid models for predicting soil nutrients using hyperspectral data.

Methods

The study employed a multi-faceted approach to predict soil nutrient variations:

  • Data Preprocessing and Dimensionality Reduction: Hyperspectral images were pre-processed to remove atmospheric and sensor errors. To handle the high dimensionality and redundancy of hyperspectral bands, Pearson correlation analysis and Principal Component Analysis (PCA) were used to select the most informative spectral variables.
  • Prediction Models:
    • Linear Model: Stepwise Linear Regression (SLR) was used to establish linear relationships between selected spectral variables and soil nutrients.
  • Non-linear Machine Learning Models: Support Vector Machine (SVM), Random Forest (RF), and a Back-Propagation Neural Network (BPNN) with two hidden layers were employed to capture complex non-linear relationships.
  • Hybrid Model: A novel hybrid approach, Back-Propagation Neural Network-Ordinary Kriging (BPNNOK), was developed. This model combines the predictive power of the BPNN with the spatial interpolation capabilities of Ordinary Kriging to account for both spectral information and spatial autocorrelation of soil nutrients.
  • Validation: The predictive performance of all models was rigorously evaluated using an independent validation dataset and metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), and Ratio of Performance to Deviation (RPD).

Datasets

  • Soil Samples: 1297 soil samples were collected from Zengcheng, north of the Pearl River Delta, China. These samples were analyzed for soil total nitrogen (TN), soil available phosphorus (AP), and soil available potassium (AK).
  • Hyperspectral Remote Sensing Images: Images from the Chinese Environmental 1A satellite, with 115 bands covering the spectral range of 450–950 nm and a spatial resolution of 100 m × 100 m, were used. The images were acquired in October 2008, coinciding with the field sampling period.
  • Validation Dataset: 324 independent soil samples were randomly selected from the total dataset for model validation.

Key Insights and Ideas

  • Hyperspectral Advantage: Hyperspectral data, with its high spectral resolution, offers a significant advantage over multispectral data in capturing subtle spectral changes related to soil nutrients.
  • Non-linearity is Key: The relationship between hyperspectral variables and soil nutrients is predominantly non-linear, making non-linear machine learning models (especially BPNN) more effective than linear models (SLR).
  • Hybridization for Superior Performance: Combining a powerful non-linear predictor (BPNN) with a spatial interpolation technique (Ordinary Kriging) in the BPNNOK model significantly enhances predictive accuracy and spatial mapping performance by accounting for both spectral information and spatial dependencies.
  • Importance of Specific Bands: Certain spectral bands within the visible and near-infrared (VNIR) region (specifically 464–517 nm) were identified as most sensitive and responsive to soil nutrient variations, providing insights for future sensor design and band selection.
  • Regional Scale Applicability: The study demonstrates the feasibility of using hyperspectral remote sensing for rapid, efficient, and accurate mapping of soil nutrients at a regional scale, which is critical for effective agricultural management.

Key Results

  • The BPNNOK model achieved the highest predictive accuracy among all tested models for all three soil nutrients:
  • Soil TN: RMSE = 0.292 mg kg⁻¹, R² = 68.51%
    • Soil AP: RMSE = 29.62 mg kg⁻¹, R² = 69.30%
    • Soil AK: RMSE = 49.67 mg kg⁻¹, R² = 70.55%
  • The BPNN model (without kriging) also showed strong performance, outperforming SLR, SVM, and RF, with R² values ranging from 44.24% (TN) to 48.53% (AK).
  • The SLR model exhibited the lowest predictive accuracy, with R² values below 25% for all nutrients, highlighting the limitations of linear models.
  • Spatial mapping results from the BPNNOK model indicated higher concentrations of soil TN in the north-central region, soil AP in the central and southwest regions, and soil AK in the central and southeast regions.

Limitations and Future Work

  • Limited Spectral Range: The study primarily utilized visible-near-infrared (VNIR) bands. Incorporating shortwave-infrared (SWIR) bands could potentially further improve prediction accuracy.
  • Environmental Factors: While the study acknowledges the influence of elevation and human agricultural management, future work could explore integrating these and other environmental factors (e.g., climate, topography, vegetation indices) more comprehensively with hyperspectral data for even more robust predictions.
  • Model Complexity vs. Efficiency: While the BPNN and BPNNOK models demonstrated superior accuracy, their computational complexity might be a consideration for very large-scale applications. Further research into optimizing these models for efficiency is warranted.
  • Generalizability: The study was conducted in a specific region of China. Further validation across diverse geographical and agricultural settings would be beneficial to assess the generalizability of the proposed methods.

Summary updated

BibTeX
@misc{song2018predicting,
  author = {Ying-Qiang Song and Xin Zhao and Hui-Yue Su and Bo Li and Yue-Ming Hu and Xue-Sen Cui},
  title = {{Predicting Spatial Variations in Soil Nutrients with Hyperspectral Remote Sensing at Regional Scale}},
  year = {2018},
  doi = {10.3390/s18093086},
}
Figure