Application of Deep Belief Network in Forest Type Identification using Hyperspectral Data

Xianxian Luo, Xu Song-ya, Hong YanAdvances 2020 · DOI

Abstract

Forest mapping by remote sensing is a critical task in forestry. While many researchers focus on forest type classification using various machine learning methods, the application of Deep Belief Networks (DBNs) remains relatively unexplored. This study assesses the feasibility of forest mapping in Dehua County, China, using HJ-1A hyperspectral data. By applying a DBN, we classified four forest types: coniferous, broad-leaved, mixed, and non-forest. Our results indicate that a DBN with three hidden layers and 256 nodes per layer achieves an overall accuracy of 85.8% and a Kappa coefficient of 0.785, significantly outperforming Support Vector Machines (SVMs).

Introduction

Accurate forest mapping is essential for conservation and ecological monitoring. However, hyperspectral imagery presents a unique challenge known as the "Hughes phenomenon": as the number of spectral bands increases, the training data required to maintain accuracy grows exponentially. Traditional classifiers like Support Vector Machines (SVMs) often struggle with this high-dimensional data without extensive manual feature engineering. This research explores whether Deep Belief Networks (DBNs) can automatically extract meaningful features from raw hyperspectral data to improve classification accuracy.

Methodology

The study utilized a Deep Belief Network (DBN) architecture, which consists of stacked Restricted Boltzmann Machines (RBMs). The training process involves two stages:

  1. Unsupervised Pre-training: The network learns the hierarchical structure of the hyperspectral data layer-by-layer, establishing a robust initial parameter state.
  2. Supervised Fine-tuning: Using labeled data and backpropagation, the model is refined to associate learned features with specific forest categories.

The researchers conducted a comprehensive hyperparameter search, testing network depths from 2 to 7 layers and hidden node counts from 16 to 512.

Datasets and Study Area

The study focused on the western part of Dehua County, Fujian Province, China. Data was acquired from the HJ-1A Hyper Spectral Imager (HSI), featuring 115 spectral bands (459–956 nm). After removing 29 bands due to stripe noise, 86 bands were used. The classification targeted four classes: Coniferous forest, Broad-leaved forest, Mixed forest, and Non-forest, using a dataset of 28,000 labeled pixels.

Key Insights and Results

The experimental results highlighted the importance of network architecture in deep learning for remote sensing.

The "Sweet Spot" for Depth: Contrary to the assumption that deeper networks always perform better, this study found that a depth of 3 hidden layers was optimal for the HJ-1A dataset. Beyond this point, accuracy began to decline, likely due to overfitting on the specific satellite dataset.

Overall Accuracy and Kappa coefficient across different network depths.

Optimal Node Configuration: The number of nodes in each hidden layer also significantly impacted performance. The model reached its peak effectiveness with 256 nodes per layer, showing a clear plateau in performance gains beyond that point.

Impact of hidden layer node count on classification performance.

Performance Comparison: The DBN model achieved an Overall Accuracy (OA) of 85.8% and a Kappa coefficient of 0.785. In comparison, a traditional SVM reached only 73% OA and a Kappa of 0.6447. This demonstrates the DBN's superior ability to handle high-dimensional hyperspectral data through automated feature extraction.

Conclusion

Deep Belief Networks offer a promising alternative to traditional machine learning for hyperspectral forest mapping. While challenges such as data noise and the need for large labeled datasets remain, the DBN's ability to outperform SVMs suggests it could become a dominant method in the field. Future work may involve using Generative Adversarial Networks (GANs) to address sample scarcity and testing the model's generalizability across different geographic regions.

Summary updated

Abstract

Forest mapping by remote sensing is a hot topics in forestry. At present, many researchers focus on the research of forest type classification or tree species identification using different machine learning methods and try to improve the accuracy of classification of satellite image. However, forest type classification using deep belief network (DBN) is still limited in previous literatures. Our research focuses on forest mapping in the western part of Dehua county in southern China. Most important objective was to assess the feasibility of forest mapping from hyperspectral data using deep learning. The HJ-1A hyperspectral data was adopted in this paper. We applied deep belief network and got a thematic map of four forest types, such as coniferous forest, broad-leaved forest, mixed forest and nonforest. Our finding shows that optimal network depth of DBN model is 3 and best node in each layer is 256 in our experiment. Overall accuracy is 85.8% and kappa coefficient is 0.785 with best-fit parameters in DBN model, while for SVM is 73% and 0.6447 respectively. DBN obtain better performance compared with support vector machine. Furthermore, network depth and number of nodes in each hidden layer in DBN model has a significant effect on overall accuracy and Kappa coefficient. In general, DBN is promised to be dominant method of forest mapping by hyperspectral data.
BibTeX
@inproceedings{luo2020application,
  author = {Xianxian Luo and XU Song-ya and Hong Yan},
  title = {{Application of Deep Belief Network in Forest Type Identification using Hyperspectral Data}},
  booktitle = {Advances},
  year = {2020},
  doi = {10.25046/aj0506186},
}
Figure