The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B3-2022
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B3-2022, 913–920, 2022
https://doi.org/10.5194/isprs-archives-XLIII-B3-2022-913-2022
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B3-2022, 913–920, 2022
https://doi.org/10.5194/isprs-archives-XLIII-B3-2022-913-2022
 
30 May 2022
30 May 2022

DETECTING AND EVALUATING DISTURBANCE IN TEMPERATE RAINFOREST WITH SENTINEL-2, MACHINE LEARNING AND FOREST PARAMETERS

E. Kutchartt1, J. Hernández2, P. Corvalán2, Á. Promis3, and F. Pirotti1,4 E. Kutchartt et al.
  • 1TESAF, Department of Land, Environment, Agriculture and Forestry, University of Padova, Via dell’Università 16, 35020 Legnaro (PD), Italy
  • 2Departamento de Gestión Forestal y su Medio Ambiente, Universidad de Chile, Santa Rosa 11315 La Pintana, Santiago Chile, Chile
  • 3Departamento de Silvicultura y Conservación de la Naturaleza, Universidad de Chile, Santa Rosa 11315 La Pintana, Santiago Chile, Chile
  • 4CIRGEO, Interdepartmental Research Center of Geomatics, University of Padova, Via dell’Università 16, 35020 Legnaro (PD), Italy

Keywords: forest alterations, anthropogenic drivers, spectral indices, basal area, aboveground tree biomass

Abstract. Earth observation via remote sensing imagery provides a fast way to define alteration levels. In this work 12 stands of Araucaria-Nothofagus forests were selected in southern Chile, which represented four alteration levels: (i) None (ii) Low (iii) Medium and (iv) High. The stands were surveyed measuring 379 field plots and Google Earth Engine was used to collect a composite of Sentinel-2 images over a one-year range, from June 2019 to June 2020. The following approaches were tested: (i) aggregating the normalized difference vegetation index (NDVI) of each image and selecting the 95th and 99th percentile values of NDVI for each pixel; (ii) creating a composite imagery with best pixels over one year timeline using NDVI as weighting factor and NDVI value band itself (NDVI) – this is similar to the 99th percentile in the previous point, but with maximum values of NDVI; (iii) aggregating the composite as in the previous approach, but using the full spectral information of Sentinel-2 and then random forest machine learning for classification over alteration areas with k-fold validation with k=5. Results show that the 95th and 99th percentile of NDVI values from approach (i) do not discriminate the four classes correctly. The maximum NDVI from approach can distinguish all four classes. It must be noted through that statistical significance does not necessarily imply a strong practical significance; medium and high alterations have very similar NDVI distributions. Random forest results provided an F-score for each class higher than 80% except for the “medium alteration” class.