The advantages of using drones over space-borne imagery in the mapping of mangrove forests

Monika Ruwaimana; Behara Satyanarayana; Viviana Otero; Aidy M Muslim; Muhammad Syafiq A; Sulong Ibrahim; Dries Raymaekers; Nico Koedam; Farid Dahdouh-Guebas

doi:10.1371/journal.pone.0200288

. 2018 Jul 18;13(7):e0200288. doi: 10.1371/journal.pone.0200288

The advantages of using drones over space-borne imagery in the mapping of mangrove forests

Monika Ruwaimana ^1,^2,^3,^‡,^*, Behara Satyanarayana ^1,^4,^‡, Viviana Otero ¹, Aidy M Muslim ⁴, Muhammad Syafiq A ⁴, Sulong Ibrahim ⁴, Dries Raymaekers ⁵, Nico Koedam ², Farid Dahdouh-Guebas ^1,²

Editor: Shijo Joseph⁶

PMCID: PMC6051606 PMID: 30020959

Abstract

Satellite data and aerial photos have proved to be useful in efficient conservation and management of mangrove ecosystems. However, there have been only very few attempts to demonstrate the ability of drone images, and none so far to observe vegetation (species-level) mapping. The present study compares the utility of drone images (DJI-Phantom-2 with SJ4000 RGB and IR cameras, spatial resolution: 5cm) and satellite images (Pleiades-1B, spatial resolution: 50cm) for mangrove mapping—specifically in terms of image quality, efficiency and classification accuracy, at the Setiu Wetland in Malaysia. Both object- and pixel-based classification approaches were tested (QGIS v.2.12.3 with Orfeo Toolbox). The object-based classification (using a manual rule-set algorithm) of drone imagery with dominant land-cover features (i.e. water, land, Avicennia alba, Nypa fruticans, Rhizophora apiculata and Casuarina equisetifolia) provided the highest accuracy (overall accuracy (OA): 94.0±0.5% and specific producer accuracy (SPA): 97.0±9.3%) as compared to the Pleiades imagery (OA: 72.2±2.7% and SPA: 51.9±22.7%). In addition, the pixel-based classification (using a maximum likelihood algorithm) of drone imagery provided better accuracy (OA: 90.0±1.9% and SPA: 87.2±5.1%) compared to the Pleiades (OA: 82.8±3.5% and SPA: 80.4±14.3%). Nevertheless, the drone provided higher temporal resolution images, even on cloudy days, an exceptional benefit when working in a humid tropical climate. In terms of the user-costs, drone costs are much higher, but this becomes advantageous over satellite data for long-term monitoring of a small area. Due to the large data size of the drone imagery, its processing time was about ten times greater than that of the satellite image, and varied according to the various image processing techniques employed (in pixel-based classification, drone >50 hours, Pleiades <5 hours), constituting the main disadvantage of UAV remote sensing. However, the mangrove mapping based on the drone aerial photos provided unprecedented results for Setiu, and was proven to be a viable alternative to satellite-based monitoring/management of these ecosystems. The improvements of drone technology will help to make drone use even more competitive in the future.

Introduction

With their luxuriant growth in tropical and subtropical latitudes along the land-sea interface, bays, estuaries, lagoons and backwaters [1], mangroves provide important ecosystem services, ranging from coastal protection to fisheries production, ecotourism, phytoremediation, carbon sequestration, and other services [2–4]. The value of mangroves were estimated to be between 4,185 and 50,349 USD per km² [5]^, with a major part of these coming from indirect services [6]. Because of the less-valued direct services like firewood, timber and honey [6], mangrove ecosystems are often unappreciated or underestimated [7,8], resulting in conversion into settlements, agriculture, and aquaculture areas in many parts of the world [9–11]. In light of the decreasing mangrove cover and subsequent consequences [2,12–14], the world’s remaining mangroves need appropriate conservation and management, for which accurate and updated information is necessary [15].

Remote sensing is indispensable for mangrove research due to its time saving and cost-effective nature compensating for the fieldwork, which is difficult to carry out, especially in areas of low accessibility [11,15–18]. Regular monitoring/mapping of the mangroves based on remote sensing can provide authentic information (along with spatial-temporal dynamics) that is needed for better management [19,20]. In addition, past and present vegetation maps are useful to reconstruct past events and predict future development scenarios [21–23]. To date, both satellite images and airplane-based aerial photographs have been used for mangrove mapping [17,19,21,23–26]. Various types of satellite images–obtained from very low- to very high-resolution (VHR) sensors, have been used in relation to the scientific targets, including estimation of global mangrove cover [27,28] and regional/local species’ level distribution [29–31]. While some low- to moderate-resolution satellite data (e.g. Landsat, Sentinel) have still been available for free [28], the VHR images are expensive, and researchers using them are often constrained by their budget [11,32]. In addition, availability of cloud-free satellite images for mangrove mapping is a known difficulty [15,33].

Unlike satellite imagery, aerial photos are not challenged by cloud cover because their acquisition (aircraft flight) time can be adjusted to local weather conditions [34]. Their spatial resolution ranges from submeter to meter to centimeter level, depending on the flight altitude [15,34,35]. Aerial photos have probably been the only source of image documentation prior to satellite technology for identifying the past land-use/cover, and to date still produce imagery with the best spatial resolution for retrospective studies [36–38]. The significance and usefulness of archived and recently taken aerial photographs for genus to species level classification of the mangroves is evident from diverse publications [15,21,39].

The main limitation of an aircraft’s aerial photography for vegetation mapping is its high operational cost [15]. However, this situation is changing, with the revolutionary mode of data acquisition through Unmanned Aerial Vehicles (UAVs) or drones, which are not only reducing the cost of aerial photography, but also the cost of equipment, due to the production of updated models in the market [34]. Besides having higher spatial resolution [40,41], the sequential drone images can provide point-cloud 3D views to estimate tree height and crown diameter [42,43]. Although drone images have been used for monitoring and management of the terrestrial vegetation—from tall canopy trees to invasive weeds, agricultural crop yields, and other purposes [34,44–48]—the implications of their use for mangrove research are still limited, except for a few recent publications on leaf area index and the inventory of production forests [43,49]

In this study, we mapped the mangroves at Setiu Wetland, Malaysia, based on drone (DJI-Phantom-2) and satellite (Pleiades-1B) images, and compared the respective results. Being one of the first attempts in scientific literature to apply a drone for mangrove species’ mapping, our objective was not only to identify the potential of drone technology for mangrove research, but also to compare both drone and satellite imagery in terms of image quality (i.e. spatial, spectral, radiometric and temporal resolution), efficiency (i.e. coverage area, data acquisition/processing time and user-cost), and land-use/cover (i.e. object- and pixel-based) classification accuracy.

Materials and methods

Study area

Setiu Wetland is located in the State of Terengganu on the Peninsular Malaysia (05⁰36’30”-05⁰42’30”N, 102⁰40’30”-102⁰48’30”E) (Fig 1). The entire wetland covers 230km² of non-mangrove wetland, 8.8km² of water, and 4.18km² of mangrove vegetation [50,51]. This ecosystem is represented by several coastal features, including beaches, sea, mudflats, lagoons, estuaries, rivers, islands, seagrass beds, and coastal vegetation, including mangroves [51]. The mangroves are composed of 23 true and 38 associate species [52]. Major mangrove species in this wetland include Avicennia alba Blume, A. lanata Ridley, Ceriops decandra (Griff.) Ding Hou, Bruguiera sexangula (Lour.) Poir., B. gymnorrhiza (L.) Lamk., B. cylindrica (L.) Blume, Lumnitzera racemosa Willd., Rhizophora apiculata Bl. and Nypa fruticans (Thunb.) Wurmb. [51], along with the critically endangered B. hainesii C.G. Rogers [7,16,53].

Setiu Wetland plays a significant role for local livelihoods, especially in terms of aquaculture and related activities [54,55]. The local communities visit the wetland regularly to collect clams (e.g. Anadara and Placuna spp.), crabs (e.g. Scylla spp.), and honey, and Nypa palm leaves for personal and commercial usage (pers. obs.). Currently, the local government is taking the necessary measures to declare this area as a State Park Reserve [56]. Of the 4.18km² of mangrove cover at Setiu, the present analysis of vegetation mapping focused on an area of 0.12km² (Fig 1) where the species composition varies, and could represent different spectral reflectance conditions. In recent years, the increase of oil palm plantations in the vicinity is believed to have decreased the freshwater input into this lagoon, especially on the northern side (Cik Azmi, village head, pers. comm.) where several muddy-sand areas with a suboptimal growth of the mangrove trees (height <3m) are noticeable (pers. obs.).

Fieldwork

Ground inventory

The fieldwork was carried out in June-July 2015, under the permit from Institute of Oceanography and Environment as an authority which managed research on Setiu Wetland. (S4le). Plot-based (5×5m²) measurements were obtained from mangroves (from the waterfront to the terrestrial edge, based on a pre-determined grid), and the existing land-use/cover (e.g. Casuarina, Pandanus, aquaculture, building and other features) was recorded with camera and GPS. Altogether, 101 mangrove plots and 155 ground verification points were investigated from the entire wetland area. However, for the present paper, only the ground inventory details of the area that corresponded to drone data coverage/analysis were considered.

Aerial photos acquisition

An overview of the remote sensing approaches, from mangrove aerial photos acquisition using a drone, to drone/satellite images processing/analysis and their results comparison, is represented in a schematic flowchart (Fig 2).

Fig 2 — The 10 ROI sets were named 1A-1B to 5A-5B. Except the manual rule-set algorithm, the remaining algorithms i.e., automatic, maximum likelihood and spectral angle mapping, were used 10 times (10×) for running the object- and pixel-based classification approaches (grey and white shades are for visualization purposes).

For aerial photography, a DJI-Phantom-2 drone and Zenmuse gimbal (to stabilize the camera) was used. Flight tracks at 100m altitude were planned in DJI Ground Station v.04 software, and uploaded through a DJI2.4G Datalink. Two cameras—a regular red-green-blue (RGB) SJ4000wifi (12megapixel, wide angle, focal length 24mm and CMOS sensor 22mm×36mm), and another infrared (IR) SJ4000 (modified by IRPro with 22Blu dual-pass-bands filter) were used. Besides the known spectral range of the CMOS sensor (i.e. 350–1100nm) [57], both RGB and IR cameras were tested again in the Laboratory of Photonics Research at the Vrije Universiteit Brussel using Spectro 320 of the Instrument Systems and Image J software [58]. As the CMOS sensor has no specific spectral range for individual RGB bands, we considered the RGB values of each pixel as its band values.

The IR camera was adjusted to take pictures through a fastest time lapse interval of 5s at 2.5m/s flying speed, while the RGB camera was adjusted for 2s at 4m/s flying speed. The difference in time lapse between IR and RGB cameras was due to model specifications/restrictions. External data logger, Flytrex Core 2, was installed to record the flight track information (i.e. altitude, geographic coordinates and time of acquisition) [59].

Data processing

Drone images

Detailed image processing protocol is available in S2 File and protocol.io (https://dx.doi.org/10.17504/protocols.io.qh7dt9n). GPS log data was imported from the drone using Gpsvisualizer [60], the individual photos were tagged by synchronising camera time and flight track information in Geosetter [61]. The tagged images was georeferenced and mosaiced in Agisoft Photoscan [62], with at least five Ground Control Points (GCPs) of prominent features from Google Earth^TM. The mosaic images were exported as a regular image with RGB values and Digital Elevation Model (DEM). The coordinate reference system was assigned to EPSG:32648 WGS84/UTMzone48N. The produced RGB, IR and DEM images were further rectified using Quantum Geographic Information System (QGIS v.2.12.3-Lyon software), with 12 GCPs from satellite image and visual observations based on field work. The images were then merged into a single image with seven bands (Table 1). In addition, edges of the image affected by the parallax effect (showing tall objects as elongated with invalid DEM) were cropped out.

Table 1. List of bands in a RGB, IR and DEM merged imagery of the DJI-Phantom-2 drone.

Band number	Source camera	Information	Digital value	Unit
Band 1	RGB	DEM	-10–30	Meter
Band 2	IR	NIR	0–255	8-bit colour value
Band 3	IR	NIR	0–255	8-bit colour value
Band 4	IR	NIR	0–255	8-bit colour value
Band 5	RGB	Red	0–255	8-bit colour value
Band 6	RGB	Green	0–255	8-bit colour value
Band 7	RGB	Blue	0–255	8-bit colour value

Open in a new tab

Satellite imagery

A Pleiades-1B image of the Setiu Wetland purchased and authorized by University of Malaysia Terangganu (S3 File), which is dated on 16 August 2013 with the spatial resolution of 2m –duly corrected by Astrium Services (the image distributor company) for radiometric and sensor distortions–was used for the present study. This multispectral image, covering 100km², was first pan-sharpened in QGIS (via Ratio Component Substitution (RCS) algorithm), and then cropped to generate the same mangrove coverage area as that of the drone imagery for comparison of the results.

Region of Interest (ROI) creation

The ROIs for land cover training sites and classification accuracy assessment were generated based on the ground-truthing data (using Semi-Automatic Classification Plugin (SCP) in QGIS) [63]. Three ROI groups—one for Pleiades, and two for drone image—were considered for the classification. While two ROI groups—one meant for Pleiades and another meant for the drone, contained the same number of (dominant) land-cover classes, the third ROI group of drone images has extra classes representing other visible (non-dominant) features on the ground. According to Mather (2004) [64], the number of pixels per land-use/cover class selected for training sites must be at least 30 times the spectral dimension. Therefore, we selected 360 pixels per class in Pleiades. However, in the case of higher resolution drone data, each corresponding class was represented by 36,000 pixels. All training sites were then separated into 20 polygons of the same size in both images. By random selection, 50% of the polygons in each class were selected as training sites. The remaining 50% unselected polygons were used for accuracy assessment. This random selection was repeated to create at least five paired ROI sets (named after 1A-1B to 5A-5B as shown in Fig 2), resulting in ten iterations of cross validation. [65,66].

Object-based classification

Both the Pleiades and drone images were segmented by Large-Scale Mean-shift Classification (LSMC) region growing algorithm [67,68] in QGIS (Orfeo toolbox). The vector segments were classified based on OpenGIS simple features reference implementation (OGR) classifier with a Support Vector Machine (SVM) algorithm, as well as a Manual rule-set (MAN) classification. While OGR is a free automatic (AUT) application to run the classification [67], the MAN is based on signature details and spectral distances, together with visual interpretation. The MAN was applied to assign a class for each object/segment, and produced the classification. With the AUT, both Pleiades and drone images were classified ten times by each ROI set, but only once using the MAN (Fig 2). The classified images were exported into raster files for accuracy assessment.

Pixel-based classification

Pixel-based classification was carried out in SCP QGIS using the ROI training sites. Since the accuracy of a handheld GPS (Garmin 45, USA) is ca. 5–6 m, field data were superimposed on the drone imagery, and visually checked for authenticity of the training sites. After the classification attempts using three algorithms i.e., Minimum Distance, Spectral Angle Mapping (SAM), and Maximum Likelihood (MLI), the results obtained from SAM and MLI were found to be appropriate for this study [63,69]. Both Pleiades and drone images were classified ten times by each ROI set under the MLI and the SAM classification scenarios (Fig 2), and exported for accuracy assessment.

Accuracy assessment

Accuracy analysis in the form of an error matrix was generated by comparing the classified (object- and pixel-based) images against the paired ROIs as cross validation (using SCP QGIS). For instance, the image classified by ROI 1A was validated by ROI 1B and vice versa (Fig 2) [65,66]. Based on this error matrix, overall accuracy (OA) and Kappa index were derived for the entire classification of each image; and specific producer accuracy (SPA) and specific user accuracy (SUA) for each land-cover class [63,69]. The error matrix was also incorporated into the Pontius Matrix to estimate ‘quantity’, ‘exchange’ and ‘shift’ parameters for identifying the source of classification error [70,71]. While ‘quantity’ (%) represents the amount of pixels that differed between training sites and classification per class, ‘exchange’ (%) shows the allocated error by the number of pixels that interchanged between two classes, and ‘shift’ (%) denotes the other allocation differences that were not included in the quantity and exchange differences [71].

Results comparison

The capability of Pleiades-1B and DJI-Phantom-2 was evaluated through their image quality in terms of spatial, spectral, radiometric and temporal resolution; efficiency in terms of area coverage, data acquisition/processing time and user-cost, and accuracy in terms of object- and pixel-based classification approaches. Results of the accuracy assessment (i.e., OA, Kappa, SPA, SUA and Pontius Matrix—quantity, exchange and shift) were represented by boxplots using the R-Studio [72]. Statistical variations among the classification accuracies and approaches were identified through Kruskal-Wallis and Mann-Whitney tests using Past v.3.14 software [73].

Results

Image quality

The drone image had a higher spatial resolution compared to the pan-sharpened Pleiades image (2.8cm vs 50cm, Table 2 & Fig 3). In term of spectral resolution, the satellite image performed better due to specific sensors with a definite wavelength. Although spectral range of the drone cameras were tested (RGB = 450–675nm and IR = 875–1100nm), each specific band wavelength was not known. Pleiades has a higher radiometric resolution compared to drones (12bit vs. 8bit). With a daily revisit schedule of the Pleiades satellite, the data is available for every day. In addition, a drone can deliver the aerial photos daily, but not under rainy or stormy conditions.

Table 2. Spatial, spectral, radiometric and temporal resolutions of the Pleiades-1B satellite and DJI-Phantom-2 drone images (source for Pleiades-1B information: Pleiades user guide [74]).

	Spatial resolution (cm)	Spectral resolution (nm)	Radiometric resolution	Temporal resolution
Pan-sharpened Pleiades imagery	50	Blue: 430–550 Green: 500–620 Red: 590–710 NIR: 740–940	12-bit (0–4095)	Daily
DJI-Phantom-2 drone imagery:
i) Red-Green-Blue (RGB) bands (SJ4000wifi camera)	2.8	450–675	8-bit (0–255)	Daily if no rain/storm
ii) Infrared (IR) (SJ4000 camera)	3.5	875–1100	8-bit (0–255)
iii) Digital Elevation Model (DEM)	10	-	-

Open in a new tab

Fig 3 — (A-B) The images for mangrove vegetation mapping at the Setiu Wetland. Red boxes show the zoomed-in subsets of–(C) Pleiades-1B and, (D) DJI-Phantom-2 drone, revealing mangrove and non-mangrove details on the ground at 50cm and 5cm spatial resolutions respectively.

Data efficiency

Cost and coverage

The cost of the DJI-Phantom-2 drone (with an RGB SJ4000 camera) used in this study was 950USD (April 2015). In addition, the cameras and other accessories (gimbal, GPS logger, batteries) cost 1003USD. Together with the fieldwork expenses of 500USD, the total budget spent was 2453USD. On the other hand, the Pleiades-1B imagery was procured for 1750USD. Out of the two months’ fieldwork, two weeks were focused on the aerial photos acquisition, and obtained 19 composite (RGB, IR and DEM) images, covering area of 1.81km² (Fig 4). If the mangrove surface area being covered in the present aerial photos is considered, then the drone data acquisition cost was ca. 1355USD per km² (2453USD ÷ 1.81km²), which is more expensive than the Pleiades-1B satellite data (17.5USD per km²). The Pleiades data is still economical, even after adding the cost of present fieldwork expenses (22.5USD per km²). Although each of our 15 minutes drone flights, corresponding to the average battery run-time at a given speed/altitude, typically covered 0.12km², the parallax cropping reduced it to 0.09–0.11km² (Fig 4).

Fig 4 — While new tasking/purchasing order of Pleiades images requires at least 100km² coverage, the archived data of each image is available for a minimum of 25 km². A drone is expected to work efficiently (if it does not crash or have technical problems) for 500 flights. If an efficient drone flight for 15 min (corresponding to average battery run-time) can cover approximately 0.1 km², the total drone flights would be able to cover ca. 50 km². However, with an improved battery run-time of up to 20 min these days, the same drone can deliver aerial photos of an area covering up to 75 km² (box dimensions are arbitrary, and the colours are for visualization purposes only).

Image acquisition and processing time

Under emergency conditions, the Pleiades satellite data is available to the user within 24 hours. Drone aerial photos can also be acquired daily under favourable weather conditions. The mangrove area of 0.12km² chosen for the present mapping study was represented by 2,800 and 12MB sizes of drone and Pleiades images respectively. Due to the large data size of the drone image (ca. 230 times greater than satellite data), its processing time—especially in the case of pixel-based classification, was found to be ten times greater than the satellite image (Table 3). Meanwhile, segmentation process in the object-based classification has greatly reduced the drone imagery size (from 2,800 into 194MB) as well as the processing time.

Table 3. Data processing time of DJI-Phantom-2 drone and Pleiades-1B satellite images.

For the drone, the accuracy analysis of pixel-based classification was conducted twice due to there being two different training sites–one representing dominant land-use/cover classes, and another representing both dominant and non-dominant classes visible on the ground.

	Time taken for drone imagery(h)	Time taken forsatellite imagery.(h)
Image preparation
Geotagging	2	-
Mosaicking	8	-
Joining	6	-
Pansharpening	-	0.5
Cropping	0.5	0.1
ROI creation
Data interpretation	1	1
Creating random polygons	2	1
Creating random sets	2	1
Object-based classification
Segmentation	4	1
Object based automatic classification (10 iterations)	2	1
Manual rule-set classification (1 iteration)	2	1
Dissolve segment	6	1
Convert to raster	2	1
Accuracy analysis (10 iterations)	2	1
Total time	18	6
Pixel-based classification
Maximum Likelihood (10 iterations)	10	1
Spectral Angle Mapping (10 iterations)	10	1
Accuracy Analysis (20 iterations)	30	3
Total time	50	5
Total time for entire analyses	89.5	14.6

Open in a new tab

Data classification

Land-cover categories

Based on the visual interpretation and the ground-truthing, it was possible to distinguish six dominant (i.e. water, land, non-mangrove Casuarina equisetifolia, and mangrove species A. alba, N. fruticans and R. apiculata) and four non-dominant (B. cylindrica, L. racemosa, S. alba and dead trees) land-cover classes at Setiu. R. apiculata and A. alba spectral signatures show relatively high homogeneity, while the N. fruticans signature shows the highest heterogeneity. While the dominant classes are visually clear in both Pleiades and drone images, the non-dominant classes are best observed in the drone imagery (Fig 5).

Fig 5 — (A) Land-cover classes based on Pleiades-1B satellite (on the left) and DJI-Phantom-2 drone (on the right). Due to the poor demarcation of some features in the Pleiades, only 1–6 land-cover classes were considered for its image classification. (B) Locations of the ten (1–10) land-use/cover classes marked on the Pleiades-1B (top) and DJI-Phantom-2 drone (bottom) images.

Image classification and accuracy

Besides the classification of drone and satellite images for dominant land-cover classes, the higher resolution drone images were additionally classified for both dominant and non-dominant classes (Table 4). The object-based (with AUT and MAN algorithms) and the pixel-based (with MLI and SAM algorithms) classification approaches produced 93 maps in total, of which the ones with the best OA were used for the present publication (Table 4). The OA, Kappa, SPA and SUA values of all object- and pixel-based classification iterations and the error matrix of each classified map used in this publication are available in the S1 parts 1 and 2.

Table 4. Details of the classification approaches and resultant maps (with accuracy iterations) using DJI-Phantom-2 drone and Pleiades-1B satellite images for the Setiu Wetland.

Each classified map was given a unique identification code that starts with ‘D’ for drone and ‘S’ for satellite, followed by a number of the land-use/cover classes used (10 = all ten land-use/cover categories and 6 = dominant six classes), classification approach (O = object-based and P = pixel-based), and the algorithm (MAN = Manual rule-set, AUT = Automatic, MLI = Maximum Likelihood and SAM = Spectral Angle Mapping) (OA = Overall Accuracy) (ROI = Region of Interest) (*training site codes follow those used in Fig 2).

Image source	No. of class-es	Classification approach	Classification.Algorithm	Abbre-viation	No. of classification iterations	No. of accuracy iterations	*ROI training sites that produced the highest accuracy maps and its OA	Mean and standard deviation of each classification method
Drone	10	Object-based (O)	Manual rule-set (MAN)	D10OMAN	1	10	4B(71.8%)	69.7±1.2%
(D)			Automatic (AUT)	D10OAUT	10	10	5A(68.7%)	65.8±3.4%
	10	Pixel-based (P)	Maximum Likelihood (MLI)	D10PMLI	10	10	4A(79.4%)	77.6±1.3%
			Spectral Angle Mapping (SAM)	D10PSAM	10	10	3B(55.0%)	51.8±1.6%
	6	Object-based (O)	Manual rule-set (MAN)	D6OMAN	1	10	2B(94.8%)	94.0±0.5%
			Automatic (AUT)	D6OAUT	10	10	5A(85.9%)	78.4±5.3%
	6	Pixel-based (P)	Maximum Likelihood (MLI)	D6PMLI	10	10	2A(93.0%)	90.0±1.9%
			Spectral Angle Mapping (SAM)	D6PSAM	10	10	3B(76.0%)	72.0±1.9%
Satellite	6	Object-based (O)	Manual rule-set (MAN)	S6OMAN	1	10	5A(77.8%)	72.2±2.7%
(S)			Automatic (AUT)	S6OAUT	10	10	1B(75.9%)	70.8±8.1%
	6	Pixel-based (P)	Maximum Likelihood (MLI)	S6PMLI	10	10	3B(88.3%)	82.8±3.5%
			Spectral Angle Mapping (SAM)	S6PSAM	10	10	4B(77.3%)	73.3±5.1%

Open in a new tab

For the object-based classification, the MAN algorithm provided more accurate results than AUT (Fig 6), whereas for the pixel-based classification, MLI was more competent than SAM (Fig 7). The drone image classified through MAN for six dominant land-cover classes (D6OMAN) provided a higher OA (94.0±0.5%) in contrast to the Pleiades (S6OMAN: 72.2±2.7%) or the drone with ten classes (D10OMAN: 69.7±1.2%) (Table 4 & Fig 8). In addition, in the case of pixel-based classification, the drone image classified through MLI for dominant land-cover classes (D6PMLI) provided good accuracy (90.0±1.9%) as compared to the Pleiades (S6PMLI: 82.8±3.5%) or the drone with ten classes (D10PMLI: 77.6±1.3%) (Table 4 & Fig 8).

Fig 7 — Map sbased on Maximum Likelihood algorithm (A-C) and the Spectral Angle Mapping algorithm (D-F). The images shown have the highest Overall Accuracy from 10 iterations. (Abbreviations and genus names follow those used in Fig 6).

Fig 8 — (A) Overall Accuracy, (B) Kappa Index, (C) Specific Accuracy and, (D) Specific Reliability. Each abbreviation along the X-axis follows the map identification codes in Table 4.

Both AUT (for object-based classification) and SAM (for pixel-based classification) algorithms provided low (D10PSAM: 51.8±1.6%) to moderate accuracy maps (D10OAUT: 65.8±3.4%; S6OAUT: 70.8±8.1%; D6PSAM: 72.0±1.9%; S6PSAM: 73.3±5.1%), except for the drone imagery that was classified through AUT for dominant classes (D6OAUT: 78.4±5.3%) (Table 4 & Fig 8). While there was an inconsistent performance of the MAN algorithm (D6OMAN, S6OMAN, D10OMAN), the MLI showed rather high OA, Kappa, SPA and SUA values (D10PMLI, D6PMLI, S6PMLI) (Fig 8). Among others, the higher threshold values of OA and Kappa were confined only to the drone image classified through the MAN (D6OMAN) and MLI (D6PMLI) algorithms. Overall, the statistical variations among these classification accuracies and approaches were found to be significant (S1 part 1)

In terms of the classification errors, identified from the first three highest OA observed land-cover maps (Fig 9), the drone image subjected to object-based classification through MAN for dominant classes (D6OMAN) showed an overestimation of R. apiculata and A. alba. In the case of pixel-based classification, both drone and satellite images classified through MLI for dominant classes (D6PMLI and S6PMLI) had their classification errors spreading among the various vegetation classes.