Pine plantation structure mapping using WorldView-2 multispectral image

Estimation of Boreal Forest Attributes from Very High Resolution Pléiades Data

“…The Pléiades-based models estimating BA achieved the lowest accuracy of the different target features, with RMSEs of 5.1 m 2 ·ha −1 (22%) and 6.6 m 2 ·ha −1 (25%), respectively. This is slightly better than Shamsoddini et al [19] and St-Onge et al [16], who reported the BA in a pine plantation at 7.7 m 2 ·ha −1 (23%) and 7.1 m 2 ·ha −1 (16%), respectively. However, it is higher both in absolute and relative terms compared to Yu et al [7], who reported the BA with 4.3 m 2 ·ha −1 (16%), and Ozdemir and Karnieli [17], who reported the BA with 1.8 m 2 ·ha −1 (17%).…”

“…However, it is interesting that the latter two received such low RMSE values from models based merely on spectral/textural features. The pine plantation used as field data by Shamsoddini et al [19] might have increased the textural importance, as the forest is similarly managed. This hypothesis is strengthened by the results from Kayitakire et al [15], whose test site almost entirely consisted of a well-managed spruce plantation.…”

“…The accuracy of the height-modelled metrics is similar or slightly better than past studies based on other VHR sensors. When image-matched heights from the Ikonos sensor were used, the dominant forest height was estimated with 1.7 m RMSE [16], and when models based purely on spectral/textural features were used, Shamsoddini et al [19] estimated the tree height with 1.9 m RMSE using WorldView-2 imagery; correspondingly, Kayitakire et al [15] estimated the top height with 2.2 m, using Ikonos imagery. However, it is interesting that the latter two received such low RMSE values from models based merely on spectral/textural features.…”

“…A few studies have shown these to be of similar importance as the image-matched height metrics [21,22] and some other are based purely on textural analysis [15,17,19]. The literature honour co-occurrence analysis, and specifically the grey-level co-occurrence matrix (GLCM) method with its related attributes.…”

“…It is a way of extracting second-order statistical texture features, while the spectral derivatives can be considered first-order features, as they do not consider pixel neighbour relationships. GLCM was developed by Haralick, Shanmugam, and Dinstein [31], and has commonly been applied in remote sensing studies [15,17,19,22,32,33]. Ten common textural attributes (contrast, dissimilarity, homogeneity, second moment, energy, max probability, entropy, average, variance, and correlation) were computed using the Sentinel Application Platform (SNAP; http://step.esa.int/main/toolboxes/snap) commissioned by the European Space Agency (ESA).…”