Rights statement: This is the author’s version of a work that was accepted for publication in Remote Sensing of Environment. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Remote Sensing of Environment, 261, 2021 DOI: 10.1016/j.rse.2021.112470
Accepted author manuscript, 5.07 MB, PDF document
Available under license: CC BY-NC-ND
Final published version
Research output: Contribution to Journal/Magazine › Journal article › peer-review
Research output: Contribution to Journal/Magazine › Journal article › peer-review
}
TY - JOUR
T1 - Tracking small-scale tropical forest disturbances
T2 - Fusing the Landsat and Sentinel-2 data record
AU - Zhang, Y.
AU - Ling, F.
AU - Wang, X.
AU - Foody, G.M.
AU - Boyd, D.S.
AU - Li, X.
AU - Du, Y.
AU - Atkinson, P.M.
N1 - This is the author’s version of a work that was accepted for publication in Remote Sensing of Environment. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Remote Sensing of Environment, 261, 2021 DOI: 10.1016/j.rse.2021.112470
PY - 2021/8/31
Y1 - 2021/8/31
N2 - Information on forest disturbance is crucial for tropical forest management and global carbon cycle analysis. The long-term collection of data from the Landsat missions provides some of the most valuable information for understanding the processes of global tropical forest disturbance. However, there are substantial uncertainties in the estimation of non-mechanized, small-scale (i.e., small area) clearings in tropical forests with Landsat series images. Because the appearance of small-scale openings in a tropical tree canopy are often ephemeral due to fast-growing vegetation, and because clouds are frequent in tropical regions, it is challenging for Landsat images to capture the logging signal. Moreover, the spatial resolution of Landsat images is typically too coarse to represent spatial details about small-scale clearings. In this paper, by fusing all available Landsat and Sentinel-2 images, we proposed a method to improve the tracking of small-scale tropical forest disturbance history with both fine spatial and temporal resolutions. First, yearly composited Landsat and Sentinel-2 self-referenced normalized burn ratio (rNBR) vegetation index images were calculated from all available Landsat-7/8 and Sentinel-2 scenes during 2016–2019. Second, a deep-learning based downscaling method was used to predict fine resolution (10 m) rNBR images from the annual coarse resolution (30 m) Landsat rNBR images. Third, given the baseline Landsat forest map in 2015, the generated fine-resolution Landsat rNBR images and original Sentinel-2 rNBR images were fused to produce the 10 m forest disturbance map for the period 2016–2019. From data comparison and evaluation, it was demonstrated that the deep-learning based downscaling method can produce fine-resolution Landsat rNBR images and forest disturbance maps that contain substantial spatial detail. In addition, by fusing downscaled fine-resolution Landsat rNBR images and original Sentinel-2 rNBR images, it was possible to produce state-of-the-art forest disturbance maps with OA values more than 87% and 96% for the small and large study areas, and detected 11% to 21% more disturbed areas than either the Sentinel-2 or Landsat-7/8 time-series alone. We found that 1.42% of the disturbed areas indentified during 2016–2019 experienced multiple forest disturbances. The method has great potential to enhance work undertaken in relation to major policies such as the reducing emissions from deforestation and forest degradation (REDD+) programmes.
AB - Information on forest disturbance is crucial for tropical forest management and global carbon cycle analysis. The long-term collection of data from the Landsat missions provides some of the most valuable information for understanding the processes of global tropical forest disturbance. However, there are substantial uncertainties in the estimation of non-mechanized, small-scale (i.e., small area) clearings in tropical forests with Landsat series images. Because the appearance of small-scale openings in a tropical tree canopy are often ephemeral due to fast-growing vegetation, and because clouds are frequent in tropical regions, it is challenging for Landsat images to capture the logging signal. Moreover, the spatial resolution of Landsat images is typically too coarse to represent spatial details about small-scale clearings. In this paper, by fusing all available Landsat and Sentinel-2 images, we proposed a method to improve the tracking of small-scale tropical forest disturbance history with both fine spatial and temporal resolutions. First, yearly composited Landsat and Sentinel-2 self-referenced normalized burn ratio (rNBR) vegetation index images were calculated from all available Landsat-7/8 and Sentinel-2 scenes during 2016–2019. Second, a deep-learning based downscaling method was used to predict fine resolution (10 m) rNBR images from the annual coarse resolution (30 m) Landsat rNBR images. Third, given the baseline Landsat forest map in 2015, the generated fine-resolution Landsat rNBR images and original Sentinel-2 rNBR images were fused to produce the 10 m forest disturbance map for the period 2016–2019. From data comparison and evaluation, it was demonstrated that the deep-learning based downscaling method can produce fine-resolution Landsat rNBR images and forest disturbance maps that contain substantial spatial detail. In addition, by fusing downscaled fine-resolution Landsat rNBR images and original Sentinel-2 rNBR images, it was possible to produce state-of-the-art forest disturbance maps with OA values more than 87% and 96% for the small and large study areas, and detected 11% to 21% more disturbed areas than either the Sentinel-2 or Landsat-7/8 time-series alone. We found that 1.42% of the disturbed areas indentified during 2016–2019 experienced multiple forest disturbances. The method has great potential to enhance work undertaken in relation to major policies such as the reducing emissions from deforestation and forest degradation (REDD+) programmes.
KW - Deep learning
KW - Downscaling
KW - Forest disturbance
KW - Landsat and Sentinel-2
KW - Small-scale clearing
KW - Forestry
KW - Image enhancement
KW - Vegetation
KW - Down-scaling
KW - Fine resolution
KW - Forest disturbances
KW - LANDSAT
KW - Landsat and sentinel-2
KW - Ratio images
KW - Small scale
KW - Tropical forest
KW - Tropics
U2 - 10.1016/j.rse.2021.112470
DO - 10.1016/j.rse.2021.112470
M3 - Journal article
VL - 261
JO - Remote Sensing of Environment
JF - Remote Sensing of Environment
SN - 0034-4257
M1 - 112470
ER -