TY - JOUR

T1 - Beyond tree cover

T2 - Characterizing southern China's forests using deep learning

AU - Li, Qian

AU - Yue, Yuemin

AU - Liu, Siyu

AU - Brandt, Martin

AU - Chen, Zhengchao

AU - Tong, Xiaowei

AU - Wang, Kelin

AU - Chang, Jingyi

AU - Fensholt, Rasmus

N1 - Publisher Copyright: © 2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.

PY - 2023

Y1 - 2023

N2 - Mapping forests with satellite images at local to global scale is done on a routine basis, but to go beyond the mapping of forest cover and towards characterizing forests according to their types, species and use, requires a dense time-series of images. This knowledge is important, because ecological and economic values differ between forests. A new generation of low cost very high spatial resolution satellite images and the advent of deep learning enables improved abilities for distinguishing objects based on their structure, which could potentially also be applied to map different forest classes related to type, species and use. Here we use GF-1 images at 2 m resolution and map six forest classes including different planted species for the karst region in southwest China, covering 806,900 km2. We compare the results with field data and show that accuracies range between 78% and 90%. We show a dominance of plantations (15%) and secondary forests (70%), and only remnants of natural forests (6%). The possibility to map forest classes based on their crown structure derived from low cost very high-resolution satellite imagery paves the road towards sustainable forest management and restoration activities, supporting the creation of connected habitats, increasing biodiversity and improved carbon storage. No temporal information is needed for our approach, which saves costs and leads to rapid results that can be updated at a high temporal frequency.

AB - Mapping forests with satellite images at local to global scale is done on a routine basis, but to go beyond the mapping of forest cover and towards characterizing forests according to their types, species and use, requires a dense time-series of images. This knowledge is important, because ecological and economic values differ between forests. A new generation of low cost very high spatial resolution satellite images and the advent of deep learning enables improved abilities for distinguishing objects based on their structure, which could potentially also be applied to map different forest classes related to type, species and use. Here we use GF-1 images at 2 m resolution and map six forest classes including different planted species for the karst region in southwest China, covering 806,900 km2. We compare the results with field data and show that accuracies range between 78% and 90%. We show a dominance of plantations (15%) and secondary forests (70%), and only remnants of natural forests (6%). The possibility to map forest classes based on their crown structure derived from low cost very high-resolution satellite imagery paves the road towards sustainable forest management and restoration activities, supporting the creation of connected habitats, increasing biodiversity and improved carbon storage. No temporal information is needed for our approach, which saves costs and leads to rapid results that can be updated at a high temporal frequency.

KW - Deep learning

KW - forest types

KW - karst

KW - monoculture plantations

KW - remote sensing

U2 - 10.1002/rse2.292

DO - 10.1002/rse2.292

M3 - Journal article

AN - SCOPUS:85135593571

VL - 9

SP - 17

EP - 32

JO - Remote Sensing in Ecology and Conservation

JF - Remote Sensing in Ecology and Conservation

SN - 2056-3485

IS - 1

ER -