A model for forest type identification and forest regeneration monitoring based on deep learning and hyperspectral imagery
Feng‐Cheng Lin, Yi-Shiang Shiu, Pei-Jung Wang, Uen-Hao Wang, Jhe-Syuan Lai, Yung-Chung Chuang
Abstract
Traditional ground-based forest survey methods involve high labor costs, and their inefficiency makes comprehensive forest resource surveys challenging. With the development of new sensors and vehicles in recent years, more diverse and novel remote sensing detection and survey techniques have emerged. This study aims to use hyperspectral imagery to classify forest types containing representative tree species. To verify the feasibility of the proposed methods, we used hyperspectral imagery from the Taiwan Forestry Experiment Institute's Liugui Research Forest in southern Taiwan, which has an area of 9882 ha and an altitude of 250–2600 m. Hyperspectral imagery offers several advantages compared to traditional multispectral imagery; it captures a broad spectrum of contiguous, narrow spectral bands, providing highly detailed spectral information, enabling differentiation of tree components that appear similar in multispectral imagery. Eight identifiable forest types were selected for the models considered, and three different deep learning algorithms, VGG19, ResNet50 and a proposed combination (VGG19 + ResNet50), were used to screen the best algorithms. Data formats and pre-processing methods that can effectively improve computational performance were explored. The research results found that: (1) Band filtering is a necessary means to improve calculation performance; (2) Flattening the original convolution kernel with cubic characteristics can significantly reduce the time required for calculation. In terms of simulation results, VGG19 + ResNet50 was identified as the best model. Its overall classification accuracy can generally reach 93% to 100%. According to the calculation process set in this study, the time required for model training can be shortened to less than 30 min. The results of this research will help process more detailed and complex information in forest resource management and more accurately quantify forest ecology and woodland conditions.