TL;DR
This paper introduces a novel hybrid approach, PSO-BPNN, that integrates Particle Swarm Optimization (PSO) with Back Propagation Neural Networks (BPNN) to significantly improve the accuracy of estimating soil heavy metal (Cd, Hg, As) content using hyperspectral data. This method shows great potential for accurate, large-scale soil heavy metal mapping, even when using satellite-based hyperspectral imagery.
What Problem Does This Paper Address?
Monitoring soil heavy metal contamination is crucial for environmental and human health. Traditional methods of soil sampling and laboratory analysis are time-consuming, costly, and provide limited spatial coverage. While hyperspectral remote sensing offers a promising alternative for large-area estimation, existing machine learning models, like standard BPNN, often suffer from lower accuracy due to difficulties in optimizing initial parameters and potential local minima issues. This paper aims to bridge this gap by developing a more robust and accurate method for estimating soil heavy metal concentrations from hyperspectral data.
Methods
The core of this research is the development of an integrated PSO-BPNN model. Here's a breakdown:
- Back Propagation Neural Network (BPNN): This is a standard supervised learning algorithm used for its ability to model complex, non-linear relationships between input spectral data and soil heavy metal content. It consists of an input layer, a hidden layer, and an output layer.
- Particle Swarm Optimization (PSO): This is a metaheuristic optimization algorithm inspired by the social behavior of bird flocking or fish schooling. In this study, PSO is used to optimize the initial weights and thresholds of the BPNN. This is the novel aspect, as it addresses the common problem of BPNNs being sensitive to initial parameter choices, which can lead to suboptimal performance or getting stuck in local minima.
- Spectral Variable Selection: To improve model performance, the researchers employed a two-step process:
- Correlation Analysis: Pearson product moment correlation was used to identify spectral bands with the highest correlation to heavy metal content.
- VIF Analysis: Variance Inflation Factor (VIF) was used to eliminate multicollinearity among selected spectral variables.
- Combined Spectral Variables: Further, combinations of spectral variables (through addition, subtraction, multiplication, and division) were explored to capture more complex relationships.
The PSO algorithm iteratively searches for the optimal set of weights and thresholds for the BPNN, aiming to minimize the error between predicted and actual heavy metal concentrations.
Datasets
The study utilized two primary datasets:
- Field-collected Hyperspectral Data:
- Training Data: 50 soil samples from Guangdong, China, were used to train the models.
- Test Data: An additional 20 soil samples from the same region were used to validate the accuracy of the trained models.
- Total Field Samples: 90 soil samples were collected for determining heavy metal content (Cd, Hg, As) and acquiring hyperspectral reflectance data.
- Satellite Hyperspectral Imagery:
- HJ-1A Satellite Image: A hyperspectral image from the HuanJing-1A satellite, acquired on October 30, 2017, covering a part of Guangzhou, Guangdong province. This data was used to assess the feasibility of the PSO-BPNN model for regional-scale mapping.
Key Insights and Ideas
- Spectral Transformations Matter: Simple raw spectral bands are less effective than transformed spectral variables (like first derivative, continuum removal) for predicting soil heavy metal content. These transformations help highlight subtle spectral features.
- Combined Variables are Powerful: Combining different spectral transformations into new variables significantly improved correlation with heavy metal content, suggesting that the relationships are complex and not easily captured by single spectral features.
- PSO is Key to BPNN Performance: The integration of PSO with BPNN (PSO-BPNN) was crucial. It effectively optimized the network's initial parameters, leading to substantially higher estimation accuracy compared to a standard BPNN. This highlights the challenge of parameter tuning in neural networks and the effectiveness of swarm intelligence for this task.
- Satellite Data is Viable: The PSO-BPNN model demonstrated its capability to map soil heavy metal (specifically Cd) content even when using lower-resolution satellite hyperspectral imagery, opening doors for large-scale environmental monitoring.
Key Results
- Improved Accuracy: The PSO-BPNN method significantly outperformed the standard BPNN. It reduced the Mean Relative Error (MRE) by 68% to 71% and the Relative Root Mean Square Error (RRMSE) by 64% to 67% for Cd, Hg, and As. The coefficient of determination (R²) also saw significant improvements.
- Regional Mapping Success: For soil heavy metal Cd, the PSO-BPNN model, when applied to HJ-1A satellite imagery, achieved an RRMSE of 36%, indicating good accuracy for regional-scale mapping.
- Heavy Metal Levels: The study found that while average soil heavy metal concentrations (Cd, Hg, As) were within normative limits, maximum values in some samples exceeded environmental baselines, indicating localized contamination issues.
Limitations and Future Work
- Vegetation Canopy Effect: The accuracy of mapping using satellite imagery was lower than with field spectral data. The authors attribute this partly to the effect of vegetation canopies in the HJ-1A image and its coarser spectral resolution. Future work should focus on spectral unmixing techniques to isolate soil components and mitigate vegetation effects.
- Limited High-Concentration Samples: The training datasets had a prevalence of low heavy metal concentrations. This may affect the model's performance for samples with very high heavy metal content. The authors suggest collecting more samples with high concentrations to improve future model performance.
- Mapping Other Metals: This study focused on Cd mapping with satellite data due to the spectral limitations of the HJ-1A sensor. Future research could explore mapping Hg and As using hyperspectral data from satellites with wider spectral coverage, such as China's Gaofen-5 satellite.
Summary updated