1. Introduction
A mixed forest is defined as an area where at least two species coexist at any stage of development, sharing resources including light, water, and nutrients [
1]. The relative abundance of species can be quantified as a percentage proportion of the stand density, volume or coverage parameters of the canopy. For operational or forest inventory purposes it is a common practice to classify as “mixed” those forest stands where two (or more) tree species contribute each to more than 10–30% to stand basal area [
1]. Similar thresholds are used for canopy cover. For instance, the Corine Land Cover nomenclature defines “mixed forest” as the alternation of patches, groups or single trees of broadleaved and coniferous trees, over a minimum mapping unit of 25 ha, the share of both coniferous and broad-leaved species representing at least 25%, but maximum 75% of tree-covered area [
2]. Mixed-species stands have gained considerable traction in science and policy, especially in Europe for several reasons. Forests composed of several tree species are expected to beget biodiversity in the forest habitat, because tree species mixing sets the stage for variation in other structural components (e.g., tree size differentiation, tree layering). Further, several studies have demonstrated that mixed forest communities are generally more productive, resilient and capable of providing more ecosystem services than stands dominated by a single species [
3,
4,
5,
6,
7,
8,
9,
10]. Albeit that the knowledge about the ecology of mixed forests has recently expanded, almost all studies have been conducted on experimental or observational platforms, where pairs or triplets of pure and mixed stands, growing under similar environmental conditions, are compared in terms of productivity, growth stability, ecosystem services provision, without any prior knowledge on their actual the spatial extent in the investigated area. It is worth noting that the occurrence of mixed forest stands may represent an exception, rather than the rule, in many forest regions since, at least in Europe, managed forests are to a large extent human-made or secondary forests, often being established and managed as monocultures [
11]. Because of their rarity, mixed forest stands might not be sampled with sufficient intensity by random or systematic forest inventories, compared to monocultures. This limits the set-up of replicated experiments to analyze in real-world conditions the effect of mixing two or more tree species forest ecosystem properties and functions, and also to fill the existing knowledge gaps perceived by forest managers [
12]. From this perspective, mapping the actual occupancy of mixed stands on a fine spatial scale can be a fundamental tool for both experimental studies and monitoring activities such as forest inventories. Indeed, maps of occurrence of tree species mixtures vs. single species would allow, by a stratified sampling approach, the extraction of sufficiently large samples of mixed forest plots. Several projects and studies have been developed at European level to map the distribution of single tree species or forest type by remote sensing with acceptable classification results [
13,
14,
15,
16]. In this regard, mapping of forest habitats types dominated by one canopy species is relatively straightforward [
17,
18], but where forest landscape is more heterogeneous and includes mixed-species forests, the mapping procedures may become more challenging [
19].
In the scientific literature, there are several studies on forest-type mapping based on the freely-available multi-source and time-series imagery as high-resolution (HR) Sentinel-2 (S2) [
20,
21,
22]. However, the possibility to detect forest canopies dominated by different tree species at fine spatial scale greatly depends on the spatial resolution of the images. Indeed, [
23] have demonstrated that the spatial resolution of Sentinel-2A images (i.e., 10 m) may be insufficient for the classification of heterogeneous forests with fragmented species distribution and recommended combining these images with very high resolution data (VHR). VHR data have been successfully used in the identification of canopy tree species in simplified contexts such as urban areas [
24]. However, to the best of our knowledge no studies attempted to use VHR images to map mixed stands over fine spatial scales across forest landscapes. Hyperspectral data [
25,
26], especially if coupled with lidar data [
27,
28], have been successful in the identification of canopy tree species. However, we believe that much research effort still remains to be addressed to demonstrate that, for the operational task of tree species identification and mixed stands delineation, the semi-automatic classification of VHR multispectral data, can result in a consistent improvement over the business-as-usual method of photointerpretation.
One approach for developing semi-automatic procedures for mapping mixed forest stands is to rely on Object-Based Image Analysis (OBIA). Such technique is widely used to extract and classify information from high spatial detail imagery [
29]. The OBIA has been successfully applied in various fields of research [
30,
31] and, in particular, classifications for environmental studies [
32], and small-scale forest mapping [
33,
34]. The OBIA technique encompasses two main steps: (i) the “segmentation”, which is the delineation of homogeneous objects from the input imagery, following the principle of clustering neighboring image pixels into “objects”, so as to maximize the intra-object spectral homogeneity and inter-object spectral heterogeneity (ii) the “classification”, which labels and assigns each polygon to the target cover class [
35]. One of the advantages of segmentation is that it creates objects that can be associated to land cover types that may be spectrally variable at the pixel level and, thus, eliminates the “salt and pepper” effect associated with per-pixel classification [
36]. Another advantage is that OBIA delineates non-arbitrary units for analysis as opposed to pixels; objects can approximate real world features better than pixels [
36], thus resulting in better classification results, than pixel-based techniques, of high and very high spatial resolution imagery [
37,
38]. The OBIA is namely well-suited to detect the fine-scaled pattern of forest canopy and to delineate specific attributes (e.g., tree crowns, canopy gaps) [
24,
39,
40]. In addition, the use of shape features, hierarchical structures of objects and classes, and the topological features relating to the objects are other benefits of OBIA approaches. In particular, the Multi-Resolution Segmentation (MRS) [
35] can generate multiple hierarchical levels of image segmentation, i.e., a hierarchical set of image segmentations at different levels of spectral and shape homogeneity.
The main objective of this study is therefore to advance research in airborne VHR Object-Based Image Analysis (OBIA), with the goal to detect and map mixed forests of broadleaves and coniferous trees with a Minimum Mapping Unit (MMU) of 500 m2 in a forest-dominated landscape in Southern Italy. The selected MMU size approximately corresponds to the minimum area covered by forest inventory plots in National Forest Inventories in Europe.
4. Discussion
Study findings confirm the initial hypothesis, i.e., the possibility to delineate mixed stands on a fine spatial scale from VHR multispectral imagery (0.20 m). In particular, the OBIA approach here applied to the analysis of very high spatial resolution images proved to be a successful technique for detecting a fine-grained pattern of mixed forest in the investigated area, i.e., small patches (most often 500 to 2000 m
2) dispersed in a forest landscape characterized by forest tracts dominated by pure stands of broadleaved or coniferous trees. Considering that spatial variation between areas dominated by one species group or admixtures of the two species is continuous, the main strength of the proposed approach is to reduce as much as possible the elements of subjectivity in the delineation of boundaries between the three examined forest types, the main limit of the photointerpretation method, taking full advantage of multi-resolution segmentation potential. It is worth noting that, in the currently growing scientific literature reporting advances in the use of VHR imagery for forest type delineation [
53,
54], this study has demonstrated that the use of a simple and cheap single date image, despite its limited spectral resolution, allowed accurate mapping of mixed stands on the MMU of 500 m
2.
Under the examined conditions, the thematic accuracy of the map of the three forest types (conifers, broadleaves and mixed stands) achieved remarkable values, not lower than 0.73. Results obtained for the broadleaves and conifers classification were better than those for mixed forests, as expected. Indeed, many works have confirmed that pure conifer and broadleaved stands can be discriminated rather straightforwardly [
40,
55]. The results about mixed stands were less obvious, since previous studies had several difficulties in discriminating such stands from pure ones [
19,
56]. In this regard, the use of topographical variable such as “eastness” combined with the multispectral data makes it possible to resolve part of the spectral overlap between conifers and broadleaves, as a consequence of the different level of illumination of their canopy at the time of image acquisition due to aspect or slope. For example, broadleaved trees facing west or north may show a reflectance similar to that of conifers facing south. By introducing eastness, the classifier, even if the reflectance in the examined bands is the same, “learns” how to discriminate shaded deciduous trees or illuminated conifer canopies.
In this sense our results are encouraging, showing how the high resolution of the four-band orthorectified data and the OBIA methods are well suited for mapping mixed stands composed either by small groups or single trees of conifers and broadleaved trees.
Our results showed that most omission and commission errors are mainly due to a confusion between conifer and areas with mixtures of broadleaves such as chestnut or turkey oak. This could be explained by the spectral similarities between these groups of species in the period of acquisition of the image. In fact, during the late spring, the spectral signatures of chestnut or turkey oak could still be confused with the spectral reflectance of Corsican pine, because these tree species have a delayed leaf phenology compared to beech. This suggests that differences in canopy trees phenology are crucial for a successful broadleaved species discrimination. An airborne image acquired in early summer could have possibly solved most of the spectral confusion experienced in this study. Moreover, as highlighted in other studies [
40,
57], fall image tended to have a high discriminating ability when the leaf color changing process occurs. For these reasons, it is indisputable that the use of multitemporal data could bring improvements to classification accuracy. Indeed, multitemporal data could help to discriminate forest types that may be spectrally similar in any single time frame, especially if the appropriate timing of the images is selected, thus maximizing phenological differences and reducing redundant information which will not be used by the classifier. VHR commercial satellite imagery would be a suitable option to cover larger areas, but has considerable costs, especially if the mapping is required over large geographical areas (e.g., the cost of 0.5 m pan-sharpened imagery from Pleaides or WorldView is EUR 562 and 937 respectively, for 25 Km
2 minimum order area and 5% or less cloud cover). In this perspective, a possible further development of this study is to apply the proposed object-oriented classification methodology to the new generation of VHR multispectral satellite products characterized by frequent revisit times (e.g., PLANETScope monitoring products). Other possible developments of the study can be the use of alternative methods for the parameterization of multi-scale image segmentation (e.g., the Estimation of the Scale Parameter tool [
58]), or classifiers other than kNN, including for example Support Vector Machine (SVM) or Random Forest (RF).
Even with the above-mentioned limitations, the method proved to be effective to map admixtures of broadleaved deciduous and coniferous trees not only in terms of thematic accuracy, but also for the replicability of the image classification process. Indeed, the proposed method is an attempt to integrate, in a fairly transparent procedure, different for image classification algorithms (MRS, Assign Class, kNN), such that spectrally homogeneous tree crowns are identified and outlined, at the finer segmentation level, the tree species is recognized and classified, and the stand label is assigned and validated. Such an automated procedure does not exclude a careful input in the form of visual interpretation—in the form of training data and validation data. However, this support is limited to a negligible proportion of the investigated area. Consequently this (semi)automatic delineation of the target forest types from VHR imagery through MRS and classification demonstrated to be efficient also in term of time, if compared to visual interpretation.