**1. Introduction**

Ecotone was first introduced by Clements [1] as transition zones where principal species from adjacent communities meet their limits. Ecotones are transitional zones between different habitats, and they exist at all spatial and temporal scales [2–4]. Generally, these zones have a set of characteristics (e.g., physiognomy, species composition, etc. [5]) uniquely defined by space and time scales and by the strength of the interactions between adjacent ecological systems [6]. Hence, they exist beside the boundaries between biomes or ecosystems [7]. Due to highly sensitive spatiotemporal dynamics, ecotones play various vital roles in community ecology, landscape ecology, and biodiversity conservation,

**Citation:** Wang, B.; Sun, H.;

Cracknell, A.P.; Deng, Y.; Li, Q.; Lin, L.; Xu, Q.; Ma, Y.; Wang, W.; Zhang, Z. Detection and Quantification of Forest-Agriculture Ecotones Caused by Returning Farmland to Forest Program Using Unmanned Aircraft Imagery. *Diversity* **2022**, *14*, 406. https://doi.org/10.3390/d14050406

Academic Editors: Lin Zhang, Jinniu Wang and Michael Wink

Received: 30 April 2022 Accepted: 18 May 2022 Published: 20 May 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

including the implementation of landscape functions, patterns, and ecological processes, as well as the provision of evidence that "indicates" or "forewarns of" the impacts of global climate change [8–12]. The intrinsic factors driving the spatiotemporal dynamic changes in ecotones are determined by the ecological processes of species migration, settlement, reproduction, and growth at the community scale [13,14].

The Yangtse River flood of 1998 resulted in the 'Returning Farmland to Forest Program' (RFFP) in China, also known as the Sloping Land Conversion Program (SLCP). It aims to restore some farmlands to forests gradually. This policy was initiated in 1999 and expanded in 2002 to cover most of China's provinces and has received immense attention [15,16]. The policy has become an essential factor in land cover changes and forest transition [17], especially in terms of the ecological processes that are responsible for material flows and species movement [18–20]. Studies of land cover change at regional and provincial scales find gains in forest cover in jurisdictions in which the RFFP was implemented [21,22]. Yet land cover change patterns vary sharply at smaller scales [21,23]. A previous study confirmed that ecotones often occur on the local scale [24]. These ecotones are dynamic entities whose changes are always influenced by the interactions of ecological processes between the abandoned agricultural lands and the adjacent ecosystems [25]. In vegetation (or plant community) science, knowledge of ecotones is crucial in understanding basic ecological patterns and processes [26]. Ecotones (or boundary) delineation is a critical issue in vegetation ecology, which can help identify the organization rules of communities and help understand the ecological processes between two adjacent ecosystems [27].

Formerly, several edge-detecting methods were used for detecting ecotones and transition zones using one-dimensional transect data [26]. The moving split-window (MSW) technique is the most used ecotone delineation method. It is hard to detect the spatial features of ecotones using the one-dimensional ecotone delineation technique, such as location, shapes, and spatial distribution patterns [8]. The delineation of ecotones at the community level is greatly affected by topography, data sampling directions, and the size and numbers of samples using one-dimensional MSW. However, these edge detection methods often need an excessive amount of effort and meticulous sampling to obtain ecological data in the same manner (grid, regular lattice) and amount (exhaustive study area) as pixel data based on transects. Delineating ecotones accurately continues to be a significant difficulty for vegetation and landscape ecologists.

Currently, for mapping the continuous spatial distribution and quantifying the degree of ecotone occurrence, algorithms also exist for detecting edges from the aerial view, and the satellite remotely sensed image data [28,29]. Remote sensing data acquired from satellites and piloted aircraft are potent tools for measuring the state and trends of environmental changes associated with natural processes and human disturbances. However, the conventional airborne and satellite remote sensing platforms upon which most sensors are mounted have not always met the needs of researchers and environmental professionals [30,31]. Not only do data gathered using conventional remote sensing systems lack operational flexibility and variety, but they also have a poor spatial and temporal resolution. With the fast growth of photogrammetry and 3D visualization technology in recent years, approaches for creating 3D models from regular digital images have evolved [32,33]. Researchers have used conventional RGB digital cameras mounted on a UAS to determine tree heights [34,35] and crop heights [36], and for biomass estimations [37,38]. Such applications have raised the possibility of using UAS-mounted ordinary digital cameras to obtain information on the heights and 3D structures of plant communities, thereby providing new research ideas and means to determine the spatial structures of ecotones. Hou and Walz [39] developed a method for detecting ecotones based on change in plant height and landcover by integrating several remote sensing data. However, data with varying resolutions may result in estimating errors.

Previous studies have suggested that 3D or even multidimensional in space and time information are key components of ecotones [11] and incorporating landscape 3D structures into modeling and monitoring will produce outcomes that better approximate reality [40]. An inherent property of forest-agricultural ecotones is higher structural heterogeneity (changes in vegetation height) than adjacent plant communities [5,41], permitting UASbased 3D photographic techniques to be used for ecotone delineation. The objective of this study is comparing the efficacy of UAS-derived canopy height versus traditional transect methods to extract the width of forest-abandoned-land ecotones at known RFFP sites. The UAS acquired photogrammetry point cloud and high-resolution orthoimage will be derived using UAS photogrammetry techniques in addition to field investigations using the transect method as ground truth. Optimizing limited field time is conducive to monitoring and researching landscape pattern changes and monitoring the associated ecological processes, thereby leading to better assessments of the mountainous ecological restoration process after farmland abandonment.

### **2. Materials and Methods**

### *2.1. Study Area*

The study area is located in Yunnan Province, China. Weixi, in the County of the Diqing Tibetan Autonomous Prefecture (98◦54 –99◦34 E, 26◦53 –28◦02 N) (Figure 1). The study area belongs to the temperate mountain monsoon climate, with the annual temperature difference of 11.3 ◦C, annual sunshine duration of 2071.3 h, annual precipitation of 954 mm, frost-free period of 195 days, and the vertical change of mountain climate is obvious. The forest soil in Weixi county showed an obvious vertical distribution, from high altitude to low altitude, alpine meadow soil, subalpine shrub meadow soil, bleaching earth, dark brown soil, yellow brown soil, and red soil. The soil in this study area is dark brown soil. It is a steep mountain area, and the prominent disturbances are grazing and firewood collection. There is a typical landscape with ecotones across the original Yunnan Pinus (*Pinus yunnanensis*) community and an abandoned land under the influence of the RFFP. Since 2003, the impact of this policy has been here for more than a decade. Before that, the study area was mainly planted with corn. In addition to Yunnan Pine, other dominant species in this study area include *Coriaria nepalensis* and *Desmodium yunnanense*.

**Figure 1.** Location of the sample plot. The right pic is the Baijixun sample plot, Weixi, with 19,203 m2. Furthermore, here are three sample transects that were be set up along the slope.

### *2.2. In-Situ Data*

Three parallel sample transects were set along the direction perpendicular to the forest edge, with an interval of 20 m. Taking the forest margin as the starting point, quadrats were set in two opposite directions, respectively. Starting from the edge of the forest, five quadrats with an area of 5 m × 5 m were set in the direction of the abandoned land. The original Yunnan pine forest was set with eight quadrats with an area of 10 m × 10 m continuously (Figure 1 and Figure S1). The area of each quadrat on abandoned land or forest has been confirmed using the nested sampling approach, which in earlier work considered the species–area relationship [42]. Based on the sampling strategies (S1), the work of plant species investigation in each of the quadrats was completed in May 2013. At the same time, in each quadrat, multiple points were randomly and uniformly selected to collect soil, and the soil was mixed to ensure its representativeness (see S1). The pH value and the organic matter (OM) content of soil were be measured using the potentiometric method and oil-bath heating potassium dichromate volumetric method, respectively (see Table S1) [42].

To detect the ecotones based on photogrammetric technology, a DJI M600Pro UAS platform (DJI, Shenzhen, China) was used to collect aerial imagery with a spatial resolution of 0.1 m. The platform uses an intelligent flight control system, a professional tripod head, and a Sony a7r (Sony, Tokyo, Japen) digital camera taking red (R), green (G), and blue (B) images for aerial photography. To assure comprehensive and practical remote sensing data, this research covers the following important steps: (1) The camera's lens length, shutter speed, and orientation were manually tuned depending on ground surface reflectance and ambient light conditions. (2) Nine GCPs (Ground Control Point) were put in open regions of the Baijixun sample plot, and their precise locations were gathered using RTK (Real-time kinematic) (CHCNAV, Shanghai, China). (3) For automated flying, the parallel path plan is used, with 80% longitudinal and 60% lateral overlap. (4) After UAS data collection, photos are sent to the field laptop and analyzed using Photoscan software (now Metashape). In May 2016, UAS flights at 133 m yielded 822 usable photographs. 131 photos were taken above our locations.

### *2.3. Image Preprocessing and Photogrammetric Processing*

Initially, some hazy images are filtered out owing to the swaying of trees caused by wind. The lightness of some shaded pics is enhanced, and the lightness of some pics with overexposure is debased by the set Brightness tool in Photoscan software. To avoid the influence from the difference of color between images in image stitching, the histogram matching was used to adjusted bias images. Sequentially, a series of parameters (medium aligning accuracy, high-quality point cloud, and mild depth-filtering level; refer to Figure 2) were adopted to produce high-quality photogrammetric point cloud data. It is important to emphasize that the point cloud data have 'medium aligning accuracy' (Figure 2). It is because of considering the influence from the high sensibility of aligning algorithm with 'high aligning accuracy', and it may cause an underestimate of canopy height, as no aligning can be made for the tie points at the top of the canopy. Nevertheless, it cannot reflect the real height of nonvegetation if the data with low accuracy. By the way, a point cloud data set sufficient to characterize 3D vegetation communities is constructed. Based on this, according to the orthoimage generation process flow of the software (Photoscan). Finally, a high resolution (0.08 m/pix) orthoimage is produced, and the degree of overlap is bigger than 90%.

**Figure 2.** Diagram of detecting and measuring ecotones utilizing UAS images, incorporating a preprocessing flow, a processing workflow of object-based classification, and the CHM extraction workflow. Prepreprocessing is represented by the boxes with a green backdrop. The red-background boxes represent object-based classification processing. Yellow boxes reflect methods for identifying, measuring, and validating ecotones derived from UAS. The parallelograms with a blue background indicate the data or outcomes created by the study framework.

### *2.4. Object-Based Image Classification (OBIC)*

Exploiting the high resolution of UAS imagery, OBIC was chosen to map land cover for the Baijixun study area. In this study, the multiresolution image segmentation (MRIS) algorithm is used [43]. It should be emphasized that in OBIC, the accuracy of segmentation directly affects the accuracy of classification. Hence, the ESP tool was adopted to estimate the optimal value of segmentation parameters. It is an effective method to determine the optimal value of segmentation parameters by controlling variables in the MRIS [44]. In our case, the Shape and Compactness are adjusted to 0.5 and 0.5, respectively (Figures 2 and S3), and the scale parameter related to MRIS is adjusted to 20 to obtain a better result of segmentation (Figure S4)

Afterward, 12581 objects from segmentation are classified into different land cover classes using the random forest (RF) classifier. The RF classifier was chosen mainly because this machine learning algorithm does not need many samples, has fast training speed, and has good antioverfitting ability [45]. In this study, vegetation, nonvegetation, and shadow were distinguished based on orthoimage. The training samples combined the data from the previous situ investigation, the information of GCPs, and some sample points selected based on visual inspection. Combined with field survey data and visual inspection, a total of 138 training sample points were obtained, 30 of which were shaded, 38 of which were nonvegetated, and 70 of which were vegetated (Figure S2). The three bands (R, G, B) were inputted as predictive variables. Out of 113 selected samples, 101 (approximately 90% of the total) sample points were selected as training samples, and the remaining samples were used as validation samples. Then, based on the number of samples and the number of classification targets, the default setting of the RF classifier was accepted. Corresponding objects were determined according to the spatial position of investigation, and the samples of the training classifier were finally involved in the form of objects. In order to detect more small biotopes from the high-resolution orthoimage, characteristic of structure (i.e., CHM), geometrical characteristic (area and shape index), and texture information were crucial. The contrast index and the dissimilarity index were selected for the texture index, generated according to the gray level co-occurrence matrix. Moreover, a rule-based classification scheme was constructed, with different rules and thresholds for each class (Figure 2). Different subclasses have significant differences in different feature dimensions, so every subclass was judged from the parent class by the most sensitive features. We completed the processing of OBIC above in eCognition Developer software.
