Mapping local density of young Eucalyptus plantations by individual tree detection in high spatial resolution satellite images

FOREST ECOLOGY AND MANAGEMENT

卷:301

Article

Zhou, Jia^1,2  Proisy, Christophe³  Descombes, Xavier²  le Maire, Guerric^4,5  Nouvellon, Yann^4,6  Stape, Jose-Luiz⁷  Viennois, Gaelle⁸  Zerubia, Josiane²  Couteron, Pierre³ 

[1] Univ Montpellier 2, UMR AMAP, F-34398 Montpellier 5, France

[2] INRIA, Sophia Antipolis Mediterranee, F-06902 Sophia Antipolis, France

[3] IRD, UMR AMAP, F-34398 Montpellier 5, France

[4] CIRAD, UMR Eco&Sols, F-34093 Montpellier 01, France

[5] CIRAD, UMR TETIS, F-34093 Montpellier 5, France

[6] Univ Sao Paulo, Dept Atmospher Sci, BR-05508090 Sao Paulo, Brazil

[7] N Carolina State Univ, Dept Forestry & Environm Sci, Raleigh, NC 27695 USA

[8] CNRS, UMR AMAP, F-34398 Montpellier 5, France

关键词: Crown identification;    Object detection;    Stochastic point process;    Forests;    Remote sensing;    Brazil;   

DOI  :  10.1016/j.foreco.2012.10.007

来源: Elsevier

【 摘 要 】

Local tree density may vary in young Eucalyptus plantations under the effects of environmental conditions or inadequate management, and these variations need to be mapped over large areas as they have a significant impact on the final biomass harvested. High spatial resolution optical satellite images have the potential to provide crucial information on tree density at an affordable cost for forest management. Here, we test the capacity of this promising technique to map the local density of young and small Eucalyptus trees in a large plantation in Brazil. We use three Worldview panchromatic images acquired at a 50 cm resolution on different dates corresponding to trees aged 6, 9 and 13 months and define an overall accuracy index to evaluate the quality of the detection results. The best agreement between the local densities obtained by visual detection and by marked point process modeling was found at 9 months, with only small omission and commission errors and a stable 4% underestimation of the number of trees across the density gradient. We validated the capability of the MPP approach to detect trees aged 9 months by making a comparison with local densities recorded on 112 plots of similar to 590 m(2) and ranging between 1360 and 1700 trees per hectare. We obtained a good correlation (r(2) = 0.88) with a root mean square error of 31 trees/ha. We generalized detection by computing a consistent map over the whole plantation. Our results showed that local tree density was not uniformly distributed even in a well-controlled intensively-managed Eucalyptus plantation and therefore needed to be monitored and mapped. Use of the marked point process approach is then discussed with respect to stand characteristics (canopy closure), acquisition dates and recommendations for algorithm parameterization. (c) 2012 Elsevier B.V. All rights reserved.

【 授权许可】

Free